Intelligent Monitoring Method and System for the Entire Clinical Trial Cycle Based on Multidimensional Data Fusion

Abstract

This invention discloses a method and system for intelligent monitoring of the entire clinical trial cycle based on multidimensional data fusion. To address the problems of high false alarm rates, lack of interpretability, and difficulty in forming a monitoring closed loop under multidimensional evidence fusion conditions, this invention acquires and standardizes multidimensional clinical trial data and its source data, constructs an evidence layer graph and a source layer graph, and establishes cross-layer associations to form a two-layer evidence graph. Based on source features, the source credibility is calculated and assigned to evidence layer nodes and edges. Source credibility gating is introduced into the heterogeneous graph Transformer attention mechanism for risk inference and risk score output. The risk score is calibrated for confidence and grouped by trial stage and research center. Group-based shape-preserving prediction is used, and the group warning threshold is determined based on the target false alarm rate to output warning labels and risk levels. The minimum sufficient evidence chain is extracted from the triggering warning object for interpretation. Under the condition of frozen model weights, feedback is used to update the credibility parameters, calibration parameters, and thresholds to form a closed loop, achieving the technical effects of low false alarms, interpretable risk warning grading, and a closed-loop intelligent monitoring throughout the entire cycle.

Description

Technical Field

[0001] This invention relates to the field of clinical trial data management, and in particular to a method and system for intelligent monitoring of the entire clinical trial lifecycle based on multidimensional data fusion. Background Technology

[0002] Clinical trials generate a large amount of multidimensional data throughout the entire lifecycle, including subject enrollment, visit execution, investigational drug management, adverse event recording, and data entry and verification. Data sources include electronic data capture systems, laboratory testing systems, imaging systems, and on-site records from research centers. With the promotion of risk-based monitoring, the industry has gradually moved from traditional monitoring methods primarily based on on-site verification to centralized and remote monitoring methods that combine key risk indicators. Data standardization, statistical testing, and machine learning methods are being introduced to identify abnormal patterns, thereby improving monitoring efficiency and reducing compliance risks. In recent years, graph modeling and deep learning inference for complex correlated data have also begun to be used to characterize the relationships between subjects, visits, and events, supporting more refined risk identification and early warning.

[0003] The existing technology still has the following shortcomings:

[0004] 1. The integration of multi-source and multi-type data often remains at the field level or single-table statistics, making it difficult to simultaneously depict the structural relationships and temporal evolution between monitored objects, resulting in insufficient utilization of complex evidence combinations in risk identification;

[0005] 2. Early warning strategies often rely on fixed thresholds or lack reliable uncertainty control. When there are differences in research centers, changes in the experimental stage, or data distribution drift, the false alarm rate is prone to increase, making it difficult to achieve stable early warning classification based on the target false alarm rate.

[0006] 3. Existing methods do not make sufficient use of data sources and data change trajectories, lack a mechanism to introduce source credibility into risk reasoning, and the interpretation of early warning results often remains at the level of rule description or feature importance, making it difficult to form a traceable chain of evidence. Furthermore, there is a lack of a feedback-driven security closed-loop update mechanism, making it difficult to continuously optimize monitoring strategies.

[0007] Therefore, a method and system for intelligent monitoring of the entire clinical trial cycle 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 an intelligent monitoring method for the entire clinical trial cycle based on multi-dimensional data fusion. Addressing the problems in existing technologies, such as the difficulty in characterizing the relationships between monitored objects through multi-source multi-dimensional data fusion, high false alarm rates due to fixed warning thresholds, poor robustness across research centers and trial phases, and the lack of traceable evidence chains to explain warning results, making it difficult to form a controllable closed-loop optimization, this invention proposes a technical solution that constructs an evidence layer graph and a traceability layer graph, and establishes a two-layer evidence graph through cross-layer association. Based on traceability features, the source credibility is calculated and assigned to evidence layer nodes and edges. Source credibility gating is introduced into the heterogeneous graph Transformer attention calculation for risk inference and outputting a risk score. The risk score is calibrated for confidence and grouped by research center and trial phase. Shape-preserving prediction is used for grouping, and the warning threshold for each group is determined based on the target false alarm rate to output a warning label and risk level. The minimum sufficient evidence chain is extracted to explain the triggering object. Finally, under the condition of freezing model weight parameters, monitoring feedback is used to update the credibility parameters, calibration parameters, and grouping thresholds to form a closed loop. This invention achieves the effects of low false alarms, interpretability, traceability, and safe iterative optimization throughout the entire clinical trial cycle.

[0009] This invention provides a method for intelligent monitoring of the entire clinical trial cycle based on multidimensional data fusion, comprising:

[0010] S1. Obtain multidimensional clinical trial data and corresponding traceability data for the entire cycle of the same clinical trial, standardize the multidimensional clinical trial data, and generate a standardized record set;

[0011] S2. Construct an evidence layer diagram based on a standardized record set to represent the monitored objects and their relationships;

[0012] S3. Construct a source traceability layer graph based on the source traceability data to characterize the data source and data change trajectory, and associate it with the evidence layer graph to generate a two-layer evidence graph containing the evidence layer graph and the source traceability layer graph. Calculate the source credibility based on the source traceability features corresponding to the nodes and edges of the evidence layer graph in the source traceability layer graph, and assign the source credibility to at least one of the nodes and edges of the corresponding evidence layer graph in the two-layer evidence graph to generate a two-layer evidence graph with source credibility.

[0013] S4. Input the two-layer evidence graph with source credibility into the heterogeneous graph Transformer network for risk reasoning. In the attention calculation of the heterogeneous graph Transformer network, source credibility is introduced as a gating factor to adjust the contribution of nodes and edges of the evidence layer graph to the attention weight, and output the risk score for the monitored object.

[0014] S5. Perform confidence calibration on the risk score, divide the calibrated risk score into multiple groups according to the preset grouping rules, determine the warning threshold for each group based on the preset target false alarm rate, perform conformal prediction output on the calibrated risk score, and generate a risk warning result set including warning labels and risk levels.

[0015] S6. For the monitoring objects that are marked as triggered by the risk warning results, extract the minimum set of nodes and the minimum set of edges that can support the warning mark and the risk level based on the two-layer evidence graph with source credibility, and sort the minimum set of nodes and the minimum set of edges based on time information to generate the minimum sufficient evidence chain and output the interpretation result set.

[0016] S7. Obtain and interpret the monitoring feedback data corresponding to the result set, including the early warning confirmation result and the early warning handling result. Under the condition of freezing the heterogeneous graph Transformer network weight parameters, update at least one of the calculation parameters of source credibility, the calibration parameters of confidence calibration, and the group early warning threshold based on the monitoring feedback data to obtain the updated parameter set. In subsequent monitoring cycles, the updated parameter set is used to perform parameterized execution of steps S3 to S5 to form a closed loop of intelligent monitoring throughout the entire clinical trial cycle.

[0017] Optionally, S1 includes:

[0018] Acquire multidimensional clinical trial data for the entire cycle of the same clinical trial, including subject identifier, visit identifier, event identifier, data item name, data item value, unit, event occurrence time, and record generation time;

[0019] Obtain source traceability data corresponding to each item of the multidimensional clinical trial data. The source traceability data includes data source identifier, source record identifier, data entry time, data version information, and data modification records.

[0020] Based on a preset data dictionary, the multidimensional clinical trial data is mapped to generate data items with unified field names. Based on a preset unit conversion table, the values ​​of the data items are unified in units. Based on preset encoding rules, the names of the data items and the values ​​of the data items are encoded in a unified manner.

[0021] Based on a preset time base, the event occurrence time and the record generation time are time-aligned to obtain the aligned event time;

[0022] The results of performing the field mapping, unit unification, encoding unification, and time alignment, along with the subject identifier, visit identifier, event identifier, and source record identifier consistent with the traceability data, are then summarized to generate the standardized record set.

[0023] Optionally, S2 includes:

[0024] Based on the standardized record set, a set of monitoring target nodes is determined and the evidence layer graph is constructed. The set of monitoring target nodes includes subject nodes identified by subject identifiers, visit nodes identified by visit identifiers, research event nodes identified by event identifiers, and research drug nodes and safety event nodes generated from records in the standardized record set corresponding to research drugs and safety events.

