Clinical trial data verification method and system based on modular architecture

Abstract

The present invention proposes a clinical trial data verification method and system based on a modular architecture, the method including: obtaining a target clinical trial protocol document and mapping it into a term mapping table; generating an execution acquisition map and path confidence based on the protocol structure map and in combination with actual trial acquisition data; comparing the protocol map with the execution acquisition map, identifying the structural deviation of the acquisition path through structural alignment verification, and generating a structural deviation matrix; obtaining the basic fluctuation range of the indicator and the medical standard, and generating a consistent behavior score in combination with the generated structural deviation matrix, which is used to quantify the abnormal behavior trend score of the subject; calculating the causal deviation score of each subject in combination with the protocol structure map, execution acquisition map, structural deviation matrix and consistent behavior score, and marking the missed acquisition, wrong acquisition and timed acquisition events in the clinical trial data according to the causal deviation score.

Description

Technical Field

[0001] The present invention belongs to the field of clinical trial data verification, and in particular relates to a clinical trial data verification method and system based on a modular architecture. Background Art

[0002] Clinical trials are an indispensable part of new drug development, device evaluation, and treatment strategy validation. The quality of their data directly determines the scientific nature, compliance, and regulatory approval rate of the entire trial. As multi-center, cross-regional, and large-scale clinical trials gradually become mainstream, trial data presents characteristics such as complex structure, diverse processes, long collection chains, and dispersed responsible parties, making data verification in clinical trials particularly critical and complex. Currently, clinical trial data verification mainly relies on three traditional methods: first, manual CRF verification, that is, manual verification of collected data and original documents through offline or remote review by CRAs; second, electronic data verification systems based on rule engines, which automatically mark abnormal items by setting field-level verification rules; third, some research units try to use data visualization methods to detect deviations or trend anomalies from a statistical perspective. These methods have played an important role in the historical process of data verification, but they also face increasingly prominent limitations.

[0003] First, with the increasing complexity of clinical trial protocols, traditional data verification methods often focus solely on field values, failing to identify structural deviations from the trial design. For example, whether a particular collection was performed according to the protocol's pre-defined procedures, at the correct time, and with the correct population, makes it difficult for existing systems to systematically restore and verify the consistency between protocol procedures and data execution. Second, regarding data authenticity, current anomaly detection methods typically rely on set thresholds or statistical models (such as mean-variance, z-score, etc.) to identify single-point anomalies. This ignores the temporal order and inherent behavioral patterns of individual continuous records, leading to the inability to identify more subtle data falsification, over-recording, and misrecording. Furthermore, clinical trials are inherently structured around event-response causal mechanisms, such as the expected physiological reactions after drug administration and the recovery indicators to be observed after surgery. These causal mechanisms are not modeled in current data verification systems, making it difficult for the system to detect data that appears numerically normal but may be completely illogical from a medical perspective. Furthermore, existing verification mechanisms lack a flexible modular architecture, requiring extensive manual configuration adjustments when trial protocols are changed, multi-center deployments are implemented, or new indicators are introduced. This lack of system composability and adaptability makes the system inflexible and adaptable. Therefore, building an efficient data verification system that understands trial protocol flows, integrates data behavior patterns with causal logic, and possesses a modular structure has become a critical technical challenge urgently needed in clinical trial data management. Summary of the Invention

[0004] The purpose of the present invention is to design a clinical trial data verification method and system based on a modular architecture, propose a clinical trial data verification method and system based on a modular architecture, establish an extensible and combinable data verification mechanism, and realize intelligent, structured, and multi-dimensional verification of clinical trial data at different levels.

[0005] To achieve the above objectives, the present invention provides a clinical trial data verification method based on a modular architecture in a first aspect, the method comprising:

[0006] Obtain the target clinical trial protocol document, generate a protocol structure diagram based on the time points, collection indicators, and subject populations of the target clinical trial protocol document, and map it into a term mapping table;

[0007] Based on the protocol structure graph and combined with the actual test collection data, an execution collection graph and path confidence are generated;

[0008] Compare the protocol map with the execution acquisition map, identify the structural deviations of the acquisition path through structural alignment verification, and generate a structural deviation matrix;

[0009] Obtaining the basic fluctuation range of the indicator and the medical standard, combining them with the generated structural deviation matrix, and generating a consistent behavior score for quantifying the abnormal behavior trend score of the subject;

[0010] The causal deviation score of each subject is calculated by combining the protocol structure map, execution collection map, structural deviation matrix and consistency behavior score, and the missed collection, wrong collection and timed collection events in the clinical trial data are marked according to the causal deviation score.

[0011] Furthermore, the target clinical trial protocol document is obtained, and a protocol structure diagram is generated according to the time points, collection indicators, and subject populations of the target clinical trial protocol document, and mapped into a term mapping table, specifically including:

[0012] Obtain the target clinical trial protocol document; input the target clinical trial protocol document into the BERT-based model, predict the structure label for each paragraph, and output a structure quadruple, including the time point, collection indicators, test population, and additional conditions in the original protocol that restrict the current collection behavior;

[0013] Convert the time point, collection index, and test population into three nodes and two edges to form a directed path from time point → test population → collection index;

[0014] All nodes were mapped to a standard coding system as a term mapping table; the collection indicators used LOINC; the subject population used the ICF grouping structure; and the time points were converted into a structure of visit days and allowed offset ranges.

