A Drug Lifecycle Risk Monitoring and Early Warning Method Based on Multi-Source Data Fusion

Abstract

This invention discloses a method for monitoring and early warning of drug risks throughout their entire lifecycle based on multi-source data fusion, belonging to the field of data analysis technology. The method includes: constructing a time-series heterogeneous graph and using a contrastive learning algorithm to calculate the similarity between drug and side-effect pairs in the graph, generating a candidate signal list; constructing a dataset and using a separation representation learning algorithm to identify confounding factors in the dataset that correlate drug selection with side-effect risk, obtaining an average causal effect value; drawing a spatiotemporal risk heatmap, classifying the risks throughout the drug's lifecycle based on the risk distribution characteristics in the heatmap, and triggering corresponding early warning strategies; and converting the early warning strategies into a standard-format risk assessment report using a pre-set report template. This invention quantifies the causal relationship between drugs and their side effects by accurately calculating individual and average causal effect values, laying a solid foundation for in-depth assessment of drug safety.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a method for monitoring and early warning of risks throughout the entire life cycle of pharmaceuticals based on multi-source data fusion. Background Technology

[0002] Adverse drug reaction (ADR) monitoring is a crucial link in ensuring drug safety and supporting risk management throughout the entire drug lifecycle. Current technology primarily relies on self-reporting mechanisms submitted by medical institutions, pharmaceutical companies, and patients to identify suspected ADRs. A typical process involves clinical staff identifying a suspected adverse reaction during treatment, manually filling out a standardized report form, which is then reviewed by the hospital's pharmacy department and uploaded to the adverse drug reaction monitoring system. However, this traditional monitoring model suffers from several technical bottlenecks and struggles to meet the urgent needs of modern drug safety supervision for timeliness, comprehensiveness, and intelligence.

[0003] First, the existing mechanism suffers from significant monitoring lag. The average cycle from the occurrence of an adverse reaction to its final entry into the regulatory database is typically 2–4 weeks, making it difficult to capture major risk signals in a timely manner. For example, in the case of a severe liver injury event caused by a certain antibiotic, it took more than 6 months from the first suspected case being reported to the regulatory authorities issuing a risk warning, severely delaying the risk intervention window. Second, underreporting is a prominent issue. Studies show that the overall ADR reporting rate of medical institutions is less than 10% of the actual occurrences, resulting in severely insufficient representativeness of the monitoring data and an inability to accurately reflect the safety risks in drug use. Third, the data sources are highly limited. The current system mainly relies on electronic medical records (EMR) and spontaneously reported data, failing to effectively integrate multi-dimensional real-world data (RWD) from social media, patient forums, medical insurance claim records, drug sales logs, etc., thus limiting the breadth of risk signal coverage and the diversity of scenarios. Fourth, data analysis capabilities are weak. Existing processes rely heavily on manual review and experience-based judgment, lacking automated and intelligent risk identification and signal mining tools, making it difficult to efficiently process massive amounts of data. For example, the FDA's FAERS database adds more than 2 million new reports annually. Relying solely on manual screening is not only inefficient but also prone to introducing subjective bias.

[0004] Furthermore, although some pharmaceutical companies have attempted to build internal safety monitoring systems, their technical architectures are generally limited to single or a few data sources (such as integrating only EMR and sales data). They have failed to achieve standardized access and deep integration of cross-platform, heterogeneous, multi-source real-world data. For example, they have failed to effectively integrate key information channels such as medical literature (e.g., PubMed), patient self-report platforms, and social networks, resulting in narrow risk signal identification dimensions and insufficient sensitivity. More critically, existing systems generally lack the analytical capabilities based on causal inference, making it difficult to distinguish the correlation and causal relationship between drugs and adverse events. This easily leads to false or invalid warnings, seriously affecting the accuracy and clinical applicability of risk warnings.

[0005] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a drug lifecycle risk monitoring and early warning method based on multi-source data fusion, in order to overcome the aforementioned technical problems existing in the current related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] A drug lifecycle risk monitoring and early warning method based on multi-source data fusion, comprising:

[0009] S1. Preprocess the collected multi-source data, construct a time-series heterogeneous graph based on the preprocessing results, and use a contrastive learning algorithm to calculate the similarity between drug and side effect pairs in the time-series heterogeneous graph to generate a candidate signal list.