[0025] Based on the standardized record set, a set of association edges is determined and the evidence layer graph is constructed. The set of association edges includes association edges between subject nodes and visit nodes, association edges between visit nodes and study event nodes, association edges between subject nodes and study drug nodes, and association edges between subject nodes and safety event nodes.

[0026] The aligned event times in the standardized record set are used as the time attributes of the corresponding nodes, and the data items with unified field names and their values ​​in the standardized record set are used as the attributes of the corresponding nodes, generating an evidence layer diagram to characterize the monitored object and its relationships.

[0027] Optionally, S3 includes:

[0028] Based on the source data, a source traceability layer graph is constructed, wherein the source traceability layer graph includes data source nodes, source record nodes, and change record nodes. The data source nodes are used to represent the data source identifier, the source record nodes are used to represent the source record identifier and the data version information, and the change record nodes are used to represent the data entry time and the data modification record. Source association edges between the data source nodes and the source record nodes, as well as change association edges between the source record nodes and the change record nodes, are established in the source traceability layer graph.

[0029] Based on the source record identifiers in the standardized record set, cross-layer associations are established between the source record nodes in the tracing layer graph and the corresponding nodes and edges generated by the standardized record set in the evidence layer graph to generate a two-layer evidence graph.

[0030] The source credibility is calculated based on the source tracing features associated with the corresponding nodes and edges in the evidence layer graph of the source tracing layer graph. The source tracing features include data entry time, data version information, and data modification records. The calculated source credibility is then assigned to the corresponding evidence layer graph nodes and edges in the two-layer evidence graph to generate the two-layer evidence graph with source credibility.

[0031] Optionally, S4 includes:

[0032] The node attributes and edge attributes of the evidence layer graph in the two-layer evidence graph with source credibility are encoded into node initial feature vectors and edge initial feature vectors, respectively, and the node initial feature vectors and edge initial feature vectors are input into the heterogeneous graph Transformer network.

[0033] In the heterogeneous graph Transformer network, linear transformations are performed on different node types and edge types of the evidence layer graph to obtain query vectors, key vectors and value vectors. Attention scores are calculated based on the query vectors and key vectors. At the same time, gating weights are generated based on the source credibility of the corresponding nodes and edges participating in the attention calculation. The attention scores are modulated with the gating weights and then normalized to obtain attention weights.

[0034] The value vectors of adjacent nodes are weighted and aggregated based on the attention weights to update the node representation vector of the target node, and the node representation vector of the monitored object is obtained through iterative updates of a preset number of layers.

[0035] The node representation vector is input into the risk prediction layer to obtain the initial risk score. Multiple initial risk scores are obtained by performing multiple forward calculations on the same monitored object, and the dispersion of the multiple initial risk scores is calculated to obtain the prediction uncertainty.

[0036] Based on the experimental phase information and research center information corresponding to the monitored objects in the two-layer evidence diagram with source credibility, the group identifier is determined, and the initial risk result set including the monitored object identifier, initial risk score, prediction uncertainty and group identifier is generated.

[0037] Optionally, S5 includes:

[0038] A calibration sample set is determined based on pre-acquired monitoring feedback data. The calibration sample set consists of samples that correspond to the historical initial risk result set and have early warning confirmation results.

[0039] Confidence calibration is performed on the initial risk scores in the initial risk result set to obtain a calibration function. The initial risk scores are then converted into calibrated risk scores using the calibration function to generate a calibrated risk result set containing the monitoring object identifier, the calibrated risk score, the prediction uncertainty, and the group identifier.

[0040] The calibrated risk result set is divided into multiple calibration groups based on the group identifier.

[0041] For each calibration group, based on the samples belonging to that calibration group in the calibration sample set, an inconsistency score is determined using the post-calibration risk score, and a group warning threshold is determined from the inconsistency score according to a preset target false alarm rate;

[0042] The group warning threshold is applied to the monitoring objects belonging to the corresponding calibration group in the post-calibration risk result set. When the inconsistency score corresponding to the monitoring object meets the warning triggering condition, a warning mark is generated. The risk level is determined based on the comparison result between the post-calibration risk score and the preset risk classification threshold, and the risk warning result set is generated.

[0043] Optionally, S6 includes:

[0044] From the set of risk warning results, filter the monitoring objects marked as triggered by the warning to determine the set of monitoring objects to be explained;

[0045] For each monitored object to be explained, a candidate subgraph is extracted based on the two-layer evidence graph with source credibility, with the evidence layer graph node corresponding to the monitored object to be explained as the center node and within the preset number of hops.

[0046] The contribution of each node and edge in the candidate subgraph to the initial risk score is calculated, and stepwise elimination is performed in the candidate subgraph based on the contribution to obtain the minimum set of nodes and the minimum set of edges that satisfy the following conditions: while retaining the minimum set of nodes and the minimum set of edges, the difference between the initial risk score obtained by using the heterogeneous graph Transformer network to perform risk reasoning on the monitored object to be explained and the initial risk score corresponding to the monitored object in the initial risk result set does not exceed a preset tolerance.

[0047] The event times corresponding to the nodes and edges in the minimum node set and the minimum edge set are sorted to generate the minimum sufficient evidence chain.

[0048] Output the interpretation result set, which includes the monitoring object identifier, the minimum sufficient evidence chain, and the data source identifier and time information corresponding to each piece of evidence in the minimum sufficient evidence chain.

[0049] Optionally, the S7 includes:

[0050] Obtain monitoring feedback data corresponding to each item in the interpretation result set. The monitoring feedback data includes the monitoring object identifier, the early warning confirmation result, and the early warning handling result.

[0051] The calculation parameters for updating the source credibility based on the monitoring feedback data include weight parameters for the source tracing features. When the early warning confirmation result indicates a false alarm, the weight parameter of the source tracing feature corresponding to the false alarm is reduced, and when the early warning confirmation result indicates a true positive, the weight parameter of the source tracing feature corresponding to the true positive is increased.

[0052] The calibration parameters for updating the confidence calibration are based on the monitoring feedback data. The updated confidence calibration parameters are used to improve the matching degree between the calibrated risk score and the early warning confirmation result.

[0053] The monitoring feedback data is used to update the group warning threshold. The updated group warning threshold includes re-estimating the group warning threshold within each calibration group based on a preset target false alarm rate.

[0054] With the weight parameters of the heterogeneous graph Transformer network frozen, the updated calculation parameters of the source confidence, the calibration parameters of the confidence calibration, and the group warning threshold are summarized to generate the updated parameter set, and the updated parameter set is called in subsequent monitoring cycles to execute steps S3 to S5.

[0055] Optionally, the tracing layer graph further includes data version integrity verification information for the source record node. The integrity verification information includes the hash value generated for each version of the source record and the hash chain of adjacent versions. When an inconsistency in the hash chain is detected, the source credibility of the evidence layer graph nodes and edges associated with the source record node across layers is reduced.

[0056] On the other hand, the present invention also provides an intelligent monitoring system for the entire clinical trial cycle based on multidimensional data fusion, comprising:

[0057] The data standardization module is used to acquire multidimensional clinical trial data and its corresponding traceability data for the entire cycle of the same clinical trial, and to standardize the multidimensional clinical trial data to generate a standardized record set.

[0058] The dual-layer evidence graph generation module is used to construct an evidence layer graph based on the standardized record set, construct a source tracing layer graph based on the source tracing data, establish cross-layer associations to generate a dual-layer evidence graph, calculate the source credibility based on the source tracing layer graph, and assign the source credibility to at least one of the nodes and edges of the evidence layer graph to obtain a dual-layer evidence graph with source credibility.

[0059] The risk reasoning module is used to input the two-layer evidence graph with source credibility into the heterogeneous graph Transformer network, and use the source credibility as a gating factor to adjust the attention weight in the attention calculation, and output the risk score of the monitored object.

[0060] The early warning output module is used to calibrate the confidence level of the risk score and group it according to preset rules. For each group, it determines the early warning threshold based on the preset target false alarm rate and performs conformal prediction. It outputs a risk early warning result containing early warning markers and risk levels.

[0061] The explanation module is used to extract the minimum set of nodes and the minimum set of edges from the two-layer evidence graph for the monitored object that triggered the warning, sort them by time to generate the minimum sufficient evidence chain, and output the explanation result.

[0062] The closed-loop update module is used to obtain monitoring feedback data, and under the condition of freezing the weight parameters of the heterogeneous graph Transformer network, update at least one of the source confidence parameters, calibration parameters, and group warning thresholds, and call the updated parameters to perform monitoring in subsequent monitoring cycles.

