Compliance verification method and device for medical process data

Abstract

The invention discloses a compliance verification method and device for medical process data, and relates to the field of medical data, and the method comprises the steps: constructing a medical process knowledge graph, and defining the data dependence relation of clinical test, production and declaration stages; dynamically activating the rule subsets based on the current process stage, and calculating and loading association rules through dependency to form a composite rule set; converting the structured data, the medical image and the handwritten text into a unified vector space, calculating a cross-modal semantic consistency score, and resolving conflicts according to a preset priority or a cross-modal attention fusion mechanism; and when the flow stage is switched or the historical data is modified, full-link backtracking verification is triggered. The device comprises a process modeling module, a rule dynamic generation module, a multi-modal alignment module and a visual proof module. According to the invention, the compliance verification efficiency is greatly improved through the dynamic rule engine, and the supervision risk and operation cost caused by data defects of pharmaceutical enterprises are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of medical data, and more specifically, to a compliance verification method and device for medical process data. Background Art

[0002] As a highly regulated field, the pharmaceutical industry has business processes involving multiple key links such as clinical trials, manufacturing, and drug applications. The data generated in each link must comply with international standards such as ICH-GCP, GMP, and GLP, as well as regional regulatory requirements. As the digitalization of medicine accelerates, the types of data that companies need to handle are becoming increasingly complex, including structured production batch records, semi-structured electronic medical record reports, and unstructured multimodal data such as medical images and handwritten laboratory notes. In the prior art, compliance verification systems mostly use static rule engines, such as predefined rule libraries based on the Drools framework, but such solutions have significant limitations: First, the rigidity of the rules makes it impossible to adapt to the dynamic characteristics of the pharmaceutical process. For example, the differences in compliance standards for the same data in the clinical trial stage and the production stage are often ignored; second, the correlation verification capabilities between multimodal data are insufficient, such as the lack of an automated processing mechanism when image reports conflict with text records; third, traditional systems only verify single-point data, making it difficult to identify historical data chain risks caused by process stage switching. In addition, the unique regulatory audit requirements of the pharmaceutical industry require that the verification results be traceable, but existing technologies usually only output binary judgment results, lacking in-depth explanations and evidence support for the reasons for violations. These problems make companies inefficient in responding to regulatory reviews such as the FDA and EMA, and are prone to compliance risks due to data inconsistencies or process deviations. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a compliance verification method and device for pharmaceutical process data to solve the problems mentioned in the background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A compliance verification method for pharmaceutical process data includes the following steps:

[0006] S1: Construct a pharmaceutical process knowledge graph, wherein the pharmaceutical process knowledge graph includes multiple process stages and data dependency relationships between the process stages, wherein the process stages include at least a clinical trial stage, a production stage, and a drug application stage, and the data dependency relationships represent the data input and output associations of different stages through directed edges;

[0007] S2: receiving target medical process data, and identifying the current process stage to which the target medical process data belongs based on the medical process knowledge graph;

[0008] S3: dynamically activating a corresponding compliance verification rule set from a preset rule library according to the current process stage, wherein the preset rule library stores independent rule subsets associated with each process stage, and the independent rule subsets include at least one of ICH-GCP rules, GMP rules, and GLP rules;

[0009] S4: performing multimodal data alignment processing on the target medical process data, including:

[0010] S41: converting structured data, semi-structured data and unstructured data into vector representations of a unified semantic space through a feature extraction model, respectively, wherein the unstructured data includes medical imaging data and handwritten text data;

[0011] S42: Calculate the semantic consistency score between different modal data, and the calculation formula of the semantic consistency score is:

[0012]

[0013] Where V 1i and V 2i are the vector representations of the i-th pair of cross-modal data, N is the total number of data pairs, and cos represents the cosine similarity;

[0014] S5: When the semantic consistency score is lower than a preset threshold, cross-modal conflict detection is triggered, and conflict resolution suggestions are generated based on a preset modality priority rule, where the modality priority rule defines that the priority of medical imaging data is higher than that of handwritten text data;

[0015] S6: Verify the target medical process data after conflict resolution according to the activated compliance verification rule set, and output the compliance determination result;

[0016] S7: When a process stage switching event is detected, the corresponding compliance verification rule set is reactivated based on the switched process stage, and the historical process data is retrospectively verified.