[0015] Furthermore, the protocol structure graph is also provided with an edge weight regularization mechanism, specifically including:

[0016] The additional conditions that restrict the current collection behavior in the original agreement extracted from the target clinical trial protocol document are fed into the BERT model to obtain the attention weight distribution of each token, and its entropy value is calculated as the semantic concentration index; the edge weight is constrained according to the semantic concentration index.

[0017] Furthermore, each data of the actual collected data of the experiment includes the actual collection time, the actual indicator name, and the actual subject ID;

[0018] Then there is,

[0019] The generation of the execution acquisition graph and path confidence based on the protocol structure graph and combined with the actual test data collection specifically includes:

[0020] Performing terminology standardization processing on the actual indicator name based on the term mapping table to generate actual standardized terms; performing time alignment processing on the actual collection time to obtain aligned actual time;

[0021] Based on the results of term standardization and time alignment processing, as well as the actual subject ID, an execution acquisition graph is constructed, and the path confidence is calculated; wherein, the execution acquisition graph includes several directed acquisition paths, a directed path of aligned actual time → actual subject ID → actual standardized term; the path confidence is jointly determined by time matching and semantic consistency; the time matching is determined based on the actual time node and time point; the semantic consistency is determined by the standardized indicator node and the acquisition indicator.

[0022] Furthermore, the execution acquisition graph is also provided with a local protocol coverage penalty term, which is used to control the deviation between the actual number of acquisition paths at a certain point in time and the number specified by the protocol.

[0023] Furthermore, the protocol map is compared with the execution acquisition map, and structural deviations of the acquisition path are identified through structural alignment verification to generate a structural deviation matrix, specifically including:

[0024] Obtaining the paths of the protocol graph and the execution acquisition graph and performing structural similarity analysis; wherein the structural similarity analysis includes node semantic consistency, population consistency, and time tolerance matching; if the protocol graph satisfies the structural similarity analysis, selecting the path with the smallest deviation as the optimal matching path;

[0025] Outputting the minimum deviation of all paths in the protocol graph and forming a structural deviation matrix;

[0026] The node semantic consistency indicates that there is a mapping between the collection indicator and the actual standardized term in the term mapping table, or the cosine similarity is ≥0.8; the population consistency indicates that the actual subject ID is the subject population; the time tolerance match indicates that the time tolerance match indicates that the actual collection time difference does not exceed ±2 days.

[0027] Furthermore, the deviation is calculated as follows:

[0028] ;

[0029] in, is the deviation, is the confidence level of executing the acquisition atlas path; It is the collection indicator of the protocol map The local coverage penalty term represents the degree of over-mining / under-mining of the acquisition structure; is the offset aggregation regularization term, which measures whether there are a large number of non-standard paths around the time point; 、 is the regularization term adjustment coefficient.

[0030] Furthermore, the obtained indicator basic fluctuation range and medical standards are combined with the generated structural deviation matrix to generate a consistent behavior score for quantifying the abnormal behavior trend score of the subject, specifically including:

[0031] Extract each subject’s time series dataset from all acquisition paths that execute the acquisition atlas path;

[0032] Constructing a behavioral trend tensor of the subject according to the time series data set to characterize the change rate pattern of the subject at adjacent time points;

[0033] Calculate the trend perturbation density regularization term based on the behavioral trend tensor and time series dataset , expressed as:

[0034] ;

[0035] Among them, T is the total time window, t is the current time window, Indicates the The variance of the change rate of all indicators at a time point, is the correlation between the two indicator time series; is the expected correlation of the pair of indicators under normal test population; is the indicator collaborative deviation adjustment factor;

[0036] The trend disturbance density regularization term and the protocol structure deviation matrix are combined to obtain a final residual term; and the consistency behavior score is determined by the final residual term.

[0037] Furthermore, the causal bias score for each subject is calculated , expressed as:

[0038] ;

[0039] in, Indicates the subject Does it meet the causal trigger conditions; is the total acquisition window, is the threshold for abnormal behavior; Indicates that the causal path is The acquisition confidence in ; is the structure-behavior joint weight;

[0040] in, The representative condition is met but the collection is weak or missing; The representative condition is not met but it is still taken and the behavior is abnormal.

[0041] In a second aspect of the present invention, a clinical trial data verification system based on a modular architecture is provided, the system comprising:

[0042] A protocol parsing module is used to obtain the target clinical trial protocol document, generate a protocol structure map based on the time points, collection indicators and subject population of the target clinical trial protocol document, and map it into a term mapping table;

[0043] A data fusion module is used to generate an execution collection map and path confidence based on the protocol structure map and the actual test collection data;

[0044] The structure verification module is used to compare the protocol map with the execution acquisition map, identify the structural deviation of the acquisition path through structural alignment verification, and generate a structural deviation matrix;

[0045] A behavior modeling module is used to obtain the basic fluctuation range of the indicator and the medical standard, and generate a consistent behavior score in combination with the generated structural deviation matrix to quantify the abnormal behavior trend score of the subject;

[0046] The causal reasoning module is used to combine the protocol structure map, execution collection map, structural deviation matrix and consistency behavior score to calculate the causal deviation score of each subject, and mark the missed collection, wrong collection and timed collection events in the clinical trial data according to the causal deviation score.