[0063] The beneficial effects of this invention are:

[0064] 1. By constructing an evidence layer graph and a source tracing layer graph and establishing cross-layer associations to form a two-layer evidence graph, the source credibility is calculated based on the source tracing features, and the contribution of nodes and edges is adjusted in a gating manner in the heterogeneous graph Transformer attention. This enables risk reasoning to suppress the interference of low-credibility source evidence, improve the robustness and accuracy of risk identification, and thus reduce false alarms.

[0065] 2. By calibrating the confidence level of the risk score and grouping them according to the experimental stage and research center, the grouped shape-preserving prediction is adopted and the grouped early warning threshold is determined based on the target false alarm rate. This ensures that the early warning trigger has a controllable upper limit for false alarms, adapts to the differences in data distribution between different centers and different stages, and achieves stable risk early warning and hierarchical output.

[0066] 3. For the monitored objects that trigger the early warning, the minimum set of nodes and the minimum set of edges that satisfy the tolerance constraints are extracted from the two-layer evidence graph and sorted by time to generate the minimum sufficient evidence chain. At the same time, the corresponding data source identifier and time information are output to make the early warning results interpretable and traceable. Under the condition of freezing the main model weight parameters, the confidence parameters, calibration parameters and thresholds are updated by feedback to form a safety closed loop, supporting continuous optimization throughout the entire cycle. Attached Figure Description

[0067] 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:

[0068] Figure 1 This is a flowchart of a method and system for intelligent monitoring of the entire clinical trial cycle based on multidimensional data fusion, as proposed in this invention. Detailed Implementation

[0069] 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.

[0070] refer to Figure 1 A method for intelligent monitoring of the entire clinical trial cycle based on multidimensional data fusion, comprising:

[0071] S1. Obtain multidimensional clinical trial data and corresponding traceability data for the entire cycle of the same clinical trial, standardize the multidimensional clinical trial data, and generate a standardized record set;

[0072] S2. Construct an evidence layer diagram based on a standardized record set to represent the monitored objects and their relationships;

[0073] S3. Construct a source traceability layer graph based on the source traceability data to characterize the data source and data change trajectory, and associate it with the evidence layer graph to generate a two-layer evidence graph containing the evidence layer graph and the source traceability layer graph. Calculate the source credibility based on the source traceability features corresponding to the nodes and edges of the evidence layer graph in the source traceability layer graph, and assign the source credibility to at least one of the nodes and edges of the corresponding evidence layer graph in the two-layer evidence graph to generate a two-layer evidence graph with source credibility.

[0074] S4. Input the two-layer evidence graph with source credibility into the heterogeneous graph Transformer network for risk reasoning. In the attention calculation of the heterogeneous graph Transformer network, source credibility is introduced as a gating factor to adjust the contribution of nodes and edges of the evidence layer graph to the attention weight, and output the risk score for the monitored object.

[0075] S5. Perform confidence calibration on the risk score, divide the calibrated risk score into multiple groups according to the preset grouping rules, determine the warning threshold for each group based on the preset target false alarm rate, perform conformal prediction output on the calibrated risk score, and generate a risk warning result set including warning labels and risk levels.

[0076] S6. For the monitoring objects that are marked as triggered by the risk warning results, extract the minimum set of nodes and the minimum set of edges that can support the warning mark and the risk level based on the two-layer evidence graph with source credibility, and sort the minimum set of nodes and the minimum set of edges based on time information to generate the minimum sufficient evidence chain and output the interpretation result set.

[0077] S7. Obtain and interpret the monitoring feedback data corresponding to the result set, including the early warning confirmation result and the early warning handling result. Under the condition of freezing the heterogeneous graph Transformer network weight parameters, update at least one of the calculation parameters of source credibility, the calibration parameters of confidence calibration, and the group early warning threshold based on the monitoring feedback data to obtain the updated parameter set. In subsequent monitoring cycles, the updated parameter set is used to perform parameterized execution of steps S3 to S5 to form a closed loop of intelligent monitoring throughout the entire clinical trial cycle.

[0078] In this specific embodiment, S1 includes:

[0079] The system retrieves full-cycle multidimensional clinical trial data from the electronic data acquisition system, laboratory testing system, imaging system, and research center on-site recording system of the same clinical trial through a pre-set acquisition interface. Each multidimensional clinical trial data is bound to its corresponding traceability data one by one using the source record identifier as the primary key. The multidimensional clinical trial data includes at least the subject identifier, visit identifier, event identifier, data item name, data item value, unit, event occurrence time, and record generation time in each record. The traceability data includes at least the data source identifier, source record identifier, data entry time, data version information, and data modification record in each record.

[0080] The system constructs a structured object for each bound record in memory and performs standardization processing. The standardization processing consists of four sub-processes in sequence: field mapping, unit unification, encoding unification, and time alignment.

[0081] Field mapping is performed based on a preset data dictionary. The preset data dictionary uses the data source identifier and the original data item name as the joint search key and returns a unified field name, field data type, standardized unit identifier, and encoding field identifier. The system replaces the original data item name with the unified field name and formats the data item value according to the field data type. The field data types include numeric, categorical, boolean, date / time, and string types, and correspond to floating-point parsing, code table retrieval, truth value normalization, timestamp parsing, and string standardization, respectively.

[0082] The unit conversion is based on a preset unit conversion table. The preset unit conversion table uses a unified field name and the original unit as the search key and returns the standard unit identifier, the scaling factor and the translation factor. The system converts the data item value from the original unit to the standard unit and writes the unit field to the standard unit identifier. The scaling factor is used to perform multiplicative scaling on the value of the numerical data item and the translation factor is used to perform additive translation on units with zero-point offset.

[0083] The encoding is uniformly executed based on preset encoding rules. The preset encoding rules are indexed by encoding field identifiers and correspond to a set of controlled term code tables. After the system performs whitespace removal, case normalization, and full / half-width normalization on the values ​​of categorical and Boolean data items, it retrieves the uniform encoding value from the corresponding controlled term code table and writes the uniform encoding value into the data item value field. At the same time, the original data item value is retained for audit traceability. The values ​​of date and time data items are uniformly converted into timestamps in Coordinated Universal Time after parsing and written into the data item value field.

[0084] Time alignment is performed based on a preset time benchmark, which is fixed in the system parameters as a benchmark time with the subject identifier as the granularity. The benchmark time is determined by the randomized event time corresponding to the subject and stored in the form of a timestamp. The system parses the event occurrence time and record generation time into timestamps respectively, and then subtracts them from the benchmark time to obtain the relative time. Records with missing event occurrence times are then filled in according to the record generation time to obtain the aligned event time. The aligned event time satisfies the formula:

[0085] ;

[0086] in This indicates the aligned event time in seconds. This indicates the event time timestamp used for alignment, which is either parsed from the event occurrence time or, if the event occurrence time is missing, parsed from the record generation time. This represents the baseline time stamp of the subject and is determined by the randomized event time;

[0087] After completing the above four sub-processes, the system summarizes the subject identifier, visit identifier, event identifier, unified field name, data item value after unit unification and coding unification, standardized unit identifier, aligned event time, and source record identifier consistent with the traceability data into a standardized record. It also performs deduplication on duplicate records with the same subject identifier, visit identifier, event identifier, unified field name, and aligned event time, and retains the record generated later. At the same time, records that fail to complete field mapping in the preset data dictionary or fail to complete unit unification and coding unification in the unit conversion table and controlled terminology code table are written into the verification log and removed from the standardized record set, thus generating the standardized record set.

[0088] In this specific embodiment, S2 includes:

[0089] The system takes a standardized record set as input and constructs an evidence layer graph data structure in memory. This evidence layer graph represents the monitored objects and their relationships and is organized in a heterogeneous graph manner. The evidence layer graph follows the formula:

[0090] ;

[0091] in Represents the evidence layer diagram. This represents the set of nodes in the evidence layer graph. Describes the set of edges in the evidence layer graph. This represents a node type mapping function used to map each node to one of the following node types: subject node, visit node, study event node, study medication node, or safety event node. This represents a node attribute mapping function used to map each node to a set of attribute key-value pairs consisting of uniform field names and data item values. This represents a time attribute mapping function used to map each node to an aligned event time.