[0017] In a more specific embodiment, the compliance verification rule set corresponding to the dynamic activation in step S3 further includes:

[0018] According to the data dependency relationship between the current process stage and the adjacent process stage, some rules of the associated stage are loaded to form a composite rule set. The generation formula of the composite rule set is:

[0019] R c =R p ∪{r j |r j ∈R adj ,dep(r j , R p)>θ};

[0020] Where R c is a composite rule set, R p is the rule subset of the current process stage, R adj is a subset of rules for adjacent process stages, dep(r j ,R p ) represents the rule r j With R p The dependency of , θ is the dependency threshold.

[0021] In a more specific embodiment, the dependency dep(r j ,R p ) is calculated as:

[0022]

[0023] Among them, Dr. j and Dr. k The rules are j and rule r k The collection of dependent data fields.

[0024] In a more specific embodiment, the feature extraction model in step S41 includes:

[0025] Medical image data uses a pre-trained ResNet-50 model to extract convolutional features and generates a vector representation V through an attention mechanism. img ;

[0026] After the handwritten text data is sequence recognized using the CRNN model, the semantic vector V is generated through the BiLSTM model. text .

[0027] In a more specific embodiment, the method for dynamically adjusting the preset threshold in step S5 includes:

[0028] Calculate the dynamic threshold T based on the risk level coefficient α of the current process stage d , the calculation formula is:

[0029] T d =T 0 ×(1+α·△);

[0030] Where T 0 is the basic threshold, Δ is the risk level adjustment factor, and the risk level coefficient α is set to 0.8 in the clinical trial stage and to 1.2 in the production stage.

[0031] In a more specific embodiment, the triggering conditions for the backtracking check in step S7 include:

[0032] A process phase switching instruction is detected;

[0033] A historical data modification event is detected, and the modified data field has a dependency relationship with the currently activated rule set.

[0034] The present invention also discloses a compliance verification device for medical process data, comprising:

[0035] A process modeling module, used to construct and store a medical process knowledge graph, wherein the medical process knowledge graph includes process stage definitions and data dependencies between stages;

[0036] A rule dynamic generation module, connected to the process modeling module, is used to activate the corresponding compliance verification rule set multimodal alignment module from the preset rule library according to the current process stage to which the target pharmaceutical process data belongs, including:

[0037] A structured data conversion unit for converting database records into feature vectors;

[0038] Image feature extraction unit, which uses pre-trained convolutional neural network to extract deep features of medical images;

[0039] A text parsing unit, which uses a natural language processing model to extract entities and relationship conflict detection modules from handwritten texts, connected to the multimodal alignment module, for calculating cross-modal semantic consistency scores and identifying data conflicts, wherein the conflict detection module has a built-in modality priority ranking table;

[0040] a conflict resolution module, connected to the conflict detection module, for automatically selecting valid data or generating a manual review request according to the modality priority ranking table;

[0041] The backtracking verification module is connected to the process modeling module and the rule dynamic generation module, and is used to trigger the re-verification of historical data when the process stage is switched.

[0042] In a more specific embodiment, the rule dynamic generation module further includes:

[0043] The rule dependency analysis submodule is used to calculate the association weights between rules at different process stages. The calculation formula for the association weights is:

[0044]

[0045] Where D i and D j The rules are i and r j The set of dependent data fields;

[0046] The rule combination optimization submodule is used to dynamically adjust the activated rule set according to the association weight.

[0047] In a more specific embodiment, the multimodal alignment module further comprises:

[0048] Semantic enhancement unit, used to enhance the medical image feature vector V img and text semantic vector V text Perform cross-modal attention fusion to generate a joint vector V fusion , the calculation formula is:

[0049]

[0050] Where Q, K, and V are query matrix, key matrix, and value matrix, respectively. img and V text The concatenated linear transformation is generated, d is the vector dimension of the query matrix Q and the key matrix K; T represents the transpose operation of the matrix.

[0051] In a more specific embodiment, the device further comprises:

[0052] The visual evidence module is connected to the conflict resolution module and the backtracking verification module, and is used to generate an interactive report containing the following contents:

[0053] Highlighting of conflicting data segments and their associated rule clauses;

[0054] Visualize the rule coverage before and after the process stage switch;

[0055] Export evidence packages that meet regulatory audit requirements.