[0047] The beneficial technical effects of the present invention are at least as follows:

[0048] To address the challenges of existing technologies, such as weak structural process verification capabilities, crude behavioral pattern analysis, lack of causal logic, and poor system scalability, this paper proposes a clinical trial data verification method and system based on a modular architecture. This system establishes a scalable and composable data verification mechanism that enables intelligent, structured, and multi-dimensional verification of clinical trial data at different levels. The present invention adopts a modular construction concept in its system design, breaking down the data verification process into several functionally independent submodules with standardized interfaces. These modules can be flexibly combined and dynamically adapted to the actual needs of the trial protocol. In terms of verification logic, a collaborative analysis is proposed based on three dimensions: "protocol process execution consistency," "intra-individual behavioral pattern consistency," and "medical causal logic consistency." This creates a multi-layered closed-loop verification system, spanning planning and execution, structure and content, and rules and reasoning. This design not only enables accurate identification of structural deviations, potential forgeries, and medical anomalies, but also enhances the system's adaptability to protocol version adjustments, multi-center deployment, and large-sample parallel processing scenarios. This significantly improves the efficiency and automation of clinical trial data quality control, providing a solid data infrastructure for future digital and intelligent clinical trials. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0050] Figure 1 This is a flow chart of the clinical trial data verification method based on a modular architecture of the present invention. DETAILED DESCRIPTION

[0051] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0052] In one or more embodiments, Figure 1 As shown, a clinical trial data verification method based on a modular architecture is disclosed, and the method includes the following S1 to S5:

[0053] S1. Obtain the target clinical trial protocol document, generate a protocol structure diagram based on the time points, collection indicators, and subject populations of the target clinical trial protocol document, and map it into a term mapping table.

[0054] Specifically, this step aims to extract and structure the content of the clinical trial protocol documents regarding collection time points, indicators, and subjects, and construct a protocol map that will be used uniformly in subsequent processes. .

[0055] Because clinical trial protocols are often written in natural language, they often suffer from inconsistent linguistic styles, complex nested collection logic, frequent conditional collection, and numerous synonymous terms. These problems make them difficult to directly process using keyword extraction or rule templates. Therefore, this paper designs a multi-task structure recognition model that classifies the structural intent and identifies term activation within protocol paragraphs. Furthermore, it incorporates semantic fuzzy constraints when generating the graph, generating a protocol graph that exhibits both medical logic consistency and semantic expression completeness.

[0056] Model input is the segmented text of the agreement document , each Represents a semantic segment, typically at the natural paragraph level. This paper first uses a BERT-based model to predict structural labels for each paragraph and simultaneously identify whether keywords from the standard terminology are triggered. This model has two output heads: one is a classifier for paragraph type (outputting structural categories: time point, indicator item, demographic condition, or irrelevant content), and the other is a binary classifier for term activations. The loss function for joint training is as follows:

[0057] ;

[0058] in, categorical cross entropy loss representing the structural type, represents the binary cross entropy loss of the term activation annotations, Is a balance parameter, set to 0.5. The model uses BERT-base as the backbone, each layer output is 768 dimensions, and two sets of fully connected layers are used for classification after the output. It is used as a supervisory signal for trigger word detection.

[0059] The structural labels and term hit results output by the model are aggregated into structural quads, forming the following set:

[0060] ;

[0061] Each quadruple represents the definition of a collection event. A description of the time point (such as "Day14"), To collect indicators (such as "ALT"), is the subject population (e.g., “intervention group”), is an additional condition in the original protocol that restricts the collection behavior. To enhance operability, the present invention introduces a sliding window mechanism in the processing process to establish the correlation between paragraphs. For example, if The paragraph states "Day 7", while the If the segment "Collect ALT and WBC" appears, the model automatically aggregates these two segments into a four-tuple .

[0062] The fourth element It does not participate in the construction of graph nodes, but only serves as a conditional attribute of the graph edge, used for fuzzy edge weights, collection dependencies and causal relationship extraction.

[0063] Take the following protocol snippet as an example:

[0064] "On Day 7, ALT, WBC and other indicators need to be collected from all intervention group subjects. If the patient experiences adverse reactions, an additional electrocardiogram (ECG) is required."

[0065] The model recognizes two quadruples:

[0066] ;

[0067] ;

[0068] Then enter the graph generation phase. The present invention uses the following function to convert the quaternary set into a protocol graph (the fourth element Does not participate in graph node construction):

[0069] ;

[0070] Each quadruple in the graph is transformed into three nodes (time point, test population, collection index) and two edges, forming a → → To express the uncertainty of conditional logic, this paper designs an edge weight regularization mechanism: the conditional sentence extracted from the original text Sent into the BERT model to obtain the attention weight distribution of each token , and calculate its entropy value as the semantic concentration index. The edge weight is defined as follows:

[0071] ;

[0072] in, represents the attention distribution entropy, is the adjustment parameter (set to 0.6). For example, when the conditional sentence is "If there are adverse reactions, ECG can be added", the attention is focused on fuzzy words such as "can be added", and the entropy value is high. If the probability of edge saturation is low, the edge will be marked as a "low certainty edge," and the alignment requirements will be appropriately relaxed during subsequent graph matching. This semantic mapping mechanism based on text entropy is an innovative structural expression for the semantic ambiguity problem in clinical protocols.