[0092] The system establishes an index table for the standardized record set and completes node and edge generation in one traversal. The index table uses the subject identifier, visit identifier, and event identifier as the primary index key and the node identifier as the return value. The node identifier of a subject node is uniquely determined by the subject identifier, and the node type of the subject node is determined by… The fixed mapping is to the subject node, the node identifier of the visit node is uniquely determined by the visit identifier, and the node type of the visit node is determined by... The fixed mapping is the visit node. The node identifier of the research event node is uniquely determined by the event identifier, and the node type of the research event node is determined by... The fixed mapping is used to study event nodes;

[0093] The study drug use node and safety event node are generated from records corresponding to study drugs and safety events in a standardized record set. The system stores a mapping table from fields to node types in the configuration file. The mapping table uses a unified field name as the search key and returns a node category identifier. The node category identifier takes a fixed value from either study drug use or safety event. The system reads the unified field name for each standardized record and searches the mapping table. When a study drug use node is returned, the combination of the subject identifier and the event identifier is used as the aggregation key to generate the study drug use node. When a safety event is returned, the combination of the subject identifier and the event identifier is used as the aggregation key to generate the safety event node. This ensures that multiple drug use field records for the same subject under the same event identifier are written to the same study drug use node and multiple safety field records for the same subject under the same event identifier are written to the same safety event node.

[0094] Regarding node attribute assignment, the system uses the unified field name as the attribute key for each standardized record and writes the data item value and the standardized unit identifier as the attribute value into the system. The managed node attribute key-value set, if the same node appears multiple times under the same attribute key, forms a value list according to the order of record arrival, and each value element retains the source record identifier for subsequent cross-level associations. The system writes the aligned event time from the standardized records to... The time attributes of the managed nodes are as follows: the time attribute of the visit node is the minimum value of the aligned event time of all standardized records belonging to the visit identifier to represent the visit start time; the time attributes of the study event node, study medication node and safety event node are the minimum value of the aligned event time of all standardized records under the corresponding event identifier to represent the first occurrence time of the event; the time attribute of the subject node is the minimum value of the aligned event time of all standardized records under the subject identifier to represent the earliest observation time after the subject enters the trial.

[0095] Regarding edge generation, the system obtains the subject node, visit node, and research event node associated with the current record based on the index table when processing each standardized record and writes them into the edge set. When the visit identifier is not empty, an association edge is generated between the subject node and the visit node, and the edge type is fixed as subject to visit. When the event identifier is not empty and the visit identifier is not empty, an association edge is generated between the visit node and the study event node, and the edge type is fixed as visit to study event. When the mapping table of the field to node type returns the study drug, an association edge is generated between the subject node and the study drug node, and the edge type is fixed as subject to study drug. When the mapping table of the field to node type returns the safety event, an association edge is generated between the subject node and the safety event node, and the edge type is fixed as subject to safety event.

[0096] The system performs deduplication on the edge set to ensure that only one edge with the same start node identifier, end node identifier, and edge type is retained. The source record identifier list associated with the edge is written into the edge attribute for subsequent cross-layer binding in the tracing layer, resulting in an evidence layer graph that represents the monitored object and its associated relationships.

[0097] In this specific embodiment, S3 includes:

[0098] The system takes source data and evidence layer diagrams as input and constructs source layer diagrams and cross-layer associations to generate a two-layer evidence diagram;

[0099] The source tracing layer graph adopts a heterogeneous graph data structure and includes three types of nodes: data source nodes, source record nodes, and change record nodes, as well as two types of edges: source association edges and change association edges.

[0100] The system establishes data source nodes based on data source identifiers for the traceability data and writes the data source identifiers into the data source node attributes to form a searchable source index;

[0101] The system then establishes source record nodes for the traceability data according to the source record identifier and writes the source record identifier and data version information into the source record node attributes. The data version information is parsed into an integer version number and stored together with the largest version number that appears under the same source record identifier to support version consistency judgment.

[0102] The system then establishes a change record node for each piece of traceability data and writes the data entry time and data modification record into the change record node attributes. The data modification record is structured as a sequence of change events, and each change event includes the modification time, operator identifier, modified field identifier, value before modification, and value after modification.

[0103] In the source tracing layer graph, the system creates a source association edge pointing to its corresponding data source node for each source record node and writes the source record identifier corresponding to the edge in the edge attribute to ensure reversible retrieval of the source tracing link. At the same time, the system creates a change association edge pointing to its corresponding source record node for each change record node in the source tracing layer graph and writes the change event sequence summary of the change record node in the edge attribute to support subsequent feature calculation.

[0104] In terms of cross-layer association, the system reads the source record identifier list of each node and each edge in the evidence layer graph and retrieves the corresponding source record node in the tracing layer graph for each source record identifier in the list. When the retrieval is successful, a cross-layer association edge is established between the node in the evidence layer graph and the source record node, and a cross-layer association edge is established between the edge in the evidence layer graph and the source record node. When the retrieval fails, the source record identifier is written to the missing tracing log and the source credibility of the corresponding evidence layer graph node or edge is directly set to 0 to block unreliable evidence from entering subsequent reasoning.

[0105] Regarding source credibility calculation, the system calculates source tracing features and generates source credibility for each source record node based on the attributes of its associated change record nodes. These source tracing features are derived from the data entry timeliness feature. Version consistency features With modified stability features Composed of and all fall within interval;

[0106] The timeliness feature of data entry By comparing data entry times The event time aligned with the evidence layer graph node or edge corresponding to the source record identifier. Time difference obtained Determine and assign values ​​according to a fixed mapping, when Seconds ,when Seconds ,when Seconds ,when Seconds ,in This indicates the data entry time in the traceability data and is stored in the form of a timestamp. It represents the aligned event time corresponding to a node or edge in the evidence layer graph and is stored in the form of a timestamp. express The non-negative value;

[0107] Version consistency features The version number is determined based on the relationship between the integer version number of the source record node and the maximum version number under the same source record identifier. When the version number of the source record node equals the maximum version number and the version number sequence under that source record identifier is consecutive, then... When a version is missing but the current version number is still equal to the maximum version number, let When the current version number is less than the maximum version number, ;

[0108] The modified stability feature Based on the data modification record, the length of the change event sequence is modified. Confirmed, when season ,when season ,when season ,when season ,when season ,in This indicates the number of modifications to the same source record identifier, and is directly counted from the length of the change event sequence.

[0109] The system inputs the above three features into a confidence calculation model with fixed parameters to obtain the source confidence of the source record node and writes it into the source record node attributes. The source confidence satisfies the formula:

[0110] ;

[0111] in Indicates the credibility of the source and its value range is This represents an exponential function with the natural constant as its base. This indicates the weight of the timeliness feature to be entered and is fixed in the system parameters. This represents the weight of the version consistency feature and is fixed in the system parameters. This indicates that the stability feature weights have been modified and fixed in the system parameters. Indicates the timeliness characteristics of data entry. Indicates version consistency characteristics, This indicates a modification to the stability trait. This represents the bias term and is fixed at -1.0 in the system parameters;

[0112] The system then backfills the source credibility of the source record node along the cross-layer association edge to the evidence layer graph and writes the source credibility attribute to the evidence layer graph node and the evidence layer graph edge respectively. The backfilling rule is to take the minimum value of the source credibility of all source record nodes associated with the same evidence layer graph node or edge to obtain the source credibility of the evidence layer graph node or edge. This allows any low credibility source to suppress the corresponding evidence from entering the subsequent risk reasoning and obtain a two-layer evidence graph containing the evidence layer graph, the source tracing layer graph and the cross-layer association with the source credibility attribute.

[0113] In this specific embodiment, S4 includes:

[0114] The system takes a two-layer evidence graph with source credibility as input and performs risk reasoning only on the evidence layer graph. The system first encodes the node attributes and edge attributes of the evidence layer graph into node initial feature vectors and edge initial feature vectors, respectively.

[0115] The encoding of the initial feature vector of a node is performed according to a preset feature template. This preset feature template consists of 40 uniform numerical field names and 20 uniform categorical field names, and is permanently stored in the system configuration file. For each node, the system reads the fields matching the preset feature template from its attribute key-value set and constructs the original feature vector. The numerical fields are calculated using the mean and standard deviation obtained through offline statistics. The scores are normalized and zero values ​​are written when missing, and missing indicator bits are written synchronously. The categorized fields use a controlled term code table to map the unified encoded values ​​to integer indices and convert them into fixed-length vectors through a field-specific embedding table. The embedding dimension of the field-specific embedding table is fixed at 8 and unknown indices are fixedly mapped to zero vectors.