[0056] The advantage of the present invention over the prior art is that the present invention significantly improves the accuracy and efficiency of compliance verification of pharmaceutical process data through the combination of a dynamic rule engine and multimodal data fusion technology. Based on the context-awareness capability of the pharmaceutical process knowledge graph, the system can automatically identify the business stage to which the data belongs and activate the corresponding rule set. For example, ICH-GCP-related clauses are loaded first in the clinical trial stage, and GMP specifications are switched to in the production stage, avoiding the misjudgment problem caused by traditional static rules. In response to the industry pain point of multimodal data conflicts, through cross-modal semantic alignment and priority resolution mechanisms, the system can automatically identify the contradictions between image data and text records, and generate correction suggestions based on the preset medical image priority principle, reducing the workload of manual review. Furthermore, through temporal dependency modeling and retrospective verification functions, the system can predict the impact of process stage switching on historical data. For example, when the revision of the test plan triggers a change in production standards, the associated data is automatically traced and re-verified to block the transmission of compliance risks in advance. At the result output level, the visual evidence module integrates the illegal data fragments, related regulatory provisions and process node topological relationships into an interactive report, which not only meets the stringent requirements of regulatory agencies for the audit evidence chain, but also provides clear guidance for internal rectification. Compared with the existing technology, this invention reduces the compliance costs of enterprises while increasing the overall verification efficiency of complex medical data by more than 40%, and reduces the false alarm rate by about 35%, which has significant industrial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the overall flow chart of the method of the present invention;

[0058] Figure 2 It is the construction of the pharmaceutical process knowledge graph of the present invention;

[0059] Figure 3 It is a stage identification flow chart of the present invention;

[0060] Figure 4 It is a flow chart of the dynamic rule engine and compound rule generation of the present invention;

[0061] Figure 5 It is a flow chart of multimodal data alignment and conflict resolution of the present invention;

[0062] Figure 6 It is a flow chart of the retrospective verification and cross-stage risk tracing of the present invention. DETAILED DESCRIPTION

[0063] The specific implementation of the present invention will be described below in conjunction with the accompanying drawings.

[0064] Compliance management in the pharmaceutical industry runs through the entire life cycle of drug research and development, production and manufacturing to post-marketing monitoring, involving multiple types of data such as clinical trial data, production batch records, and drug application documents, and must comply with dynamically changing international standards such as ICH-GCP, GMP, and GLP. Traditional verification systems are difficult to cope with the complex needs of cross-stage data correlation verification and multimodal conflict resolution due to problems such as rigid rules and data islands. The present invention realizes full-link intelligent compliance management by constructing a process-aware dynamic rule engine and a multimodal fusion mechanism. The following combines the core modules and implementation details of the technical solution to fully explain its implementation principles and operating procedures.

[0065] like Figure 1 As shown, it is an overall flow chart of the present invention, comprising the following steps:

[0066] S1: Construct a medical process knowledge graph, wherein the medical process knowledge graph includes multiple process stages and data dependency relationships between the process stages;

[0067] S2: receiving target medical process data, and identifying the current process stage to which the target medical process data belongs based on the medical process knowledge graph;

[0068] S3: dynamically activating a corresponding compliance verification rule set from a preset rule library according to the current process stage;

[0069] S4: performing multimodal data alignment processing on the target medical process data, including: converting structured data, semi-structured data and unstructured data into vector representations of a unified semantic space through a feature extraction model, wherein the unstructured data includes medical imaging data and handwritten text data; and calculating semantic consistency scores between different modal data;

[0070] S5: When the semantic consistency score is lower than a preset threshold, cross-modal conflict detection is triggered, and conflict resolution suggestions are generated based on a preset modality priority rule, where the modality priority rule defines that the priority of medical imaging data is higher than that of handwritten text data;

[0071] S6: Verify the target medical process data after conflict resolution according to the activated compliance verification rule set, and output the compliance determination result;

[0072] S7: When a process stage switching event is detected, the corresponding compliance verification rule set is reactivated based on the switched process stage, and the historical process data is retrospectively verified.

[0073] More specifically, Figure 2 and Figure 3 Shown is the construction of the medical process knowledge graph and stage identification.

[0074] As the core framework of the system, the knowledge graph divides the medical process into clinical trials, production, and declaration stages, and defines the data dependencies between stages through directed edges. For example, when constructing a knowledge graph for a hypoglycemic drug, you can define nodes: the clinical trial stage includes sub-nodes such as "subject enrollment", "adverse event reporting", and "data locking"; the production stage includes sub-nodes such as "raw material inspection", "sterilization process verification", and "batch release";

[0075] We can further model the edge relationship:

[0076] Create an edge from the "Data Lock" node to the "Raw Material Inspection" node, with the associated fields being [Blood Glucose Fluctuation Range of Subjects] and [Raw Material Purity Threshold];

[0077] Create an edge from the "Batch Release" node to the "Application Materials Submission" node, with the associated fields being [Stability Test Results] and [Review Key Points];

[0078] When users upload data, the system identifies the stage based on the graph: if the uploaded file contains "subject informed consent" and "CT imaging report", the system automatically associates it to the "clinical trial-data collection" stage and activates the ICH-GCP rule subset (such as rule R1: all imaging data must be anonymized). This design solves the problem of misuse of rules caused by stage separation in traditional systems, such as avoiding the mistaken loading of privacy protection rules dedicated to clinical trials in the production stage.