[0073] Ultimately, all nodes will be mapped to a standard encoding system. (Indicators) Use LOINC, (Crowd) Using the ICF grouping structure, The data is uniformly converted into a structure consisting of visit day and allowed offset range. Graph nodes and edges will record the original text index, protocol version number, and sentence position to support subsequent cross-version comparison.

[0074] S2. Based on the protocol structure graph and combined with the actual test collection data, an execution collection graph and path confidence are generated.

[0075] Specifically, the goal of this step is to start from the actual collected data and combine it with the protocol structure map generated in the previous step. and standard terminology mapping table , build an execution collection graph that is consistent with the protocol structure and has the ability to express real data This graph is the key bridge that transforms "experimental design" into "data execution" within the entire patent system. It is not only used for structural consistency analysis in step three but also serves as the behavioral information source for subsequent behavioral modeling and causal logic analysis. Its quality directly determines the accuracy and robustness of the verification process.

[0076] Compared to the protocol structure, the actual data collected in clinical trials is highly uncertain, including:

[0077] Time point offset (e.g., data collected on Day 7 occurred on Day 6 or Day 9);

[0078] Indicator names are not uniform ("alanine aminotransferase" vs "ALT");

[0079] The subject number was separated from the population grouping;

[0080] The same indicator is collected multiple times (e.g., additional collection due to data loss);

[0081] The central difference leads to inconsistent data structure and field encoding.

[0082] Therefore, this paper proposes a two-layer mapping mechanism based on graph structure reconstruction + collection credibility regularization, and introduces a local agreement coverage penalty term to control the problem of graph structure imbalance caused by "over-collection", thereby constructing a graph with dynamic fault tolerance, structural alignment and behavior traceability. .

[0083] The inputs for this step include:

[0084] Protocol structure diagram , from S1; term mapping table , generated by S1, used for indicator standardization; actual collection record set , each , indicating the actual collection time, actual indicator name, and actual subject ID.

[0085] Furthermore, first of all use Perform terminology standardization. If the keyword is not in the standard terminology set, the BERT+CRF joint naming recognition model is used to extract the keywords, and the cosine similarity is used to complete the fuzzy matching, and the matching distance is recorded as the semantic consistency weight. , value range The closer it is to 1, the more consistent the indicator name is with the protocol, and the actual standardized terminology is obtained. .

[0086] Then for all Using the Protocol Map The time nodes defined in Perform time alignment to generate the actual time after alignment. The alignment method is based on an asymmetric time distance function to prevent early acquisition from being mistakenly identified as normal:

[0087] ;

[0088] in, is the actual collection time, The standard date corresponding to the time point of the protocol design, is an adjustment factor (default 0.4). This design emphasizes the realistic clinical preference of "delayed collection is tolerable, early collection requires caution", reflecting the semantic embedding understanding of actual collection behavior.

[0089] After completing the term and time alignment, the present invention constructs the execution acquisition map Each collection record corresponds to a collection path in the graph, which is represented by the time node , subject node , Standardized indicator nodes Three nodes form a directed path The path confidence is determined by the combination of temporal matching and semantic consistency:

[0090] ;

[0091] This confidence is used as an edge attribute Recorded on the side For "soft edge matching" in subsequent atlas comparison.

[0092] Furthermore, the present invention designs a penalty term for the structural distortion problem caused by "repeated collection" or "out-of-range collection" in clinical trials, namely the local agreement coverage penalty term. , used to control a certain point in time The deviation between the actual number of acquisition paths and the number specified in the protocol is as follows:

[0093] ;

[0094] in, For nodes Actual connection Quantity (the type of indicators actually collected at a certain time point in the execution graph), In the protocol diagram The number of metrics that should be collected. This is used as a structural regularization signal in the subsequent behavioral scoring model to avoid the situation where "center habit multiple collections" do not mark protocol changes.

[0095] Take specific data as an example:

[0096] If the protocol requires ALT and WBC to be collected on Day 7, and a center uploads records showing that patient P001 had ALT collected on Day 9 and AST, WBC, CRP, and GLU collected on Day 7, the system will process the data as follows:

[0097] ALT and Day9 → delayed collection, , but the terminology is consistent ;

[0098] WBC and Day 7 → normal collection, ;

[0099] AST, CRP, GLU no agreement mapping, low confidence or classified , by semantic matching Control whether to retain.

[0100] Finally, build When the high confidence path (such as ALT, WBC) is retained, the fuzzy path (such as AST) is marked, and a , prompting the possibility of "over-collection".

[0101] The final output includes: execution acquisition map , all collection paths and their edge weights; local agreement covers the penalty item set , which is used as a structural imbalance signal in subsequent consistency analysis.

