An artificial intelligence-based rwe design automation method and system

Abstract

The application provides an RWE design automation method and system based on artificial intelligence, and belongs to the technical field of artificial intelligence aided design. The method comprises the following steps: analyzing a research target text input by a user through a BioBERT-BiLSTM fusion model, extracting PICO elements and classifying research types; constructing a dynamic medical relationship graph based on medical entity data, and generating a 256-dimensional node feature vector; using an XGBClassifier model trained by a federal learning framework, combining a seven-dimensional feature vector to recommend an optimal research design scheme; calculating a candidate data source adaptation score through a graph attention network, and outputting a radar chart visualization report and a data source combination scheme; based on a regulatory rule knowledge base, using an adversarial verification mechanism to optimize covariant selection, and automatically generating a complete RWE design scheme containing executable code, SAP documents and regulatory checklists. The application realizes the full-process automation of real-world evidence design, effectively improves research efficiency and reduces the artificial error rate.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-aided design technology, and in particular to an AI-based RWE design automation method and system. Background Technology

[0002] In today's data-driven healthcare field, real-world evidence (RWE) has become a core basis for drug development, medical device evaluation, and medical decision-making. Compared with traditional randomized controlled trials, RWE originates from massive amounts of heterogeneous data in real-world medical scenarios, including electronic health records, medical insurance claim databases, disease registry systems, and wearable device monitoring data. Its research conclusions are more likely to reflect actual treatment effects and long-term patient prognoses. However, existing technological systems have many structural flaws in the entire RWE research process design, leading to low research efficiency and insufficient reliability of conclusions, severely restricting the application value of RWE in regulatory decision-making and clinical practice. In the healthcare field, medical data sources are extremely dispersed, including medical records and test reports from hospital information systems, real-time physiological data collected by wearable devices, and patient-reported health logs and follow-up feedback. These data vary greatly in form, ranging from structured code and numerical values ​​to unstructured text descriptions and image files, and even semi-structured tables and forms. In the traditional model, staff need to manually screen and organize this data, which is not only time-consuming and labor-intensive but also prone to data deviation due to human error. More importantly, medical data is closely linked to patient privacy, and how to build a strong security barrier during data transmission, storage, and analysis to prevent information leakage while not affecting the normal flow and use of data has become a long-standing and difficult problem to balance within the industry. In addition, the lack of unified data standards among different medical institutions and the incompatibility of data formats exported from some older systems with existing analysis tools further exacerbate the difficulty of data integration, making the cross-institutional and cross-regional application of RWE extremely challenging. Furthermore, the RWE analysis chain, from data cleaning and feature engineering to model training, involves switching between multiple tools, such as SQL, Python, and specialized statistical software. The lack of an end-to-end automated platform results in research cycles that can last for months, making it difficult to meet regulatory timeliness requirements.

[0003] In view of this, there is an urgent need for an AI-based RWE design automation method and system to at least address the above-mentioned shortcomings. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based automated design process (RWE) method and system to address the problems in existing technologies, such as low efficiency in data integration and processing, difficulty in ensuring data quality and consistency, fragmented processes, lack of collaboration, and low repeatability and standardization. The specific technical solution is as follows:

[0005] This invention provides an AI-based method for automating Redesign for Web Entries (RWE), comprising:

[0006] Step 1: Parse the research target text input by the user using the BioBERT-BiLSTM fusion model, extract PICO elements, and output a structured set of research elements and research type classification results;

[0007] Step 2: Based on the medical entity data in the structured research element set, a graph convolutional network is used to perform node embedding calculation on the knowledge graph to generate a dynamic medical relationship graph containing 256-dimensional feature vectors;

[0008] Step 3: Based on the research type classification results and the seven-dimensional feature vector, recommend the optimal research design scheme type using the XGBClassifier model trained by the federated learning framework;

[0009] Step 4: Calculate the feature matching degree between the structured research element set and the dynamic medical relationship graph, and use a multi-head attention mechanism to output the candidate data source fit score, radar chart visualization report and optimal data source combination scheme;

[0010] Step 5: Based on the optimal research design scheme type and regulatory rule knowledge base, optimize the selection of covariates through adversarial verification mechanism, and automatically generate a complete RWE design scheme containing executable code, SAP documents and regulatory checklists.