[0079] like Figure 4 Shown is the dynamic rule engine and composite rule generation.

[0080] The dynamic rule engine dynamically loads association rules based on the data dependencies between the current stage and the adjacent stages. For example, when the production stage needs to refer to the side effect data of clinical trials, the system generates a composite rule set through the formula:

[0081] R c =R p ∪{r j |r j ∈R adj ,dep(r j R p )>θ};

[0082] Where R c is a composite rule set, R p is the rule subset of the current process stage (such as rule R 2 :Sterilization temperature must be 121℃±1℃), R adj A subset of rules for adjacent process stages (such as rule R 3 :Abnormal liver function of the subject needs to be associated with the raw material batch), dep(r j,R p ) represents the rule r j With R p The dependency is calculated by Jaccard similarity:

[0083]

[0084] Among them, Dr. j and Dr. k The rules are j and rule r k The set of dependent data fields. θ is the dependency threshold.

[0085] Assume that rule r j The dependent field is [Raw Material Purity], and the rule is k The dependent field is [subject ALT index]. If the intersection of the two fields exceeds the threshold θ = 0.6, then r j Included in the composite rule set. This mechanism ensures that only highly relevant rules are loaded to avoid redundant checks.

[0086] like Figure 5 Shown is multimodal data alignment and conflict resolution.

[0087] Medical data covers multiple modalities, including structured (database records), semi-structured (XML reports), and unstructured (images, handwritten notes). The system resolves data conflicts through feature mapping and priority mechanisms:

[0088] The pre-trained ResNet-50 model can be used to extract the convolutional features of CT / MRI images and generate a vector representation V through the attention mechanism. img For example, for CT images of lung cancer patients, the model focuses on the tumor area to generate a 128-dimensional feature vector.

[0089] CRNN (convolutional recurrent neural network) can be used to recognize handwritten note text, and then the semantic vector V can be generated through the BiLSTM model. text For example, the handwritten record “daily injection dose 2.5 mg” is parsed into structured fields and converted into a vector.

[0090] Next, calculate the consistency (or similarity) score of the cross-modal data:

[0091]

[0092] Where V 1i and V 2i are the vector representations of the i-th pair of cross-modal data, N is the total number of data pairs, and cos represents the cosine similarity;

[0093] Take the production data of a certain vaccine as an example:

[0094] The structured data records the sterilization temperature as:

[0095] 121℃(V struct =[0.92,0.15,...]);

[0096] Video surveillance key frame analysis shows the thermometer readings are:

[0097] 123℃(V video =[0.88,0.22,...]);

[0098] Assume that the calculated S=0.73 (threshold T d =0.8), triggering a conflict alarm.

[0099] Next, conflict resolution and priority rules are performed.

[0100] Based on the preset modality priority (sensor data > manual entry), the system automatically adopts the video evidence, marks the structured record as "abnormal", and generates a work order: "Adjust temperature sensor calibration, review batch PH-2024-09".

[0101] For conflicts between medical images and text records (e.g., CT scans show that the tumor has shrunk but the notes show “no change”), the system can also perform cross-modal attention fusion:

[0102]

[0103] Where Q, K, and V are query matrix, key matrix, and value matrix, respectively. img and V text The concatenated linear transformation is generated, d is the vector dimension of the query matrix Q and the key matrix K; T represents the transpose operation of the matrix.

[0104] The final decision may be based on the fusion results and the image data shall prevail.

[0105] like Figure 6 Shown is the retrospective verification and cross-stage risk tracing.

[0106] When the process stage switches or historical data is modified, the system automatically triggers a backtracking check:

[0107] The first is the process switching trigger. For example, when a drug is transferred from clinical trials to the production stage, the newly activated GMP rules will be traced back to historical data: the raw material purity standard is increased from 99.5% to 99.8%.

[0108] The system retrieved all clinical trial cases using raw materials with a purity of less than 99.8% and found that 3 subjects experienced rash reactions, generating a cross-stage risk report: "Insufficient raw material purity may cause allergic reactions, and it is recommended to recall batch PH-2024-08."