[0116] The system maps node types to fixed-length vectors with a fixed embedding dimension of 8 through a node type embedding table. Then, the node type embedding vector, the normalized values ​​of all numerical fields and missing indicator bits, and the embedding vectors of all partial type fields are concatenated in a fixed order to form the original node vector, which is then input into the node encoder to obtain the initial feature vector of the node. The node encoder is a two-layer fully connected network with the first layer having an output dimension of 256 and using the ReLU activation function, and the second layer having an output dimension of 128 and using linear output. The initial feature vector of the node has a fixed dimension of 128.

[0117] The initial feature vector encoding of an edge is based on the edge type and edge attributes to construct the original edge vector. The edge attributes include the difference between the length of the source record identifier list and the time attributes of the nodes at both ends of the edge. The system maps the edge type to an 8-dimensional vector through an edge type embedding table and truncates the difference between the source record identifier list length and the time attribute, then normalizes it to the maximum value. The interval values ​​are concatenated with the edge type embedding vector to form the original edge vector, which is then input into the edge encoder to obtain the initial feature vector of the edge. The edge encoder is a single-layer fully connected network with a fixed output dimension of 64 and uses linear output.

[0118] The system inputs the initial feature vectors of nodes and edges into a heterogeneous graph Transformer network and performs a 4-layer iterative update. Each layer contains a multi-head attention sublayer and a feedforward network sublayer, both with residual connections and LayerNorm normalization. The number of heads in the multi-head attention sublayer is fixed at 8, and the dimensions of the query vector, key vector, and value vector of each head are fixed at 32. The system sets independent linear transformation parameters for different node types in the evidence layer graph to generate the query vector, key vector, and value vector of each attention head. The system sets independent edge transformation parameters for different edge types in the evidence layer graph to map the initial feature vector of the edge to a 32-dimensional edge modulation vector and sums the edge modulation vector to the key vector and value vector of the corresponding neighbor node to achieve edge information injection.

[0119] The system incorporates source credibility as a gating factor in attention calculation, with gating weights... From the target node Source credibility, neighboring nodes The minimum value among the source credibility of the two and the source credibility of the edge between them is obtained, and the value range is [value range missing]. This is used to adjust the attention contribution of neighboring nodes to the target node, and the attention weights satisfy the formula:

[0120] ;

[0121] in Indicates the first Under the attention of the target node Assigned to neighboring nodes Attention weights Indicates at node The index on the set of all adjacent nodes The Softmax function is used for normalization. Indicates the gating weight, Represents the target node In the The query vector under each attention head Indicates adjacent nodes In the The key vector under each attention head This represents the transpose of a vector. This represents the vector dimension of each attention head, which is fixed at 32. Indicates to The scaling factor obtained by taking the square root Indicates the target node index. Indicates the index of adjacent nodes. Indicates the attention head index;

[0122] The system calculates the aggregate vector of the target node by weighted summation of the value vectors of adjacent nodes based on attention weights, and outputs the updated node representation vector of the target node through linear transformation. The feedforward sublayer adopts a two-layer fully connected network with a fixed intermediate layer dimension of 512 and ReLU activation function, and a fixed output layer dimension of 256 and linear output. The system enables Dropout at the output of each attention sublayer and the output of each feedforward sublayer with a fixed dropout rate of 0.2, and keeps it enabled during the inference phase to support uncertainty estimation.

[0123] After completing four iterations, the system obtains the node representation vector corresponding to the monitored object and inputs it into the risk prediction layer to obtain the initial risk score. The risk prediction layer is a two-layer fully connected network with the first layer having an input dimension of 256 and an output dimension of 128, using the ReLU activation function, and the second layer having an output dimension of 1, mapping the output to a sigmoid function. The interval is used as the initial risk score;

[0124] With Dropout enabled, the system performs 20 forward calculations on the same monitored object to obtain 20 initial risk scores, uses the arithmetic mean of these scores as the initial risk score for the monitored object, and uses the standard deviation of these scores as the prediction uncertainty.

[0125] The system reads the experimental phase information and research center information corresponding to the monitored object from the two-layer evidence map with source credibility, and generates a group identifier according to a fixed splicing rule. The system outputs an initial risk result set, and each record in the initial risk result set contains the monitored object identifier, initial risk score, prediction uncertainty and group identifier.

[0126] In this specific embodiment, S5 includes:

[0127] The system takes the initial risk result set and the pre-acquired monitoring feedback data as input and generates a risk warning result set;

[0128] The system filters samples with early warning confirmation results from monitoring feedback data and matches them one-to-one with the historical initial risk result set according to the monitored object identifier to form a calibration sample set. Each calibration sample includes an initial risk score. Grouping identifiers And the binary label determined by the early warning confirmation result. , and when This indicates that the monitored object has been confirmed to pose a risk. This indicates that the monitored object has been confirmed to pose no risk.

[0129] The system uses a set of calibration samples. A confidence calibration model is trained and a calibration function is obtained. The confidence calibration model employs one-dimensional logistic regression calibration and contains only learnable parameters. and and the initial risk score Mapped to calibrated risk score To improve ratings and tags For consistency, the calibration function satisfies the formula:

[0130] ;

[0131] in This indicates the risk score after calibration and its range is [value range missing]. This represents the initial risk score and its range is [value range missing]. This indicates the calibration slope parameter and its initial value is set to... This indicates the calibration bias parameter and its initial value is set to... Represents an exponential function with the natural constant as its base;

[0132] The system uses binary cross-entropy as the loss function and adjusts the parameters. and Perform batch gradient descent training, with a learning rate during training. The value is fixed at 0.01, the number of iterations is fixed at 500, and the... and The application coefficient is of Regular terms and Fixed as After training, a definitive result is obtained. and And solidify it into calibration parameters;

[0133] The system applies the calibration function to each record in the current initial risk result set to generate a calibrated risk result set. Each record in the calibrated risk result set includes the monitored object identifier and the calibrated risk score. Prediction uncertainty and grouping identifier ;

[0134] The system is based on group identifiers The calibrated risk result set is divided into multiple calibration groups, and conformal prediction is performed within each calibration group to determine the group warning threshold.

[0135] Specifically, the system in each calibration group From the calibration sample set Filter out tags The samples form a negative class calibration subset. And the calibrated risk score for each sample within the negative class calibration subset. The system will directly define the non-consistency score of the sample and then calibrate the negative class subset. All inconsistency scores are sorted in ascending order and based on the target false alarm rate. Select the sorted sequence number as The non-consistency score, rounded up, is used as the group warning threshold. ,in This represents the preset target false alarm rate, which is fixed in the system parameters. Represents a negative class calibration subset The number of samples, This indicates the grouped early warning threshold and its value range is [value range missing]. ;

[0136] The system will group warning thresholds The risk results set after calibration belong to the corresponding group identifier. The monitored object is identified and an early warning marker is generated. A risk score is then calculated after the monitored object is calibrated. satisfy The warning flag will be set to triggered if the warning is triggered, otherwise it will be set to not triggered.

[0137] The system further scores the calibrated risk based on preset risk grading thresholds. Risk levels are classified, and the preset risk classification threshold is fixed in the system parameters. and and will Corresponding to low risk Corresponding to medium risk Corresponding to high risk This corresponds to a serious risk;

[0138] Output a risk warning result set, and each result includes the monitoring object identifier, the calibrated risk score, the group identifier, the group warning threshold, the warning mark and the risk level.

[0139] In this specific embodiment, S6 includes:

[0140] The system takes the risk warning result set as input and filters the monitoring objects triggered by the warning markers to form a set of monitoring objects to be explained. For each monitoring object to be explained, the system locates the central node of the evidence layer graph that matches the identifier of the monitoring object to be explained from the two-layer evidence graph with source credibility, and reads the initial risk score corresponding to the central node in the initial risk result set as the benchmark score. At the same time, the heterogeneous graph Transformer network is placed in the explanation reasoning mode to ensure that the score can be reproduced. The explanation reasoning mode has a fixed random seed. The 20 forward computations from step S4 are used again, and Dropout is enabled to generate a random mask sequence consistent with the benchmark score, making the subsequent score comparison between the candidate subgraph and the minimum subgraph deterministic.