[0011] Furthermore, the BioBERT-BiLSTM fusion model in step 1 includes: a BioBERT layer extracting text semantic representations, a BiLSTM layer capturing contextual dependencies, and a CRF layer optimizing sequence labeling results; the PICO elements include: Population, Intervention, Comparator, and Outcome.

[0012] Furthermore, the dynamic medical relationship graph in step 2 is constructed using a two-layer graph convolutional network.

[0013] Furthermore, the seven-dimensional feature vector in step 3 includes: timeliness requirements, sample size, exposure features, result frequency, promiscuous control requirements, cost constraints, and regulatory level; the federated learning framework adopts a differential privacy aggregation algorithm, iterates 5 times locally, and has a learning rate of 0.01.

[0014] Furthermore, the fit score in step 4 is based on four-dimensional evaluation indicators: data integrity, timeliness, quality certification, and privacy compliance; the optimal data source combination scheme includes a primary data source, a supplementary data source, and a verification data source.

[0015] Furthermore, in step 5: the adversarial verification mechanism adopts a GAN architecture: the generator is a multilayer perceptron, the discriminator is a CNN, and the loss function is the Wasserstein distance; the executable code is generated based on the R language, and the survival package and the geepack package are called to implement statistical analysis.

[0016] In a second aspect, the present invention also discloses a system for the aforementioned AI-based RWE design automation method, comprising:

[0017] The structured element extraction module is used to parse and extract PICO elements based on the research target text description input by the user, through the BioBERT-BiLSTM fusion model in the natural language processing module, and output a set of structured research elements and research type classification results.

[0018] The graph generation module is used to generate a dynamic medical relationship graph containing 256-dimensional feature vectors by performing node embedding calculations on the knowledge graph through a graph convolutional network based on the medical entity data in the structured research element set.

[0019] The scheme recommendation module is used to recommend the optimal research design scheme type based on the research type classification results and seven-dimensional feature vectors, using the XGBClassifier model trained by the federated learning framework.

[0020] The data source adaptation module is used to calculate the candidate data source adaptation score based on the feature matching degree between the structured research element set and the dynamic medical relationship graph, through the multi-head attention mechanism of the graph attention network, and output a radar chart visualization report with weighted interpretation and the optimal data source combination scheme.

[0021] The regulatory compliance generation module is used to optimize the selection of covariates and automatically generate a statistical analysis plan that meets regulatory requirements based on the optimal research design scheme type and the regulatory rule knowledge base through an adversarial verification mechanism. It outputs a complete RWE design scheme that includes executable code, SAP documents and regulatory checklists.

[0022] Furthermore, the constraint exception handling mechanism of the data source adaptation module includes:

[0023] When the primary data source score is <0.85, initiate the supplementary data source matching algorithm or redesign the instruction; when the verification data source score is <0.7, initiate the data quality enhancement process: increase the privacy compliance weight to 0.25 and activate the data integrity audit submodule.

[0024] Furthermore, the radar chart visualization report includes:

[0025] The four-quadrant radar chart displays scores for data integrity, timeliness, quality certification, and privacy compliance.

[0026] The weights for each dimension are labeled as percentages, along with the reasons for the matching algorithm's recommendations.

[0027] An electronic device includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method.