[0102] S3. Compare the protocol map with the execution acquisition map, identify the structural deviations of the acquisition path through structural alignment verification, and generate a structural deviation matrix.

[0103] Specifically, the goal of this step is to output the protocol structure graph of step 1. The execution acquisition graph output in step 2 Conduct structural alignment verification to identify whether the actual execution is consistent with the agreement plan, whether there are behavioral deviations such as "non-collection", "wrong collection", "early or delayed collection", "non-structural redundant collection", and finally construct a structural deviation matrix as output.

[0104] Understandably, structural consistency verification plays a key role in the entire patent solution and is the core comparison bridge between "static design" and "dynamic execution". Its accuracy not only affects the input basis of subsequent behavioral models and causal paths, but also directly affects whether the entire verification system is operable. This step introduces three innovative mechanisms in design: (1) structural path fuzzy matching mechanism; (2) joint modeling of confidence and edge coverage penalty; (3) acquisition path aggregation drift regularization term, which is used to solve various structural deviations caused by ambiguous protocol language expression, data upload delays and different acquisition habits between centers in clinical trials.

[0105] Furthermore, the present invention first establishes an alignment matrix at the path level. For each path in the protocol graph , the system searches for paths with similar structures from the execution graph The structural similarity criteria consist of the following three conditions:

[0106] Node semantic consistency: and A mapping exists in the term mapping table, or cosine similarity ;

[0107] Crowd consistency: , that is, the same population;

[0108] Time tolerance matching: , indicating that the collection time difference does not exceed ±2 days.

[0109] If multiple paths meet the above criteria, the system selects the one that minimizes the deviation function as the optimal matching path. The deviation function is as follows:

[0110] ;

[0111] in: is the deviation,

[0112] yes Confidence level (given in step 2, representing execution quality);

[0113] Is a protocol node The local coverage penalty term represents the “over-mining / under-mining degree” of the collection structure;

[0114] is the offset aggregation regularization term, which measures whether there are a large number of non-standard paths around the time point (for example, a large number of "nearby" indicators are collected in a concentrated manner);

[0115] 、 is the regularization term adjustment coefficient (recommended to be 0.3 and 0.2).

[0116] Offset aggregation regularization term The calculation method is:

[0117] ;

[0118] in:

[0119] Indicates the time point of the agreement All acquisition paths within ±2 days;

[0120] is the confidence of each path;

[0121] is a threshold (e.g., 0.6), indicating a “low confidence path”;

[0122] This statistic measures the "density of suspicious collection paths". If there are abnormally large numbers of clusters, it means that the node may have systematic deviations or non-structural expansion behaviors.

[0123] Finally, the system calculates each path in the protocol graph. Output its minimum deviation to form the structural deviation matrix:

[0124] ;

[0125] Indicates the protocol path The degree of deviation from the path with the minimum structural difference among all execution paths. Higher values ​​indicate more severe deviations. The system uses this matrix as a structural risk heat map and inputs it into downstream models.

[0126] Take a specific scenario as an example: If On Day 14, ALT and WBC were collected. ALT appears on Day 16 ( ), WBC appeared on Day 13 and several non-planned indicators such as GLU, AST, CRP were also collected (resulting in ), and there are 8 paths with confidence < 0.5 in this time window ( ), then the final path , the system automatically determines it as a "high structural deviation path".

[0127] Output includes: Structural deviation matrix : Each protocol path corresponds to a deviation value.

[0128] S4. Obtain the basic fluctuation range of the indicator and the medical standard, combine them with the generated structural deviation matrix, and generate a consistent behavior score for quantifying the abnormal behavior trend score of the subject.

[0129] Specifically, the goal of this step is to build a behavioral consistency scoring mechanism that integrates structural deviation driving, trend trajectory analysis, and indicator collaborative anomaly modeling. , which is used to quantify the degree of behavioral deviation of each subject during the actual data collection process. Behavioral deviation is not equivalent to the numerical abnormality of a certain indicator, but refers to whether the fluctuation of indicators, the trend direction, and the linkage between indicators over a period of time are "reasonable", "stable", and "conform to the logic of medical changes in the experimental context." Unlike traditional anomaly detection methods such as z-score and sliding window outlier detection, this solution performs behavioral-level modeling entirely based on the structural map, execution map, and structural scoring results in the previous steps, reflecting the coupled modeling concept under the systematic patented structure.

[0130] The input of this step includes: executing the acquisition map : From step 2, including each subject's collection behavior, actual indicator value and path confidence ; Protocol structure deviation matrix : From step 3, used to provide focus on structural anomalies; standard terminology mapping table : Used to normalize the indicator values ​​of each center; the basic fluctuation range of the indicator and the medical standard (provided by the clinical knowledge base) are used for normalization and trend anomaly modeling.

[0131] Further, the present invention firstly combines all Collection path Each subject was extracted Time series dataset ,in Indicates that the subject is at time point The collected All indicator values ​​are Match, map to standard encoding and unify the units, and normalize to the interval [0,1].