[0141] Starting from the center node of the evidence layer graph, candidate subgraphs within a preset number of hops are extracted. The preset number of hops is fixed at 2 hops, and the candidate subgraphs only contain all nodes and all edges in the evidence layer graph whose shortest path length to the center node is no greater than 2. At the same time, the system retains the cross-layer associations corresponding to each node and each edge in the candidate subgraphs so as to trace back the data source identifier and time information when interpreting the output.

[0142] The system calculates the contribution of each node and edge in the candidate subgraph to the baseline score and uses the contribution to drive the gradual elimination to obtain the minimum set of nodes and the minimum set of edges. The contribution calculation is implemented using the occlusion score difference method. For any node in the candidate subgraph, the system sets the node's initial feature vector to zero and the node's source credibility attribute to 0 to form an occlusion version of the subgraph. In the interpretation reasoning mode, risk reasoning is performed on the central node to obtain the occlusion score. The system defines the absolute value of the difference between the baseline score and the occlusion score as the contribution of the node.

[0143] For any edge in the candidate subgraph, the system removes the edge from the adjacency relationship of the evidence layer graph and sets the source credibility attribute of the edge to 0 to form an edge-masked version of the subgraph. In the interpretation reasoning mode, risk reasoning is performed on the central node to obtain the masking score. The system defines the absolute value of the difference between the baseline score and the masking score as the contribution of the edge.

[0144] After obtaining the contribution values ​​of all nodes and all edges, the system merges the nodes and edges in the candidate subgraph into a sequence to be eliminated and sorts them by contribution value from smallest to largest. Then, it performs step-by-step elimination in the sorted order, and only tries to eliminate one node or one edge at a time. When eliminating a node, it simultaneously eliminates the edges connected to that node that are missing at either end to maintain the effective adjacency structure of the subgraph. When eliminating an edge, it keeps the nodes at both ends unchanged.

[0145] After each attempt to remove a node, the system performs risk reasoning on the central node based on the current subgraph after removal to obtain the current score and determine whether the tolerance constraint is satisfied. The tolerance constraint satisfies the formula:

[0146] ;

[0147] in This represents the risk score of the central node output by the heterogeneous graph Transformer network on the current subgraph. This represents the baseline score corresponding to the central node in the initial risk result set. This indicates a preset tolerance, which is fixed in the system parameters. This represents the absolute value operation;

[0148] When the tolerance constraint is met, the system confirms the removal and updates the current subgraph to the new baseline subgraph. When the tolerance constraint is not met, the system cancels the removal and retains the node or edge that was attempted to be removed. After traversing the entire sequence to be removed, the system obtains the final minimum set of nodes and the minimum set of edges to be retained.

[0149] Based on time information, the minimum set of nodes and the minimum set of edges are sorted to generate the minimum sufficient evidence chain. The time information of a node is taken from the time attribute of the node, that is, the aligned event time. The time information of an edge is taken from the larger value of the time attribute of the two nodes of the edge to represent the latest time when the association is established. The system sorts the nodes from early to late according to the time information. In the case of the same time, it performs a stable sorting according to the node type order: subject node, visit node, research event node, research drug node, and safety event node. The edge is inserted between the two nodes it connects.

[0150] When outputting the interpretation result set, the system retrieves the corresponding source record node for each node and each edge in the minimum sufficient evidence chain and further retrieves the data source node to obtain the data source identifier. At the same time, it summarizes all source record identifiers associated with the node or edge and writes them into the interpretation result for audit traceability, so that each record in the interpretation result set contains the monitoring object identifier, the minimum sufficient evidence chain, and the data source identifier and time information corresponding to each piece of evidence in the minimum sufficient evidence chain.

[0151] In this specific embodiment, S7 includes:

[0152] The system acquires and interprets the monitoring feedback data corresponding to each record in the result set and forms a feedback record table. Each record of the monitoring feedback data includes at least the monitoring object identifier, the warning confirmation result, and the warning handling result. The warning confirmation result is a binary variable with a value of [value missing]. ,when When the warning is confirmed to be a true positive, When the warning is confirmed to be a false alarm, the warning handling result is used to identify the type of handling action that has been performed and is written as an audit field into the parameter update log, but does not change the parameter update calculation path of this step;

[0153] The system puts the heterogeneous graph Transformer network into a frozen state and calculates a hash digest of all its weight parameters and writes it into the frozen list. The frozen state means that in this step, gradient calculation and parameter assignment are not performed on any weight parameters of the heterogeneous graph Transformer network and only forward calculation is allowed.

[0154] The system updates the calculation parameters of source credibility based on the feedback record table and maintains the source credibility calculation model structure consistent with step S3. The system denotes the parameter vector of the source credibility calculation model as follows: The traceability features are denoted as ,in This indicates the weight of the timeliness feature during data entry. Indicates the weight of version consistency features. This indicates that the stability feature weights have been modified. This represents the bias term in the source credibility calculation model. This indicates the timeliness of data entry and is obtained by mapping the time difference between the data entry time and the aligned event time. This indicates version consistency and is determined by version number continuity and whether the current version is the largest version. This indicates that the stability feature has been modified and is determined by the number of modifications; the constant 1 is used to enable the bias term to participate in the update.

[0155] For each feedback record, the system relocates the cross-layer association of the monitored object in the two-layer evidence graph and collects all associated source record nodes. For each source record node, it reads the source tracing features calculated in step S3. And construct feature vectors and the parameter vector Perform deterministic increment / decrement updates to reduce the weight of false positives and increase the weight of true positives, wherein the updates satisfy the formula:

[0156] ;

[0157] in This represents the vector of source credibility calculation parameters to be updated. ← indicates an assignment / update operation. This means clipping each dimension of the vector to a closed interval. The clipping operator is used to limit the parameter range. This indicates the update step size, which is fixed in the system parameters. This indicates the confirmation result of the warning recorded in the feedback log. This represents the source feature vector corresponding to the source record node. Will Mapped to +1 and The mapping is set to -1, thereby increasing the parameter for true positives and decreasing the parameter for false positives;

[0158] After processing all records in the feedback log table, the system uses the updated version. In subsequent monitoring cycles, the source credibility of step S3 is recalculated and backfilled into the nodes and edges of the evidence layer graph to complete the credibility update loop.

[0159] The system synchronously updates the calibration parameters for confidence level calibration and maintains the calibration model as a one-dimensional logistic regression calibration. The system records the calibration slope parameter as... The calibration bias parameter is denoted as And using the monitored objects with early warning confirmation results in the feedback record table as incremental calibration samples. and Batch gradient descent retraining is performed, with a fixed learning rate of 0.005 and 200 iterations, using binary cross-entropy loss and applying a coefficient of... of The regular expression term yields the updated version. and And in subsequent monitoring cycles, it is used to convert the initial risk score into a calibrated risk score;

[0160] The system synchronously updates the group warning threshold and maintains the group shape-preserving prediction process. The system uses the group identifier as the key to summarize all negative class confirmation samples in the most recent monitoring period, which satisfies the requirement. The samples were sorted in ascending order of calibrated risk score within each group and based on the target false alarm rate. Reassess the group's alert threshold And will update Used for determining early warning triggers in subsequent monitoring cycles;

[0161] The system will update the source credibility calculation parameters. Updated calibration parameters and the updated warning thresholds for each group. The update parameter set, along with the hash summary of the frozen list, the update timestamp, and the version number, is written to the update parameter set and persisted. In subsequent monitoring cycles, this update parameter set is called to perform parameterized execution of steps S3 to S5 to form a closed loop of intelligent monitoring throughout the entire clinical trial cycle.

[0162] In this specific embodiment, when the system constructs the traceability layer graph in step S3, it adds data version integrity verification information to each source record node and persists it as an attribute field of the source record node. The data version integrity verification information includes the hash value of each version of the source record and the hash chain of adjacent versions.

[0163] The system uses the source record identifier as the aggregation key to extract all data version information under the same source record identifier from the traceability data, and then parses the version number into an incrementing integer sequence and sorts them in ascending order of version number.

[0164] The system performs normalized serialization on the source record content of each version to ensure the determinism of hash calculation. The normalized serialization includes expanding the source record content into a key-value pair structure, sorting the key names in lexicographical order, performing UTF-8 encoding and removing invisible characters on string fields, performing fixed decimal place representation on numeric fields, uniformly converting time fields into timestamp representation, and concatenating the serialization results into a single byte sequence.