[0028] The beneficial effects of this invention are as follows:

[0029] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0031] Figure 1 This is a schematic diagram of an AI-based RWE design automation method in an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of an AI-based RWE design automation system in an embodiment of the present invention. Detailed Implementation

[0033] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0034] Example 1

[0035] This embodiment provides an AI-based method for automating RWE (Research, Design, and Environment) design, such as... Figure 1 As shown, it includes:

[0036] Step 1: Based on the research objective text description input by the user, the BioBERT-BiLSTM fusion model in the natural language processing module is used to parse and extract PICO elements, and determine the structured research element set and research type classification results;

[0037] The research objective text description is: natural language text input by the user describing the clinical research objective, which includes at least one of the following: disease, intervention, study population, and study endpoint.

[0038] The BioBERT-BiLSTM fusion model in the natural language processing module is as follows: it uses a BioBERT model pre-trained in the biomedical field to obtain semantic representations of text, combines a bidirectional long short-term memory network (BiLSTM) to capture contextual dependencies, and finally optimizes the fusion model structure of sequence labeling results through a conditional random field (CRF) layer.

[0039] The PICO elements are the core framework of clinical trial design, including:

[0040] P (Population): The target population of the study refers to a patient group with specific disease characteristics, demographic characteristics, or exposure characteristics.

[0041] I (Intervention): Intervention measures refer to drugs, devices, surgeries, or treatments applied to a study population;

[0042] C (Comparator): The control measure refers to the placebo, standard treatment, or other control condition compared with the intervention.

[0043] O (Outcome): The study endpoint refers to the clinical outcome, safety indicator, or surrogate endpoint used to evaluate the effectiveness of the intervention.

[0044] The structured research element set is a data set that stores the extracted PICO elements in a machine-readable structured format, and its fields include at least: population (string or encoded value), intervention (string or encoded value), comparator (string or encoded value), and outcome (array of strings or array of encoded values).

[0045] The research type classification results are as follows: clinical research design types automatically identified based on research objective descriptions, including at least one of the following types: observational studies (cohort studies, case-control studies, cross-sectional studies), interventional studies (randomized controlled trials, single-arm trials), or real-world evidence studies (prospective registry studies, retrospective database analyses).

[0046] Step 2: Based on the medical entity data in the structured research element set, a graph convolutional network is used to perform node embedding calculations on patient attributes, medical events, and institutional relationships in the knowledge graph, generating a dynamic medical relationship graph with 256-dimensional feature vectors;

[0047] Among them, medical entity data refers to standardized medical concepts extracted from the structured research element set, which include at least disease terms (using ICD-11 coding), drug names (using RxNorm coding), surgical procedures (using CPT coding), and laboratory tests (using LOINC coding).

[0048] The knowledge graph is a representation of medical domain knowledge organized in a graph structure, and its node types include:

[0049] Patient node: Attributes must include at least age, gender, history of underlying diseases, and medication history;

[0050] Medical event nodes: Attributes must include at least the event type (diagnosis / medication / surgery), event time, and clinical description;

[0051] Institution node: The attributes must include at least the type of medical institution (hospital / clinic / pharmacy) and its level information;

[0052] The edge relationship types must include at least four types of medical relationships: "patient-diagnosis-disease", "patient-receive-treatment", "treatment-include-medication", and "disease-cause-complication".

[0053] Among them, the dynamic medical relationship graph is a medical knowledge graph representation with temporal evolution characteristics, and its dynamism is reflected in:

[0054] Time-aware edge connections: The correlation strength between medical events decays over time, with a decay factor set to λ = 0.85 / month;

[0055] Dynamic node attribute updates: The health status attributes of patient nodes are updated in real time as medical events occur;

[0056] The graph structure is updated incrementally every 24 hours, and newly added medical event data is accessed through stream processing.

[0057] The 256-dimensional feature vector is a fixed-dimensional vector representation obtained through node embedding. Each vector element represents the strength value of the node in the corresponding semantic dimension. Nodes with a cosine distance of less than 0.3 between vectors are considered to have strong medical relevance.