[0132] Then construct the subject behavior trend tensor , which is used to characterize the rate of change pattern at adjacent time points:

[0133] ;

[0134] in:

[0135] Indicates the Indicators in time The value of

[0136] is the standard fluctuation value of the indicator in the subject population (extracted from the clinical knowledge base);

[0137] is a small constant to avoid division by zero (such as );

[0138] This tensor has a time dimension and indicator dimensions Expanded to structure.

[0139] Considering that some abnormalities in real clinical data are not caused by mutations, but by collective “drift” or “abnormal stability” of indicators (such as long-term maintenance at marginal values ​​and loss of coordination between indicators), this paper introduces a trend perturbation density regularization term , used to capture the behavior state where the rate of change is obviously not concentrated:

[0140] ;

[0141] in:

[0142] Indicates the The variance of the rate of change of all indicators at a time point. A low value means that all indicators fluctuate similarly, while a too low value may indicate "abnormal stability";

[0143] is the correlation between the two indicator time series;

[0144] is the expected correlation of the pair of indicators under normal test population;

[0145] is the indicator synergy deviation adjustment factor, with an empirical value of 0.3;

[0146] This regularization term can punish both the “false stability” caused by too concentrated fluctuations and the structural anomalies caused by “co-destruction of indicators”.

[0147] Final residual Comprehensive trend change degree and coordinated disturbance, and introduce structural deviation weight As a signal for attention region regulation, a consistent behavior score was constructed:

[0148] ;

[0149] in,

[0150] The first term represents the overall fluctuation intensity of the subject under all time periods;

[0151] The second term activates the trend perturbation regularization in structural anomaly areas, further penalizing atypical trend behavior;

[0152] is the regularization adjustment coefficient (empirical value 0.5~1.0);

[0153] is the structural deviation activation threshold (e.g. 0.7).

[0154] The system finally outputs the behavior consistency score , normalized to (0,1), with higher values ​​indicating greater “conformity to medical trend logic” and lower values ​​indicating stronger behavioral deviations. This value will be directly used in subsequent causal path activation and risk fusion scoring.

[0155] For example, during a collection at a certain center, a subject's ALT, WBC, and CRP indicators showed minimal changes (<2%) from Day 5 to Day 12, with fluctuation variances <0.005, and these time points were all at high structural deviations ( ), the system judges it as "false stability under structural abnormality", through Activate the trend disturbance penalty, resulting in Significant decrease (e.g. ), and thus are classified as individuals at high risk of behavioral abnormalities.

[0156] The output of this step is: : Behavioral consistency score for each subject; Optional: Annotate the behavioral abnormality degree map by time point and indicator dimension for visualization system display.

[0157] S5. Calculate the causal deviation score for each subject based on the protocol structure map, execution collection map, structural deviation matrix, and consistent behavior score, and mark missed collection, incorrect collection, and timed collection events in the clinical trial data according to the causal deviation score.

[0158] Specifically, the goal of this step is to complete the identification and modeling of causal logic deviations in the execution of clinical trials, and output the causal consistency deviation score for each subject. , and clearly mark the interruption points of the causal path, which serve as the direct basis for subsequent automatic system review, manual review by auditors, and report generation.

[0159] In the previous steps, the system has completed structural alignment (step 3) and behavioral consistency modeling (step 4). However, clinical trial compliance requires more than just "correct structure" and "correct data." More importantly, it requires logical data collection under the correct preconditions, i.e., causal consistency.

[0160] For example, the protocol defines: "If a serious adverse event (SAE) occurs, blood pressure and ECG measurements must be collected within 48 hours." This type of collection design with preconditions is extremely common, but in actual implementation, the following may occur:

[0161] The trigger condition is met but no sampling occurs (missed sampling);

[0162] Trigger conditions are not met but the mining is still carried out (wrong mining);

[0163] The conditions are met but the sampling timeout occurs, the indicators are incorrect, or the behavior is abnormal;

[0164] This step is used to identify and quantify these causal logic misalignment behaviors and complete the final verification score Output.

[0165] Among them, the input includes: protocol structure map : Contains conditional collection edges; executes collection graph : Contains path collection confidence ; Protocol structure deviation matrix : Prompt offset and misplaced structure; behavioral consistency score : Provides behavioral trend anomaly scores; causal rule sets defined in the protocol ,in As the trigger condition, The adopted indicator.

[0166] Furthermore, the system analyzes The real data of each subject in the test is used to judge each rule Whether it is effectively implemented in practice.

[0167] Whether the trigger conditions are met (such as SAE ≥ grade 2, abnormal body temperature, rapid drop in WBC);

[0168] Was the data collected within the specified time? ;

[0169] Is the collection behavior high confidence? high);

[0170] Is the behavior trend reasonable? high);

[0171] Furthermore, the present invention defines each subject as The causal bias scores are as follows:

[0172] ;

[0173] in:

[0174] :Indicates the subject Does it meet the Causal trigger conditions (analysis Record in the system, if SAE event occurs);

[0175] :The causal path is The acquisition confidence in ;

[0176] :Subjects Overall behavioral consistency score (output from step 4);

[0177] The first item represents that the condition is met but the collection is weak or missing; the second item represents that the condition is not met but the collection is still carried out, and the behavior is abnormal;

[0178] is the structure-behavior joint weight (recommended value: 0.5~0.7);

[0179] For the behavioral abnormality threshold, it is recommended to set it to the 30th percentile of the behavioral score distribution.