[0165] The system performs the SHA-256 algorithm on the byte sequence to calculate the version hash value and writes it into the source record node attribute. Simultaneously, the system constructs adjacent version hash chains and writes them into the source record node attribute. The adjacent version hash chains are for versions with the following version numbers: With version number Establish a one-to-one correspondence between adjacent versions and store their version hash values ​​for verification;

[0166] The system performs integrity checks at the start of each source traceability layer update or monitoring cycle. Integrity checks include recalculating the version hash value for each version and comparing it with the version hash value stored in the source record node attributes, and checking whether the version hash value pair stored in the source record node attributes for each pair of adjacent versions is consistent with the recalculated version hash value pair. When there is any inconsistency in the version hash value, any inconsistency in the adjacent version hash value pair, discontinuous version number sequence, or missing adjacent version relationship, the source record node is marked as having an inconsistent hash chain and the inconsistency location, the version number involved, and the recalculated hash value are recorded in the audit log.

[0167] After detecting an inconsistency in the hash chain, the system performs a source credibility penalty update on all evidence layer graph nodes and edges that are cross-layered with the source record node, following the cross-layer associations between the source record node and the evidence layer graph nodes and edges. This penalty update updates the original source credibility of the evidence layer graph nodes or edges. Adjusted to source credibility after punishment And satisfy the formula:

[0168] ;

[0169] in This indicates the source credibility after penalty, and its value range is [value missing]. This indicates that step S3 calculates and backfills the original source confidence of the nodes or edges in the evidence layer graph, and the value range is [value range missing]. This represents the integrity penalty coefficient, which is fixed in the system parameters. This variable indicates hash chain inconsistency and takes a value of 1 when any source record node associated with this evidence layer graph node or edge is marked as hash chain inconsistency; otherwise, it takes a value of 0.

[0170] The system will penalize the source credibility. The source credibility attribute of the corresponding evidence layer graph node or edge is overwritten and kept from being overwritten by a higher credibility value within the same monitoring period. This ensures that the evidence corresponding to inconsistent source records in the hash chain is continuously suppressed in the subsequent heterogeneous graph Transformer attention gating and reduces risky reasoning false alarms caused by version tampering or version missing.

[0171] 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.

[0172] This invention addresses the technical problem of "achieving low-false-report and interpretable risk warning grading and forming a monitoring closed loop under multi-dimensional evidence fusion conditions." It standardizes multi-dimensional data from the entire clinical trial cycle to construct an evidence layer graph to depict the structural relationships between subjects, visits, events, investigational drugs, and safety events. Simultaneously, it constructs a source tracing layer graph to depict data sources and data change trajectories, achieving a one-to-one correspondence between evidence and source through cross-layer association. During the risk reasoning stage, source credibility is calculated based on source tracing features and used as a gating factor in the attention calculation of the heterogeneous graph Transformer to adjust the contribution of evidence layer nodes and edges to attention weights. This strengthens high-credibility evidence and suppresses low-credibility evidence, reducing false alarms caused by noisy evidence at the source. During the early warning output phase, the confidence level of the risk score is calibrated to improve the consistency between the score and the actual risk. Conformal prediction is implemented according to the experimental stage and research center groups. The early warning threshold for each group is determined based on the target false alarm rate, thereby achieving early warning triggering and risk classification with a controllable false alarm rate from a mechanism perspective. During the interpretation phase, the minimum set of nodes and the minimum set of edges that meet the tolerance constraints are extracted and sorted by time to form the minimum sufficient evidence chain. At the same time, the corresponding source identifier and time information are output, so that the early warning results have auditable and traceable interpretability, thereby supporting monitoring and handling and compliance review.

[0173] This invention improves the algorithm structure to address the aforementioned technical problems in the following ways: First, it proposes a two-layer evidence graph structure containing an evidence layer and a source tracing layer, enabling source credibility to be bound to specific evidence nodes and edges across layers. This avoids coarse-grained data quality processing only at the feature level, thus more effectively suppressing unreliable evidence introduced by factors such as entry timeliness, version changes, and frequent modifications. Second, it introduces source credibility-gated attention into the heterogeneous graph Transformer, applying credibility to the message passing and aggregation process to achieve a strongly coupled evidence weighting mechanism with graph reasoning, improving the stability of risk identification in cross-object association scenarios. Third, it adopts grouped shape-preserving prediction and determines the threshold with the target false alarm rate, solving the problem of threshold failure caused by distribution differences between different centers and different stages. Fourth, under the condition of freezing the main model weight parameters, it only updates the source credibility calculation parameters, confidence calibration parameters, and grouping thresholds, achieving safe and controllable closed-loop optimization, reducing the model drift risk caused by online updates, and continuously improving the technical effect of low false alarms and interpretable monitoring.

Claims

1. A clinical trial full-cycle intelligent monitoring method based on multi-dimensional data fusion, comprising: S1, obtaining multi-dimensional clinical trial data and corresponding traceability data of the same clinical trial full-cycle, standardizing the multi-dimensional clinical trial data to generate a standardized record set; S2, constructing an evidence layer graph for representing a monitoring object and an association relationship based on the standardized record set; S3, constructing a traceability layer graph for representing data sources and data change trajectories based on the traceability data, and associating the traceability layer graph with the evidence layer graph to generate a double-layer evidence graph containing the evidence layer graph and the traceability layer graph, calculating a source credibility based on traceability features corresponding to nodes and edges of the evidence layer graph in the traceability layer graph, and assigning the source credibility to at least one of the nodes and edges of the evidence layer graph in the double-layer evidence graph to generate a double-layer evidence graph with source credibility; S4, inputting the double-layer evidence graph with source credibility into a heterogeneous graph Transformer network for risk reasoning, introducing the source credibility as a gating factor in attention calculation of the heterogeneous graph Transformer network to adjust the contribution of nodes and edges of the evidence layer graph to attention weights, and outputting a risk score for the monitoring object; S5, calibrating the risk score for confidence, dividing the calibrated risk score into multiple groups according to a preset grouping rule, and determining a group warning threshold for each group according to a preset target false alarm rate, performing a conformal prediction output on the calibrated risk score to generate a risk warning result set including a warning mark and a risk level; S6, for the monitoring object with the warning mark triggered in the risk warning result set, extracting a minimum node set and a minimum edge set capable of supporting the warning mark and the risk level based on the double-layer evidence graph with source credibility, and sorting the minimum node set and the minimum edge set based on time information to generate a minimum sufficient evidence chain, and outputting an explanation result set; S7, obtaining monitoring feedback data corresponding to the explanation result set, including warning confirmation results and warning disposal results, updating at least one of calculation parameters of the source credibility, calibration parameters of the confidence calibration, and the group warning threshold based on the monitoring feedback data under the condition of freezing the weight parameters of the heterogeneous graph Transformer network to obtain an update parameter set, and in the subsequent monitoring cycle, parameterizing steps S3 to S5 using the update parameter set to form a clinical trial full-cycle intelligent monitoring closed loop.

2. The clinical trial full-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S1 comprising: obtaining multi-dimensional clinical trial data of the same clinical trial full-cycle, the multi-dimensional clinical trial data including subject identification, visit identification, event identification, data item name, data item value, unit, event occurrence time, and record generation time; obtaining traceability data corresponding to the multi-dimensional clinical trial data piece by piece, the traceability data including data source identification, source record identification, data entry time, data version information, and data modification record; performing field mapping on the multi-dimensional clinical trial data based on a preset data dictionary to generate data items with uniform field names, performing unit unification on the data item values based on a preset unit conversion table, performing coding unification on the data item names and the data item values based on a preset coding rule; performing time alignment on the event occurrence time and the record generation time based on a preset time reference to obtain aligned event time; and aggregating the results after performing the field mapping, unit unification, coding unification and time alignment together with the subject identifier, visit identifier, event identifier and source record identifier consistent with the traceability data to generate the standardized record set.