[0010] S2. Based on the candidate signal list and related covariates, construct a dataset and use the separation representation learning algorithm to identify confounding factors in the dataset that are associated with drug selection and side effect risk. Calculate the individual causal effect value for each patient based on the confounding factors to obtain the average causal effect value.

[0011] S3. Combining individual causal effect values, average causal effect values, and related covariates, draw a spatiotemporal risk heat map. Based on the risk distribution characteristics in the spatiotemporal risk heat map, classify the risks in the entire life cycle of the drug and trigger corresponding early warning strategies.

[0012] S4. Use the preset report template to convert the early warning strategy into a standard format risk assessment report, adjust the format of the risk assessment report according to the drug regulatory requirements, and output the adjusted risk assessment report according to the preset format.

[0013] Furthermore, the collected multi-source data is preprocessed, a temporal heterogeneity graph is constructed based on the preprocessing results, and a contrastive learning algorithm is used to calculate the similarity between drug and side effect pairs in the temporal heterogeneity graph, generating a candidate signal list including:

[0014] S11. Collect multi-source data of patients from multi-source heterogeneous platforms and map the multi-source data into a unified standardized format;

[0015] S12. Based on the standardized multi-source data, the distance algorithm is used to remove duplicate data and clean outliers. Differential privacy technology is used to perform privacy protection processing on the cleaned multi-source data. Based on the privacy protection processing results, a time-series heterogeneous graph containing drugs, side effects and time-series relationships is constructed.

[0016] S13. Use a contrastive learning algorithm to learn the temporal heterogeneous graph, calculate the similarity between drug and side effect pairs in the temporal heterogeneous graph, and filter the drug and side effect pairs based on the calculation results to generate a candidate signal list.

[0017] Furthermore, a contrastive learning algorithm is used to learn the temporal heterogeneity graph, calculate the similarity between drug and side effect pairs in the graph, and filter the drug and side effect pairs based on the calculation results to generate a candidate signal list, including:

[0018] S131. Hierarchical attention mechanism is used to capture the multidimensional dynamic correlation features between drugs, side effects and covariates over time, and contrastive learning algorithm is used to optimize the discriminativeness of graph structure embedding vectors.

[0019] S132. Using a cross-modal semantic fusion algorithm, semantically align the optimized graph structure embedding vector with the corresponding text embedding vector to generate a cross-modal embedding vector.

[0020] S133. Based on the generated cross-modal embedding vectors, calculate the Gaussian kernel embedding similarity of drug and side effect pairs in the graph structure space and the semantic similarity in the text semantic space. Based on the calculation results, use an adaptive threshold learning strategy to dynamically screen drug and side effect pairs and construct a candidate set.

[0021] S134. Analyze the candidate set using an association rule mining algorithm, calculate the support and confidence of each drug-side effect pair, and combine the preset support and confidence thresholds to select drug-side effect pairs that simultaneously meet the support and confidence thresholds, generating a candidate signal list.

[0022] Furthermore, based on the candidate signal list and related covariates, a dataset is constructed, and a separation representation learning algorithm is used to identify confounding factors in the dataset that associate drug selection with side effect risk. Based on these confounding factors, the individual causal effect value for each patient is calculated, yielding the average causal effect value, which includes:

[0023] S21. Based on the generated candidate signal list, divide the patients into a medication group and a non-medication group, and construct a dataset by combining covariates related to medication use and side effects.

[0024] S22. Decouple the constructed dataset to the prior space of potential factors, and combine it with the preset risk constraints to identify confounding factors that are associated with drug selection and side effect risk.

[0025] S23. Based on the identified confounding factors, using the representational balance technique and the counterfactual reasoning framework, predict the side effects of each patient taking the medication and not taking the medication, and calculate the individual causal effect value for each patient based on the prediction results.

[0026] S24. Statistically aggregate the individual causal effect values ​​of all patients to calculate the average causal effect value of the drug and side effects.