[0058] In one embodiment,Step 2: Based on the medical entity data in the structured research element set, a graph convolutional network is used to perform node embedding calculations on patient attributes, medical events, and institutional relationships in the knowledge graph, generating a dynamic medical relationship graph with 256-dimensional feature vectors. This includes a neural network architecture with two graph convolutional layers (GCN layers). The first layer has an input dimension of 128 and an output dimension of 256, using the ReLU activation function; the second layer has an input dimension of 256 and an output dimension of 256, using the Tanh activation function. The adjacency matrix is ​​normalized using a symmetric normalization method. The node embedding calculation involves learning the low-dimensional vector representation of nodes in the graph through the graph convolutional network. The specific calculation formula is as follows: ;

[0059] Where H (l) Let W be the feature matrix of the nodes in the l-th layer. (l) σ is a trainable weight matrix, and σ is an activation function; Step 3: Based on the research type classification results and the seven-dimensional feature vector (timeliness requirements, sample size, exposure characteristics, result frequency, confounding control requirements, cost constraints, and regulatory level), the XGBClassifier model trained by the federated learning framework recommends the optimal research design scheme type;

[0060] The research type classification results are as follows: the clinical research design type classification output obtained from step 1 includes at least one type of observational study (cohort study, case-control study, cross-sectional study), interventional study (randomized controlled trial, single-arm trial) or real-world evidence study (prospective registry study, retrospective database analysis);

[0061] The seven-dimensional feature vector consists of a combination of quantitative indicators representing the constraints of the research design, specifically including:

[0062] Timeliness requirement: The urgency of completing the research within a specified timeframe, quantified on a scale of 1 to 5 (1 indicates no time limit, 5 indicates an urgent need, to be completed within 30 days).

[0063] Sample size: The minimum sample size required for the study, calculated based on statistical power, with a range of [100, 100,000].

[0064] Exposure characteristics: The frequency and pattern of exposure to the intervention, expressed as continuous variables (daily medication frequency × treatment cycle).

[0065] Result frequency: the expected incidence of the study endpoint, ranging from [0.01, 0.5] (representing an incidence rate of 1%-50%).

[0066] Confounding control requirements: The intensity of requirements for controlling confounding factors is divided into three levels: low (Level 1), medium (Level 2), and high (Level 3).

[0067] Cost constraints: Research budget limits (in US$10,000), ranging from [10,500].

[0068] Regulatory Level: Classified according to the stringency of regulatory requirements, Level 1 (non-registration research) to Level 3 (registration application research).

[0069] The federated learning framework is a model training architecture that employs cross-institutional collaboration.

[0070] Among them, the XGBClassifier model is a classification model based on gradient boosting decision trees;

[0071] The optimal study design types are: the recommended study design types output by the model, including: prospective cohort studies (suitable for safety studies with high exposure frequency and high outcome incidence), case-control studies (suitable for rare disease studies with low outcome incidence), target trial simulations (suitable for observational studies requiring simulated RCTs), registry studies (suitable for efficacy studies requiring long-term follow-up), and hybrid design studies (combining prospective and retrospective data). In one embodiment, the federated learning framework is: a model training architecture employing cross-institutional collaboration, including:

[0072] The participants must include at least three local data nodes from medical institutions;

[0073] A security aggregation algorithm based on differential privacy is adopted, with the privacy budget ε set to 1.2;

[0074] Each training round consists of 5 local iterations with a learning rate of 0.01.

[0075] Model parameters are aggregated using a weighted average, with weights allocated proportionally to the amount of data from each institution. The XGBClassifier model is a gradient boosting decision tree-based classification model, including:

[0076] Maximum tree depth: 12 levels

[0077] Learning rate: 0.1

[0078] Subsample proportion: 0.8

[0079] Regularization parameters: λ=1.5, γ=0.3

[0080] Objective function: multi:softprob (multi-class probability output)

[0081] Number of trees: 150.