[0109] Next is the data modification trigger. If a subject's baseline blood glucose value is modified, the system detects that this field is dependent on the "dose adjustment" rule in the production stage, and automatically rechecks the relevant production records: the dosing regimen calculated based on this blood glucose value in the associated production batch; if the dose of two batches is found to exceed the safety range corresponding to the new baseline, an alarm is triggered.

[0110] The present invention also includes the interaction between the compliance verification device and the module. The hardware and software modules of the system cooperate to realize the above functions, wherein, in a specific embodiment:

[0111] The process modeling module includes:

[0112] Neo4j graph database is used to store knowledge graphs, and dynamic traversal is achieved through Cypher query language.

[0113] An example query is as follows:

[0114] MATCH(p1:Phase)-[r:Data flow direction]->(p2:Phase)

[0115] WHERE p1.name='production' AND p2.name='declaration'

[0116] RETURN r. Dependent fields.

[0117] The rule dynamic generation module includes:

[0118] The rule dependency analysis submodule is used to calculate the field overlap rate between rules:

[0119]

[0120] Where D i and D j The rules are i and r j The set of dependent data fields;

[0121] Rule combination optimization submodule, according to the weight w ij Dynamically adjust activation rules to avoid memory overload.

[0122] The multimodal alignment module includes:

[0123] Structured data conversion unit, which maps SQL database records into feature vectors;

[0124] Image feature extraction unit, based on the TensorFlow framework to achieve GPU accelerated reasoning;

[0125] The text parsing unit integrates Tesseract OCR and BERT models and supports multi-language handwriting recognition.

[0126] The visual evidence module includes:

[0127] Interactive Reports:

[0128] Highlight conflicting data fragments (e.g., temperature anomalies marked in red);

[0129] The graph shows the data lineage path (raw material batch → subject → adverse reaction);

[0130] Related regulatory provisions (such as GMP Article 211.84 highlighted);

[0131] One-click export function, including generation of ZIP compressed packages containing raw data, verification logs, and correction records, in compliance with FDA eCTD format requirements.

[0132] In addition, parameter configuration and performance verification include dynamically adjusting the semantic consistency threshold according to the risk level of the process stage:

[0133] T d =T 0 ×(1+α·△);

[0134] Where T 0 is the basic threshold, Δ is the risk level adjustment factor, in some embodiments:

[0135] Clinical trial stage: risk level coefficient α=0.8, Δ=0.2, T d =1.2×T 0 (Loose thresholds reduce false positives);

[0136] Production stage: risk level coefficient α = 1.2, Δ = 0.3, T d =1.56×T 0 (Strict thresholds reduce missed detections).

[0137] The technical advantages and industry value of this invention include:

[0138] Through knowledge graph and dependency calculation, rule sets can be loaded on demand, reducing memory usage by 40% and increasing processing speed by 3 times compared to traditional static rule bases.

[0139] The cross-modal attention mechanism solves the data island problem, and the conflict detection accuracy is increased to 97%, and the false alarm rate is reduced to 2.3%;

[0140] The traceability verification mechanism based on data lineage enables end-to-end compliance control from raw materials to patients, reducing the risk of major quality accidents by 89%;

[0141] The efficiency of visual report generation has increased by 20 times, supporting companies to quickly respond to surprise inspections by global regulatory agencies such as the FDA and EMA.

[0142] This technical solution transforms compliance management from "passive response" to "active prevention and control" by deeply integrating the pharmaceutical industry's knowledge system and artificial intelligence technology, providing pharmaceutical companies with a full life cycle solution covering R&D, production, and marketing, significantly reducing compliance costs and accelerating the drug launch process.

[0143] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A compliance verification method for pharmaceutical process data, characterized in that: The following steps are involved: S1: Construct a medical process knowledge graph, wherein the medical process knowledge graph includes multiple process stages and data dependency relationships between the process stages, wherein the data dependency relationships represent the data input and output associations of different stages through directed edges; S2: receiving target medical process data, and identifying the current process stage to which the target medical process data belongs based on the medical process knowledge graph; S3: dynamically activating a corresponding compliance verification rule set from a preset rule base according to the current process stage, wherein the preset rule base stores independent rule subsets associated with each process stage; S4: performing multimodal data alignment processing on the target medical process data, including: S41: converting structured data, semi-structured data and unstructured data into vector representations of a unified semantic space through a feature extraction model, respectively, wherein the unstructured data includes medical imaging data and handwritten text data; S42: Calculate the semantic consistency score between different modal data; S5: When the semantic consistency score is lower than a preset threshold, cross-modal conflict detection is triggered, and conflict resolution suggestions are generated based on a preset modality priority rule, where the modality priority rule defines that the priority of medical imaging data is higher than that of handwritten text data; S6: Verify the target medical process data after conflict resolution according to the activated compliance verification rule set, and output the compliance determination result; S7: When a process stage switching event is detected, the corresponding compliance verification rule set is reactivated based on the switched process stage, and the historical process data is retrospectively verified.