[0027] Furthermore, the constructed dataset is decoupled to the prior space of latent factors, and combined with pre-defined risk constraints, to identify confounding factors that correlate drug selection with side effect risk, including:

[0028] S221. The dataset is processed using a generative adversarial interpolation algorithm, and a joint input matrix containing drug selection, side effect risk, and covariates is constructed based on the processing results.

[0029] S222. Based on the pre-acquired medical knowledge, the latent factors in the joint input matrix are divided into dimensions. Based on the division results, the total dimension of the latent factors and the proportion of each type of latent factor in the total dimension are allocated, and a corresponding prior space is defined for each latent factor.

[0030] S223. Based on the defined latent factor prior space, design a three-level network model including encoder, decoupler and decoder, and train the three-level network model using the separation representation algorithm.

[0031] S224. Based on the preset risk constraints, construct a multi-objective loss function and use the multi-objective loss function to optimize the trained three-level network model, and output the latent factor matrix.

[0032] S225. Calculate the correlation strength score of each latent factor in the latent factor matrix, and filter the calculation results. Based on the filtering results, use the correlation analysis algorithm to analyze the correlation strength between the drug and side effects, and output the confounding factor.

[0033] Furthermore, the potential factors in the joint input matrix are dimensionally divided into: treatment factors related to drug selection, outcome factors related to side effect risk, and confounding factors that link drug selection and side effect risk.

[0034] Furthermore, based on the identified confounding factors, the characterization balance technique and counterfactual reasoning framework were used to predict the side effects of each patient taking the medication and not taking the medication, respectively. Based on the prediction results, the individual causal effect value for each patient was calculated, including:

[0035] S231. Using the identified confounding factors as constraints, construct a balanced representation learning model, and optimize the balanced representation learning model using a two-branch adversarial training mechanism. Generate a balanced representation vector based on the optimization results.

[0036] S232. Based on the generated balanced representation vector, construct an outcome regression model to simulate the side effects of taking the medication and not taking the medication, respectively.

[0037] S233. For each patient, the counterfactual reasoning framework is used to call the outcome regression model to predict the side effects of taking the medication and not taking the medication, and the difference between the side effects of taking the medication and not taking the medication is calculated according to different side effect types to obtain the individual causal effect value of the patient.

[0038] Furthermore, the identified confounding factors are used as constraints to construct a balanced representation learning model, and a two-branch adversarial training mechanism is used to optimize the balanced representation learning model. Based on the optimization results, balanced representation vectors are generated, including:

[0039] S2311. Using the identified confounding factors as constraints, design a neural network model containing an input layer, a hidden layer, and an output layer as a balanced representation learning model.

[0040] S2312. Construct a bipartite network architecture and iteratively optimize the bipartite network architecture through an adversarial training strategy until the bipartite network architecture converges to a preset threshold, thus completing the optimization of the balanced representation learning model.

[0041] S2313. Using the identified confounding factors as input and whether the patient takes medication as output, train a pre-built propensity score model, and use the trained propensity score model to calculate the propensity score for each patient, thereby obtaining the conditional probability of the patient receiving medication under the confounding factor conditions.

[0042] S2314. Convert the obtained conditional probabilities into corresponding weighting coefficients, and use the weighting coefficients to adjust the optimized balanced representation learning model. Output the balanced representation vector based on the weighted balanced representation learning model.

[0043] Furthermore, a bi-branch network architecture is constructed, and iterative optimization of the bi-branch network architecture is performed through an adversarial training strategy until the bi-branch network architecture converges to a preset threshold, thus completing the optimization of the balanced representation learning model, including:

[0044] S23121. Based on the designed balanced representation learning model, construct a dual-branch network architecture that includes a representation optimization branch and a hybrid discrimination branch.

[0045] S23122. Using a course learning strategy, the dual-branch network architecture is divided into multiple stages, and adversarial training is performed in each stage.

[0046] S23123. Based on the adversarial training process, a federated alternating update protocol is designed to perform collaborative optimization in a multi-center environment. The convergence status of the representation optimization branch is monitored through Nash equilibrium. When the representation optimization branch converges to a preset threshold, the optimized representation optimization branch is output to complete the optimization of the balanced representation learning model.