[0082] Step 4: Based on the feature matching degree between the structured research element set and the dynamic medical relationship graph, calculate the candidate data source fit score through the multi-head attention mechanism (k=8) of the graph attention network, and output a radar chart visualization report with weighted interpretation and the optimal data source combination scheme.

[0083] The structured research element set is a machine-readable structured data set obtained in step 1, whose fields include at least: population (string or encoded value), intervention (string or encoded value), comparator (string or encoded value), and outcome (array of strings or array of encoded values). The dynamic medical relationship graph is a medical knowledge graph representation with temporal evolution characteristics generated in step 2, whose node embedding vector dimension is 256, containing semantic information about patient attributes, medical events, and institutional relationships. The feature matching degree is quantified by using cosine similarity calculation to measure the degree of matching between research elements and data source features; the calculation formula is: ;

[0084] Where A is the feature vector of the research element, B is the feature vector of the data source, and the similarity threshold is set to 0.7; the graph attention network is a neural network that uses the attention mechanism for graph-structured data, and its attention coefficient is calculated using the following formula: ;

[0085] Where h i and h j Let W be the node feature vector, W be the trainable weight matrix, a be the attention parameter vector, and LeakyReLU be the activation function.

[0086] The multi-head attention mechanism (k=8) uses eight independent attention heads to compute attention weights in parallel, and the final output is the concatenation of the outputs from each head. The formula used is as follows:

[0087] ;

[0088] in W represents the normalized attention weight for the k-th attention head. k Let N be the weight matrix specific to the k-th head. i Let be the set of neighbors of node i, σ be the activation function, and k be the concatenation of the outputs of the 8 heads.

[0089] The candidate data source fit score is a comprehensive score based on four evaluation dimensions, specifically including:

[0090] Data integrity (weight 0.35): Key variable coverage, calculated as (number of existing variables / number of required variables) × 100%;

[0091] Timeliness (weight 0.25): Data update frequency score, calculated as the reciprocal of the update delay per day;

[0092] Quality Certification (weight 0.25): Data quality certification level, 1-5 points (1 point: no certification; 5 points: ISO 9001 certification)

[0093] Privacy compliance (weight 0.15): The degree of privacy protection compliance, 1-3 points (1 point: basic compliance; 3 points: GDPR / HIPAA compliance).

[0094] The radar chart visualization report with weighted interpretation is as follows: It uses a four-quadrant radar chart to intuitively display the scores of each data source in four evaluation dimensions. The weight of each dimension is marked as a percentage on the corresponding coordinate axis and includes a textual explanation of the recommendation reason based on the matching algorithm.

[0095] The optimal data source combination scheme is as follows: a combination of the top three data sources ranked by suitability score, including:

[0096] Primary data source: The data source with the highest fit score (score ≥ 0.85); Supplementary data source: Auxiliary data source used to fill in missing variables in the primary data source (score ≥ 0.75); Validation data source: Independent data source used to validate the results (score ≥ 0.7).

[0097] Step 5: Based on the optimal study design type and regulatory rule knowledge base (including FDA 21 CFR Part 11 and ICH E9(R1) requirements), optimize the selection of covariates through adversarial validation mechanism and automatically generate a statistical analysis plan that meets regulatory requirements, outputting a complete RWE design scheme that includes executable code, SAP documentation and regulatory checklist.

[0098] The optimal study design type is: the model recommendation results obtained from step 3, including one of the following: prospective cohort study, case-control study, target trial simulation, registry study, or hybrid design study; the regulatory rule knowledge base is: a structured, machine learning-readable database of regulatory requirements; the adversarial validation mechanism is: a covariate selection optimization algorithm using a generative adversarial network (GAN) architecture; and the covariate selection optimization involves: screening variables that are both relevant to the outcome and irrelevant to treatment allocation through adversarial validation, with the specific criteria being:

[0099] Correlation with outcome: Pearson correlation coefficient ≥ 0.3;

[0100] Independence from treatment allocation: passed the double robustness estimation test, p-value > 0.05;

[0101] The final number of covariates is controlled within the range of 5-15; the statistical analysis plan is automatically generated by a template-based intelligent filling system; the executable code is a complete analysis script generated based on R language (version 4.2.0), which includes:

[0102] Data preprocessing code: performs data cleaning, variable transformation, and format standardization;

[0103] Statistical analysis code: calls the survival package (version 3.4-0) and the geepack package (version 1.3.9).