[0180] After the system is finished running, the following results will be output:

[0181] Causal consistency bias score for each subject ;

[0182] Furthermore, for For high-risk subjects, the system will automatically:

[0183] A list of paths marked with causal trigger failures;

[0184] Give reason labels such as "not collected but should have been collected", "collected by mistake and abnormal behavior", "collected at the wrong time";

[0185] Push its number to the review interface and report generation module for manual review and correction;

[0186] Optional export of review checklists and Excel reports for review by CROs or regulators.

[0187] In one or more embodiments, a clinical trial data verification system based on a modular architecture is disclosed, the system comprising:

[0188] A protocol parsing module is used to obtain the target clinical trial protocol document, generate a protocol structure map based on the time points, collection indicators and subject population of the target clinical trial protocol document, and map it into a term mapping table;

[0189] A data fusion module is used to generate an execution collection map and path confidence based on the protocol structure map and the actual test collection data;

[0190] The structure verification module is used to compare the protocol map with the execution acquisition map, identify the structural deviation of the acquisition path through structural alignment verification, and generate a structural deviation matrix;

[0191] A behavior modeling module is used to obtain the basic fluctuation range of the indicator and the medical standard, and generate a consistent behavior score in combination with the generated structural deviation matrix to quantify the abnormal behavior trend score of the subject;

[0192] The causal reasoning module is used to combine the protocol structure map, execution collection map, structural deviation matrix and consistency behavior score to calculate the causal deviation score of each subject, and mark the missed collection, wrong collection and timed collection events in the clinical trial data according to the causal deviation score.

[0193] It is worth noting that the specific workflow of the clinical trial data verification system based on modular architecture provided in the embodiment of the present invention is the same as the workflow of the clinical trial data verification method based on modular architecture described in the above embodiment, and will not be repeated here.

[0194] An embodiment of the present invention further provides a clinical trial data verification device based on a modular architecture, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the steps in the above-mentioned embodiment of the clinical trial data verification method based on the modular architecture are implemented, for example Figure 1 or, the processor implements the functions of the modules in the above-mentioned system embodiments when executing the computer program.

[0195] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the clinical trial data verification device based on a modular architecture.

[0196] The modular clinical trial data verification device can be a computing device such as a desktop computer, laptop, PDA, or cloud server. The modular clinical trial data verification device can include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the modular clinical trial data verification device can also include input / output devices, network access devices, buses, and the like.

[0197] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor serves as the control center of the clinical trial data verification device based on a modular architecture, connecting the various components of the entire clinical trial data verification device based on a modular architecture using various interfaces and lines.

[0198] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of the modular architecture-based clinical trial data verification device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system, at least one application required for a function, and the data storage area may store data generated based on the operation of the air conditioning controller. Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a Smart Memory Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0199] If the modules integrated into the modular clinical trial data verification device are implemented as software functional units and sold or used as standalone products, they can be stored in a computer-readable storage medium. Based on this understanding, the present invention can also implement all or part of the process steps in the above-mentioned method embodiments by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a removable hard drive, a magnetic disk, an optical disk, computer memory, read-only memory (ROM), random access memory (RAM), an electrical carrier signal, a telecommunications signal, and software distribution media.

[0200] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0201] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A clinical trial data verification method based on a modular architecture, characterized in that: The method comprises: Obtain the target clinical trial protocol document, generate a protocol structure diagram based on the time points, collection indicators, and subject populations of the target clinical trial protocol document, and map it into a term mapping table; Based on the protocol structure graph and combined with the actual test collection data, an execution collection graph and path confidence are generated; Compare the protocol map with the execution acquisition map, identify the structural deviations of the acquisition path through structural alignment verification, and generate a structural deviation matrix; Obtaining the basic fluctuation range of the indicator and the medical standard, and combining them with the structural deviation matrix, to generate a consistent behavior score, which is used to quantify the abnormal behavior trend score of the subject; Calculate the causal deviation score for each subject by combining the protocol structure map, execution collection map, structural deviation matrix, and consistent behavior score, and mark missed collection, incorrect collection, and timed collection events in the clinical trial data according to the causal deviation score; The acquisition of the basic fluctuation range of the indicator and the medical standard, combined with the structural deviation matrix, generates a consistent behavior score for quantifying the abnormal behavior trend score of the subject, specifically including: Extract each subject’s time series dataset from all acquisition paths that execute the acquisition atlas path; Constructing a behavioral trend tensor of the subject according to the time series data set to characterize the change rate pattern of the subject at adjacent time points; Calculate the trend perturbation density regularization term based on the behavioral trend tensor and time series dataset , expressed as: ； Among them, T is the total time window, t is the current time window, Indicates the The variance of the change rate of all indicators at a time point, is the correlation between the two indicator time series; is the expected correlation of the pair of indicators under normal test population; is the indicator collaborative deviation adjustment factor; Combining the trend perturbation density regularization term and the structural deviation matrix to obtain a final residual term; determining the consistency behavior score through the final residual term; The causal bias score for each subject was calculated as described , expressed as: ； in, Indicates the subject Does it meet the causal trigger conditions; is the total acquisition window, is the threshold for abnormal behavior; Indicates that the causal path is in the collection graph The acquisition confidence in ; is the structure-behavior joint weight; Each subject was scored for behavioral consistency; in, The representative condition is met but the collection is weak or missing; The representative condition is not met but it is still taken and the behavior is abnormal.