[0047] Furthermore, by combining individual causal effect values, average causal effect values, and related covariates, a spatiotemporal risk heat map is drawn. Based on the risk distribution characteristics in the spatiotemporal risk heat map, the risks throughout the entire life cycle of the drug are classified and corresponding early warning strategies are triggered, including:

[0048] S31. Collect relevant spatiotemporal data based on individual causal effect values, average causal effect values, and related covariates, and construct a spatiotemporal dataset.

[0049] S32. Based on the constructed spatiotemporal dataset, calculate the comprehensive risk value of the average causal effect value and the individual causal effect value, and use the preset spatiotemporal diffusion model to simulate the propagation and evolution of the comprehensive risk value in the time and space dimensions, and draw a spatiotemporal risk heat map based on the simulation results.

[0050] S33. Use a spatiotemporal convolutional network to extract risk distribution features from the drawn spatiotemporal risk heat map, and use an attention mechanism to weight the extracted risk distribution features in order to predict future risk change trends.

[0051] S34. Based on the predicted risk change trend, design a dynamic threshold optimization strategy based on reinforcement learning, and use the dynamic threshold optimization strategy to classify the risk level in the entire life cycle of the drug, and trigger the corresponding early warning strategy based on the classification result.

[0052] The beneficial effects of this invention are as follows:

[0053] 1. This invention constructs a time-series heterogeneous graph by standardizing and preprocessing multi-source data, and accurately generates a candidate signal list by applying a contrastive learning algorithm. At the same time, it uses a separation representation learning method to identify confounding factors that affect drug selection and side effect risk, thereby accurately calculating individual and average causal effect values, and quantifying the causal relationship between drugs and their potential side effects, laying a solid foundation for in-depth evaluation of drug safety.

[0054] 2. This invention combines individual and average causal effect values ​​to create a spatiotemporal risk heat map, which visually displays the distribution of drug risks in different times and spaces. Based on the risk patterns shown in the heat map, it automatically classifies risk levels and triggers corresponding early warning mechanisms. This enables regulatory authorities to monitor risk changes in real time, respond quickly, reduce the threat of drug risks to public health, and ensure medication safety.

[0055] 3. This invention utilizes a preset template to transform customized early warning strategies into standardized risk assessment reports, and flexibly adjusts the report content and format according to regulatory needs, thereby ensuring the report's standardization, completeness, and readability, meeting the needs of diverse regulatory scenarios, and thus helping to improve the quality and efficiency of regulatory decisions and strengthen the risk management capabilities throughout the entire drug lifecycle. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a drug lifecycle risk monitoring and early warning method based on multi-source data fusion according to an embodiment of the present invention. Detailed Implementation

[0058] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0059] According to embodiments of the present invention, a method for monitoring and early warning of drug risks throughout the entire life cycle based on multi-source data fusion is provided.

[0060] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a drug life-cycle risk monitoring and early warning method based on multi-source data fusion includes:

[0061] S1. Preprocess the collected multi-source data, construct a time-series heterogeneous graph based on the preprocessing results, and use a contrastive learning algorithm to calculate the similarity between drug and side effect pairs in the time-series heterogeneous graph to generate a candidate signal list.

[0062] In this optional embodiment, the collected multi-source data is preprocessed, a temporal heterogeneity graph is constructed based on the preprocessing results, and a contrastive learning algorithm is used to calculate the similarity between drug and side effect pairs in the temporal heterogeneity graph, generating a candidate signal list including:

[0063] S11. Collect multi-source data from patients from a multi-source heterogeneous platform and map the multi-source data into a unified standardized format.

[0064] It should be further explained that the system interfaces with the hospital's EMR system via the HL7 FHIR API to collect basic patient information, medication records, laboratory test results (such as liver function indicators), and diagnostic codes (ICD-10). It also supports access to relational databases such as MySQL and PostgreSQL. Regarding regulatory data, it periodically crawls the FDA FAERS and NMPA adverse drug reaction databases, using XML parsing technology to extract reporter information, suspected drugs, and adverse reaction terms (MedDRA codes). Social media data is collected through a distributed web crawler developed based on the Scrapy framework, targeting topics related to drug side effects, medication experience notes, and adverse reaction-related content. Selenium is used to simulate login and dynamically load information. Academic literature data is accessed via PubMed. API (E-utilities) searches for literature related to "drug name + adverse reaction" from the past 5 years, extracting MeSH terms and author conclusions from the abstracts; medical insurance data is obtained by connecting to the National Healthcare Security Administration's DRG / DIP database, acquiring structured data such as drug reimbursement records, length of hospital stay, and complication diagnosis, thereby enabling structured extraction of social media data and improving the timeliness of signal recognition by 80%. For example, redness and swelling signals at the injection site of a certain biological agent can be detected 45 days in advance through monitoring on a social media platform, which is faster than traditional EMR systems.