[0104] Output code: Generates tables and charts that meet journal requirements; The SAP document is a statistical analysis plan conforming to regulatory formats, with a structure including: research overview (purpose, design, population), statistical analysis methods (primary analysis, secondary analysis, subgroup analysis), missing data handling plan, sensitivity analysis plan, and sample size calculation basis. The regulatory checklist is a machine-readable compliance verification document with fields including: clause compliance flags (yes / no / not applicable), evidence location index (pointing to specific line of code or document paragraph), and automatic verification result (pass / fail / requires manual review). The complete RWE design is a standardized package integrating all outputs, with the following file structure:

[0105] / code / analysis_main.R (Main analysis code)

[0106] / docs / SAP_v1.0.pdf (Statistical Analysis Plan)

[0107] / docs / compliance_checklist.json (Regulatory checklist)

[0108] / README.txt (Implementation instructions)

[0109] In one embodiment, the regulatory rules knowledge base is: a structured, machine learning-readable database of regulatory requirements, containing at least:

[0110] FDA 21 CFR Part 11: Compliance requirements for electronic records and electronic signatures, specifically including three sub-modules: system verification (Section 11.10(a)), audit trail (Section 11.10(e)), and electronic signatures (Section 11.50);

[0111] ICH E9(R1) requires: statistical principles for clinical trials, especially including an estimation target framework (five core elements: treatment strategy, population, variables, summary measures, and management of comorbidities).

[0112] EMA Real-World Evidence Guidelines: Data Reliability Assessment Criteria (Completeness, Accuracy, Timeliness, Consistency);

[0113] NMPA Real-World Research Guidelines: Data Provenance and Quality Control Requirements. In one embodiment, the adversarial verification mechanism is a covariate selection optimization algorithm employing a Generative Adversarial Network (GAN) architecture, specifically including:

[0114] Generator: Multilayer perceptron (hidden layer dimensions [256,128,64]), used to generate covariate combinations;

[0115] Discriminator: Convolutional Neural Network (3×3 kernel size, stride 1), used to determine whether the combination of covariates leads to the prediction of treatment group assignment;

[0116] Loss function: Wasserstein distance, gradient penalty coefficient λ=10;

[0117] Training iterations: 1000 rounds, batch size: 32.

[0118] In one embodiment, the statistical analysis plan is automatically generated based on a template-based intelligent fill system, which includes:

[0119] Core statistical methods module: Automatically selected based on research type, such as Cox proportional hazards model, logistic regression, and generalized estimation equations;

[0120] Missing data processing module: Employs multiple imputation (m=5 imputation times) or maximum likelihood estimation;

[0121] Sensitivity Analysis Module: Includes 3 alternative model configurations, specifically changing covariates, changing model parameters, and using different estimation methods;

[0122] Sample size calculation module: Calculated based on a two-sided test with power=0.9 and α=0.05.

[0123] AI-based RWE design automation systems, such as Figure 2 As shown, it includes:

[0124] The structured element extraction module is used to parse and extract PICO elements based on the research target text description input by the user, through the BioBERT-BiLSTM fusion model in the natural language processing module, and output a set of structured research elements and research type classification results.

[0125] The graph generation module is used to generate a dynamic medical relationship graph containing 256-dimensional feature vectors by performing node embedding calculations on the knowledge graph through a graph convolutional network based on the medical entity data in the structured research element set.