3. The clinical trial full-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S2 comprising: determining a monitoring object node set and constructing the evidence layer graph based on the standardized record set, wherein the monitoring object node set includes a subject node with a subject identifier as a node identifier, a visit node with a visit identifier as a node identifier, a research event node with an event identifier as a node identifier, and a research drug node and a safety event node generated by the records in the standardized record set corresponding to research drug and safety events; determining an association relationship edge set and constructing the evidence layer graph based on the standardized record set, wherein the association relationship edge set includes an association relationship edge between the subject node and the visit node, an association relationship edge between the visit node and the research event node, an association relationship edge between the subject node and the research drug node, and an association relationship edge between the subject node and the safety event node; and taking the aligned event time in the standardized record set as the time attribute of the corresponding node, taking the data items with uniform field names and their data item values in the standardized record set as the attributes of the corresponding node, to generate an evidence layer graph for representing the monitoring objects and their association relationships.

4. The clinical trial full-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S3 comprising: constructing a traceability layer graph based on the traceability data, wherein the traceability layer graph includes a data source node, a source record node and a change record node, the data source node is used to represent the data source identifier, the source record node is used to represent the source record identifier and the data version information, the change record node is used to represent the data entry time and the data modification record, and the source association edge between the data source node and the source record node and the change association edge between the source record node and the change record node are established in the traceability layer graph; establishing cross-layer association between the source record node in the traceability layer graph and the corresponding node and edge in the evidence layer graph generated by the standardized record set to generate a double-layer evidence graph based on the source record identifier in the standardized record set. Based on the provenance features associated with the corresponding nodes and edges of the evidence layer graph in the provenance layer graph, the source credibility is calculated, the provenance features including data entry time, data version information and data modification record, and the calculated source credibility is assigned to the corresponding nodes and edges of the evidence layer graph in the double-layer evidence graph to generate the double-layer evidence graph with source credibility.

5. The clinical trial whole-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S4 comprising: encoding the node attributes and edge attributes of the evidence layer graph in the double-layer evidence graph with source credibility into node initial feature vectors and edge initial feature vectors, and inputting the node initial feature vectors and the edge initial feature vectors into the heterogeneous graph Transformer network; in the heterogeneous graph Transformer network, performing linear transformation on different node types and different edge types of the evidence layer graph to obtain query vectors, key vectors and value vectors, calculating attention scores based on the query vectors and the key vectors, and generating gating weights according to the source credibility corresponding to the nodes and edges participating in the attention calculation, and performing normalization processing on the attention scores after modulation by the gating weights to obtain attention weights; based on the attention weights, weighting and aggregating the value vectors of adjacent nodes to update the node representation vector of the target node, and obtaining the node representation vector of the monitoring object through iteration update of a preset number of layers; inputting the node representation vector into the risk prediction layer to obtain the initial risk score, and obtaining multiple initial risk scores by performing multiple forward calculations on the same monitoring object, and calculating the dispersion of the multiple initial risk scores to obtain the prediction uncertainty; based on the trial stage information and the research center information corresponding to the monitoring object in the double-layer evidence graph with source credibility, determining the grouping identifier, and generating the initial risk result set including the monitoring object identifier, the initial risk score, the prediction uncertainty and the grouping identifier.

6. The clinical trial whole-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S5 comprising: determining a calibration sample set based on pre-acquired monitoring feedback data, the calibration sample set being a sample corresponding to a historical initial risk result set and having a pre-warning confirmation result; performing confidence calibration on the initial risk score in the initial risk result set to obtain a calibration function, and converting the initial risk score into a calibrated risk score using the calibration function, and generating a calibrated risk result set including the monitoring object identifier, the calibrated risk score, the prediction uncertainty and the grouping identifier; dividing the calibrated risk result set into multiple calibration groups according to the grouping identifier; for each calibration group, determining a non-consistency score based on the calibrated risk score of the sample belonging to the calibration group in the calibration sample set, and determining a group pre-warning threshold from the non-consistency score according to a preset target false positive rate; The grouping early warning threshold is applied to the monitoring objects in the calibrated risk result set that belong to the corresponding calibration grouping, an early warning mark is generated when the non-consistency score corresponding to the monitoring object meets the early warning triggering condition, and the risk level is determined according to the comparison result of the calibrated risk score and the preset risk classification threshold, and the risk early warning result set is generated.

7. The clinical trial whole-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S6 comprises: screening the monitoring objects with triggered early warning marks from the risk early warning result set to determine a set of monitoring objects to be explained; for each monitoring object to be explained, extracting a candidate subgraph centered on the evidence layer graph node corresponding to the monitoring object to be explained and within a preset hop range based on the double-layer evidence graph with source credibility; calculating the contribution of the nodes and edges in the candidate subgraph to the initial risk score, and performing step-by-step elimination in the candidate subgraph based on the contribution to obtain a minimum node set and a minimum edge set that meet the following condition: the difference between the initial risk score obtained by risk inference on the monitoring object to be explained using a heterogeneous graph Transformer network and the initial risk score corresponding to the monitoring object to be explained in the initial risk result set does not exceed a preset tolerance while the minimum node set and the minimum edge set are retained; sorting the aligned event times corresponding to the nodes and edges in the minimum node set and the minimum edge set to generate the minimum sufficient evidence chain; outputting the explanation result set, which includes monitoring object identifiers, minimum sufficient evidence chains, and data source identifiers and time information corresponding to each evidence in the minimum sufficient evidence chain.

8. The clinical trial whole-cycle intelligent monitoring method based on multi-dimensional data fusion according to claim 1, S7 comprises: obtaining monitoring feedback data corresponding to each of the explanation result sets, the monitoring feedback data including monitoring object identifiers, early warning confirmation results, and early warning disposal results; updating the calculation parameters of source credibility based on the monitoring feedback data, the updated calculation parameters of source credibility including weight parameters of traceability features, wherein the weight parameter of the traceability feature corresponding to a false positive is reduced when the early warning confirmation result indicates a false positive, and the weight parameter of the traceability feature corresponding to a true positive is increased when the early warning confirmation result indicates a true positive; updating the calibration parameters of confidence calibration based on the monitoring feedback data, the updated calibration parameters of confidence calibration being used to improve the matching degree between the calibrated risk score and the early warning confirmation result; updating the grouping early warning threshold based on the monitoring feedback data, the updated grouping early warning threshold including the re-estimation of the grouping early warning threshold in each calibration grouping according to a preset target false positive rate; Under the condition of freezing the weight parameters of the heterogeneous graph Transformer network, the updated calculation parameters of the source credibility, the calibration parameters of the confidence calibration, and the grouping early warning threshold are summarized to generate the updated parameter set, and the updated parameter set is called in the subsequent monitoring period to perform steps S3 to S5. 9.The method of claim 4, wherein, The provenance layer graph further includes data version integrity check information for the source record node, the integrity check information including a hash value generated for each version source record and a neighboring version hash chain, and when a hash chain inconsistency is detected, the source credibility of the evidence layer graph node and edge associated across layers with the source record node is reduced.

10. A multi-dimensional data fusion-based intelligent monitoring system for the whole cycle of a clinical trial, configured to perform any one of the multi-dimensional data fusion-based intelligent monitoring methods for the whole cycle of a clinical trial according to claims 1 to 9, comprising: a data standardization module configured to obtain multi-dimensional clinical trial data and corresponding provenance data for the whole cycle of a clinical trial, and to perform standardization processing on the multi-dimensional clinical trial data to generate a standardized record set; a double-layer evidence graph generation module configured to construct an evidence layer graph based on the standardized record set, construct a provenance layer graph based on the provenance data, and establish cross-layer association to generate a double-layer evidence graph, calculate source credibility based on the provenance layer graph, and assign the source credibility to at least one of the nodes and edges of the evidence layer graph to obtain a double-layer evidence graph with source credibility; a risk reasoning module configured to input the double-layer evidence graph with source credibility into a heterogeneous graph Transformer network, use the source credibility as a gating factor to adjust attention weight in attention calculation, and output a risk score of the monitoring object; an early warning output module configured to perform confidence calibration on the risk score and grouping according to a preset rule, determine an early warning threshold for each group according to a preset target false alarm rate and perform conformal prediction, and output a risk early warning result including an early warning marker and a risk level; an explanation module configured to extract a minimum node set and a minimum edge set from the double-layer evidence graph for a monitoring object triggering early warning, sort the minimum node set and the minimum edge set by time to generate a minimum sufficient evidence chain, and output an explanation result; a closed-loop update module configured to obtain monitoring feedback data, update at least one of the source credibility parameters, the calibration parameters, and the grouping early warning threshold under the condition of freezing the weight parameters of the heterogeneous graph Transformer network, and call the updated parameters to perform monitoring in the subsequent monitoring period.

Images