[0065] S12. Based on the standardized multi-source data, the distance algorithm is used to remove duplicate data and clean outliers. Differential privacy technology is used to perform privacy protection processing on the cleaned multi-source data. Based on the privacy protection processing results, a time-series heterogeneous graph containing drugs, side effects and time-series relationships is constructed.

[0066] It should be further explained that the CDISC standard is used to map multi-source data into a unified format. Specifically, drug names are mapped to WHO ATC codes (e.g., aspirin is mapped to B01AC06), adverse reaction terms are mapped to MedDRA SOC / PT hierarchical codes (e.g., rash is mapped to 10028594), and patient features are mapped to the HL7 FHIR Patient resource format. In the deduplication and cleaning process, duplicate text (e.g., repeated posts by patients) is removed based on the Levenshtein distance algorithm, and key information is extracted using regular expressions (e.g., extracting the drug as XX, the symptom as headache, and the time as 3 days from "headache occurred 3 days after taking XX drug"). In terms of privacy protection, differential privacy technology is used to add Gaussian noise to process sensitive information such as patient IDs and mobile phone numbers to meet HIPAA / GDPR requirements. In addition, a temporal heterogeneous graph containing drug nodes, side effect nodes, and medication-side effect occurrence time-series edges is constructed from the processed data to present the relationship between the three.

[0067] S13. Use a contrastive learning algorithm to learn the temporal heterogeneous graph, calculate the similarity between drug and side effect pairs in the temporal heterogeneous graph, and filter the drug and side effect pairs based on the calculation results to generate a candidate signal list.

[0068] In this optional embodiment, a contrastive learning algorithm is used to learn the temporal heterogeneity graph, calculate the similarity between drug and side effect pairs in the temporal heterogeneity graph, and filter the drug and side effect pairs based on the calculation results to generate a candidate signal list, including:

[0069] S131. Hierarchical attention mechanism is used to capture the multidimensional dynamic correlation features between drugs, side effects and covariates over time, and contrastive learning algorithm is used to optimize the discriminative power of graph structure embedding vectors.

[0070] It should be further explained that the goal is to capture the multidimensional dynamic correlation features between drugs, side effects, and covariates over time. For example, this involves analyzing the regulatory role of covariates such as age and concomitant medications on the probability of side effects during different stages of medication (such as the initiation, maintenance, and discontinuation phases), and then modeling the temporal dependency relationship among the three. Furthermore, a contrastive learning algorithm (such as a scheme based on the InfoNCE loss function) is introduced to optimize the embedding vectors obtained from graph structure learning, thereby improving the distinguishability of different types of associations in the vector space.

[0071] S132. Using a cross-modal semantic fusion algorithm, the optimized graph structure embedding vector and the corresponding text embedding vector are semantically aligned to generate a cross-modal embedding vector.

[0072] It should be further explained that by applying cross-modal semantic fusion algorithms (such as attention fusion mechanisms), the optimized graph structure embedding vector is semantically aligned with the text embedding vector generated by the BERT model from the drug instructions and side effect description text, generating a cross-modal embedding vector that combines structural association and textual semantic information, making up for the lack of information in a single modality and achieving multi-dimensional complementarity.

[0073] S133. Based on the generated cross-modal embedding vectors, calculate the Gaussian kernel embedding similarity of drug and side effect pairs in the graph structure space and the semantic similarity in the text semantic space. Based on the calculation results, use an adaptive threshold learning strategy to dynamically screen drug and side effect pairs and construct a candidate set.