[0126] The scheme recommendation module is used to recommend the optimal research design scheme type based on the research type classification results and seven-dimensional feature vectors, using the XGBClassifier model trained by the federated learning framework.

[0127] The data source adaptation module is used to calculate the candidate data source adaptation score based on the feature matching degree between the structured research element set and the dynamic medical relationship graph, through the multi-head attention mechanism (k=8) of the graph attention network, and output a radar chart visualization report with weighted interpretation and the optimal data source combination scheme.

[0128] The regulatory compliance generation module, based on the optimal research design scheme type and the regulatory rule knowledge base, optimizes covariate selection through an adversarial verification mechanism and automatically generates a statistical analysis plan that meets regulatory requirements, outputting a complete RWE design scheme including executable code, SAP documentation, and a regulatory checklist; the data source adaptation module triggers a constraint anomaly handling mechanism based on feature matching degree, including:

[0129] When the primary data source fit score is <0.85, the first type of constraint exception handling is initiated: the supplementary data source matching algorithm is called to recalculate the fit; if the supplementary data source score is ≥0.75, the optimal data source combination scheme is updated; if the supplementary data source score is <0.75, a redesign instruction is sent to the scheme recommendation module.

[0130] When the data source fit score is less than 0.7, initiate the second type of constraint anomaly handling: Based on the regulatory rule knowledge base, initiate a data quality enhancement process, including:

[0131] 1) Increase the privacy compliance weight to 0.25;

[0132] 2) Activate the data integrity audit submodule;

[0133] 3) Recalculate the fitness score until it is ≥0.7;

[0134] The adversarial verification mechanism of the regulatory compliance generation module includes dual anomaly interception:

[0135] Level 1 interception: When the covariate is correlated with the outcome (Pearson coefficient < 0.3) or independent of treatment allocation (p-value ≤ 0.05), the generator retraining process is initiated.

[0136] Update the generator's hidden layer dimension configuration to [512, 256, 128];

[0137] Increase the number of training iterations to 1500.

[0138] Second-level interception: When the number of covariates exceeds the range of [5,15], the regularization enhancement mechanism is triggered.

[0139] Adjust the discriminant's convolution kernel size to 5×5;

[0140] Increase the gradient penalty coefficient to λ=15;

[0141] The solution recommendation module implements a federated learning exception circuit breaker strategy:

[0142] If the number of participating data nodes is less than 3, then perform the following steps:

[0143] 1) Generate synthetic medical data using GAN;

[0144] 2) Privacy budget ε decreased to 0.8;

[0145] 3) The learning rate is adjusted to 0.005.

[0146] If the confidence level of the XGBClassifier model output is less than 90%, the hybrid design degradation mechanism is triggered: 1) reduce the maximum depth of the tree to 8 layers; 2) increase the number of trees to 200; 3) enable Bayesian optimization to retrain the model.

[0147] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based method for automating RWE design, characterized in that, include: Step 1: Parse the research target text input by the user through the BioBERT-BiLSTM fusion model, extract PICO elements and output a structured set of research elements and research type classification results. The BioBERT-BiLSTM fusion model in Step 1 adopts the BioBERT model pre-trained in the biomedical field to obtain the semantic representation of the text, and combines the fusion model structure of BiLSTM to capture contextual dependencies and CRF layer to optimize sequence labeling results. The PICO elements include: Population (target population), Intervention (intervention), Comparator (control measure), and Outcome (study endpoint). Step 2: Based on the medical entity data in the structured research element set, a graph convolutional network is used to perform node embedding calculation on the knowledge graph to generate a dynamic medical relationship graph containing 256-dimensional feature vectors; Step 3: Based on the research type classification results and the seven-dimensional feature vector, recommend the optimal research design scheme type using the XGBClassifier model trained by the federated learning framework; Step 4: Calculate the feature matching degree between the structured research element set and the dynamic medical relationship graph, and use a multi-head attention mechanism to output the candidate data source fit score, radar chart visualization report and optimal data source combination scheme; Step 5: Based on the optimal research design scheme type and regulatory rule knowledge base, optimize the selection of covariates through adversarial verification mechanism, and automatically generate a complete RWE design scheme containing executable code, SAP documents and regulatory checklists.