2. The clinical trial data verification method based on modular architecture according to claim 1, characterized in that: The target clinical trial protocol document is obtained, and a protocol structure diagram is generated according to the time point, collection indicators, and subject population of the target clinical trial protocol document, and mapped into a term mapping table, specifically including: Obtain the target clinical trial protocol document; input the target clinical trial protocol document into the BERT-based model, predict the structure label for each paragraph, and output a structure quadruple, including the time point, collection indicators, test population, and additional conditions in the original protocol that restrict the current collection behavior; Convert the time point, collection index, and test population into three nodes and two edges to form a directed path from time point → test population → collection index; All nodes were mapped to a standard coding system as a term mapping table; the collection indicators used LOINC; the subject population used the ICF grouping structure; and the time points were converted into a structure of visit days and allowed offset ranges.

3. The clinical trial data verification method based on modular architecture according to claim 2, characterized in that: The protocol structure graph also has an edge weight regularization mechanism, which specifically includes: The additional conditions that restrict the current collection behavior in the original agreement are extracted from the target clinical trial protocol document, and the additional conditions that restrict the current collection behavior in the original agreement are fed into the BERT model to obtain the attention weight distribution of each token, and its entropy value is calculated as the semantic concentration index; the edge weight is constrained according to the semantic concentration index.

4. The clinical trial data verification method based on modular architecture according to claim 1, characterized in that: Each data of the actual collected data of the experiment includes the actual collection time, the actual indicator name, and the actual subject ID; Then there is, The generation of the execution acquisition graph and path confidence based on the protocol structure graph and combined with the actual test data collection specifically includes: Performing terminology standardization processing on the actual indicator name based on the term mapping table to generate actual standardized terms; performing time alignment processing on the actual collection time to obtain aligned actual time; Based on the results of term standardization and time alignment processing, as well as the actual subject ID, an execution acquisition graph is constructed, and the path confidence is calculated; wherein, the execution acquisition graph includes several directed acquisition paths, a directed path of aligned actual time → actual subject ID → actual standardized term; the path confidence is jointly determined by time matching and semantic consistency; the time matching is determined based on the actual time node and time point; the semantic consistency is determined by the standardized indicator node and the acquisition indicator.

5. The clinical trial data verification method based on modular architecture according to claim 4, characterized in that: The execution acquisition graph is also provided with a local protocol coverage penalty term, which is used to control the deviation between the actual number of acquisition paths at a certain point in time and the number specified by the protocol.

6. The clinical trial data verification method based on modular architecture according to any one of claims 2 or 4, characterized in that: The protocol map and the execution acquisition map are compared to identify the structural deviation of the acquisition path through structural alignment verification to generate a structural deviation matrix, specifically including: Obtaining the paths of the protocol graph and the execution acquisition graph and performing structural similarity analysis; wherein the structural similarity analysis includes node semantic consistency, population consistency, and time tolerance matching; if the protocol graph satisfies the structural similarity analysis, selecting the path with the smallest deviation as the optimal matching path; Outputting the minimum deviation of all paths in the protocol graph and forming a structural deviation matrix; The node semantic consistency indicates that there is a mapping between the collection indicator and the actual standardized term in the term mapping table, or the cosine similarity is ≥0.8; the population consistency indicates that the actual subject ID is the subject population; the time tolerance match indicates that the actual collection time difference does not exceed ±2 days.

7. The clinical trial data verification method based on modular architecture according to claim 6, characterized in that: The deviation is calculated as follows: ； in, is the deviation, is the confidence level of executing the acquisition atlas path; It is the collection indicator of the protocol map The local coverage penalty term represents the degree of over-mining / under-mining of the acquisition structure; is the offset aggregation regularization term, which measures whether there are a large number of non-standard paths around the time point; 、 is the regularization term adjustment coefficient.

8. A system for executing the clinical trial data verification method based on modular architecture according to claim 1, characterized in that: The system comprises: A protocol parsing module is used to obtain the target clinical trial protocol document, generate a protocol structure map based on the time points, collection indicators and subject population of the target clinical trial protocol document, and map it into a term mapping table; A data fusion module is used to generate an execution collection map and path confidence based on the protocol structure map and the actual test collection data; The structure verification module is used to compare the protocol map with the execution acquisition map, identify the structural deviation of the acquisition path through structural alignment verification, and generate a structural deviation matrix; A behavior modeling module is used to obtain the basic fluctuation range of the indicator and the medical standard, and combine it with the structural deviation matrix to generate a consistent behavior score for quantifying the abnormal behavior trend score of the subject; The causal reasoning module is used to combine the protocol structure map, execution collection map, structural deviation matrix and consistency behavior score to calculate the causal deviation score of each subject, and mark the missed collection, wrong collection and timed collection events in the clinical trial data according to the causal deviation score.

Images