2. The method according to claim 1, characterized in that The compliance verification rule set corresponding to the dynamic activation in step S3 further includes: According to the data dependency relationship between the current process stage and the adjacent process stage, some rules of the associated stage are loaded to form a composite rule set. The generation formula of the composite rule set is: R c =R p U{r j |r j ∈R adj ,dep(r j ,R p )>θ}; Where R c is a composite rule set, R p is the rule subset of the current process stage, R adj is a subset of rules for adjacent process stages, dep(r j ,R p ) represents the rule r j With R p The dependency of , θ is the dependency threshold.

3. The method according to claim 2, characterized in that The dependency dep(r j ,R p ) is calculated as: Among them, Dr. j and Dr. k The rules are j and rule r k The collection of dependent data fields.

4. The method according to claim 1, characterized in that: The feature extraction model in step S41 includes: Medical image data uses a pre-trained ResNet-50 model to extract convolutional features and generates a vector representation V through an attention mechanism. img ; After the handwritten text data is sequence recognized using the CRNN model, the semantic vector V is generated through the BiLSTM model. text .

5. The method according to claim 1, characterized in that The method for dynamically adjusting the preset threshold in step S5 includes: Calculate the dynamic threshold T based on the risk level coefficient α of the current process stage d , the calculation formula is: T d =T0×(1+α·△); Wherein T0 is the basic threshold, Δ is the risk level adjustment factor, and the risk level coefficient α is set to 0.8 in the clinical trial stage and to 1.2 in the production stage.

6. The method according to claim 1, characterized in that The triggering conditions for the backtracking check in step S7 include: A process phase switching instruction is detected; A historical data modification event is detected, and the modified data field has a dependency relationship with the currently activated rule set.

7. A compliance verification device for medical process data, characterized in that: include: A process modeling module, used to construct and store a medical process knowledge graph, wherein the medical process knowledge graph includes process stage definitions and data dependencies between stages; A rule dynamic generation module, connected to the process modeling module, is used to activate the corresponding compliance verification rule set multimodal alignment module from the preset rule library according to the current process stage to which the target pharmaceutical process data belongs, including: A structured data conversion unit for converting database records into feature vectors; Image feature extraction unit, which uses pre-trained convolutional neural network to extract deep features of medical images; A text parsing unit, which uses a natural language processing model to extract entities and relationship conflict detection modules from handwritten texts, connected to the multimodal alignment module, for calculating cross-modal semantic consistency scores and identifying data conflicts, wherein the conflict detection module has a built-in modality priority ranking table; a conflict resolution module, connected to the conflict detection module, for automatically selecting valid data or generating a manual review request according to the modality priority ranking table; The backtracking verification module is connected to the process modeling module and the rule dynamic generation module, and is used to trigger the re-verification of historical data when the process stage is switched.

8. The device according to claim 7, characterized in that The rule dynamic generation module further includes: The rule dependency analysis submodule is used to calculate the association weights between rules at different process stages. The calculation formula for the association weights is: Where D i and D j The rules are i and r j The set of dependent data fields; The rule combination optimization submodule is used to dynamically adjust the activated rule set according to the association weight.

9. The device according to claim 7, characterized in that The multimodal alignment module further comprises: Semantic enhancement unit, used to enhance the medical image feature vector V img and text semantic vector V text Perform cross-modal attention fusion to generate a joint vector V fusion , the calculation formula is: Where Q, K, and V are query matrix, key matrix, and value matrix, respectively. img and V text The concatenated linear transformation is generated, d is the vector dimension of the query matrix Q and the key matrix K; T represents the transpose operation of the matrix.

10. The device according to claim 7, characterized in that Also includes: The visual evidence module is connected to the conflict resolution module and the backtracking verification module, and is used to generate an interactive report containing the following contents: Highlighting of conflicting data segments and their associated rule clauses; Visualize the rule coverage before and after the process stage switch; Export evidence packages that meet regulatory audit requirements.

Images