2. The AI-based RWE design automation method as described in claim 1, characterized in that, The dynamic medical relationship graph described in step 2 is constructed using a two-layer graph convolutional network.

3. The AI-based RWE design automation method as described in claim 1, characterized in that, The seven-dimensional feature vector in step 3 includes: timeliness requirements, sample size, exposure features, result frequency, promiscuous control requirements, cost constraints, and regulatory level; the federated learning framework adopts a differential privacy aggregation algorithm, iterates 5 times locally, and has a learning rate of 0.

01.

4. The AI-based RWE design automation method as described in claim 1, characterized in that, The fit score in step 4 is based on four-dimensional evaluation indicators: data integrity, timeliness, quality certification, and privacy compliance; the optimal data source combination scheme includes a primary data source, a supplementary data source, and a verification data source.

5. The AI-based RWE design automation method as described in claim 1, characterized in that, In step 5: the adversarial verification mechanism adopts a GAN architecture: the generator is a multilayer perceptron, the discriminator is a CNN, and the loss function is the Wasserstein distance; the executable code is generated based on R language, and the survival package and geepack package are called to implement statistical analysis.

6. The AI-based RWE design automation method as described in claim 1, characterized in that, The radar chart visualization report includes: The four-quadrant radar chart displays scores for data integrity, timeliness, quality certification, and privacy compliance. The weights for each dimension are labeled as percentages, along with the reasons for the matching algorithm's recommendations.

7. An AI-based RWE design automation system for implementing the AI-based RWE design automation method as described in any one of claims 1-6, characterized in that, include: The structured element extraction module is used to parse and extract PICO elements based on the research target text description input by the user through the BioBERT-BiLSTM fusion model in the natural language processing module, and output the structured research element set and research type classification results. The BioBERT-BiLSTM fusion model in step 1 adopts the text semantic representation obtained by the BioBERT model pre-trained in the biomedical field, and combines the fusion model structure of BiLSTM capturing contextual dependencies and CRF layer optimizing sequence labeling results. The PICO elements include: Population (target population), Intervention (intervention), Comparator (control measure), and Outcome (study endpoint). The graph generation module is used to generate a dynamic medical relationship graph containing 256-dimensional feature vectors by performing node embedding calculations on the knowledge graph through a graph convolutional network based on the medical entity data in the structured research element set. The scheme recommendation module is used to recommend the optimal research design scheme type based on the research type classification results and seven-dimensional feature vectors, using the XGBClassifier model trained by the federated learning framework. The data source adaptation module is used to calculate the candidate data source adaptation score based on the feature matching degree between the structured research element set and the dynamic medical relationship graph, through the multi-head attention mechanism of the graph attention network, and output a radar chart visualization report with weighted interpretation and the optimal data source combination scheme. The regulatory compliance generation module is used to optimize the selection of covariates and automatically generate a statistical analysis plan that meets regulatory requirements based on the optimal research design scheme type and the regulatory rule knowledge base through an adversarial verification mechanism. It outputs a complete RWE design scheme that includes executable code, SAP documents and regulatory checklists.

8. The AI-based RWE design automation system as described in claim 7, characterized in that, The constraint exception handling mechanism of the data source adaptation module includes: When the primary data source score is <0.85, initiate the supplementary data source matching algorithm or redesign the instruction; when the verification data source score is <0.7, initiate the data quality enhancement process: increase the privacy compliance weight to 0.25 and activate the data integrity audit submodule.

9. An electronic device, comprising a processor and a memory, characterized in that, The memory stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.

Images