[0074] S134. Analyze the candidate set using an association rule mining algorithm, calculate the support and confidence of each drug-side effect pair, and combine the preset support and confidence thresholds to select drug-side effect pairs that simultaneously meet the support and confidence thresholds, generating a candidate signal list.

[0075] It should be further explained that, based on cross-modal embedding vectors, the Gaussian kernel embedding similarity (kernel bandwidth set to 0.5) and cosine similarity of drug-side effect pairs in graph structure space and text semantic space are calculated respectively. An adaptive threshold learning strategy (the threshold range of 0.3-0.8 is dynamically adjusted according to the data distribution) is used to dynamically filter and construct a candidate set, thereby improving the accuracy of the initial screening and avoiding the problems of missed detections or false detections caused by fixed thresholds. The Apriori association rule mining algorithm is used on the candidate set to calculate the support (minimum preset threshold of 0.01) and confidence (minimum preset threshold of 0.8) of each drug-side effect pair. For example, the support of ibuprofen-gastric bleeding pair is 0.02 (i.e., 2% of ibuprofen users report this side effect) and the confidence is 0.85 (i.e., 85% of gastric bleeding-related reports involve ibuprofen). Finally, drug-side effect pairs that meet both threshold conditions are selected to generate a candidate signal list, providing high-quality data support for the accurate confirmation of drug-side effect associations in the future.

[0076] S2. Based on the candidate signal list and related covariates, construct a dataset and use the separation representation learning algorithm to identify confounding factors in the dataset that are associated with drug selection and side effect risk. Calculate the individual causal effect value for each patient based on the confounding factors to obtain the average causal effect value.

[0077] In this optional embodiment, a dataset is constructed based on a list of candidate signals and related covariates. A separation representation learning algorithm is then used to identify confounding factors in the dataset that correlate drug selection with side effect risk. Based on these confounding factors, the individual causal effect value for each patient is calculated, yielding the average causal effect value, which includes:

[0078] S21. Based on the generated candidate signal list, divide the patients into a medication group and a non-medication group, and construct a dataset by combining covariates related to medication use and side effects.

[0079] It should be further explained that the target drug was identified based on the candidate signal list. Using patients' electronic health records as the data source, patients were grouped according to whether they took the target drug (≥7 days) during a continuous 90-day observation period. The drug-taking group included patients who had clearly used the drug, while the non-drug-taking group was selected by propensity score matching (1:2 ratio) for patients who had not used the target drug or similar alternative drugs and who did not differ significantly from the drug-taking group in terms of covariates such as gender, age group, and number of underlying diseases. Subsequently, drug usage details, side effects, and covariate data were collected from both groups. Missing values ​​were imputed five times to construct a structured dataset, which accurately identified the research subjects, reduced confounding bias, and provided high-quality and balanced data support for subsequent causal association verification.

[0080] S22. Decouple the constructed dataset to the prior space of potential factors, and combine it with preset risk constraints to identify confounding factors that relate drug selection and side effect risk.

[0081] In this optional embodiment, the constructed dataset is decoupled to the prior space of potential factors, and combined with preset risk constraints, the confounding factors associated with drug selection and side effect risk are identified, including:

[0082] S221. The dataset is processed using a generative adversarial interpolation algorithm, and a joint input matrix containing drug selection, side effect risk, and covariates is constructed based on the processing results.

[0083] It should be further explained that the dataset was preprocessed by converting categorical variables (such as gender, drug dosage form, and underlying disease type) into numerical values ​​using one-hot encoding, and normalizing continuous variables (such as age, drug dosage, and liver and kidney function indicators) to the [0,1] interval. For missing data in the dataset, if the missing rate is ≤20%, a generative adversarial imputation algorithm was used: a model was constructed containing a generator (a 3-layer MLP structure with 128, 64, and 32 hidden layer neurons respectively) and a discriminator (a 2-layer MLP structure with 64 and 32 hidden layer neurons). This model was trained for 1000 epochs using the Adam optimizer (learning rate 0.001). The generator predicted missing values ​​based on the observed data distribution, and the discriminator distinguished between true values ​​and imputed values. The values ​​are trained adversarially to make the imputed values ​​fit the original data distribution (MAE ≤ 0.05 after imputation). Based on the imputed dataset, core feature dimensions are extracted, including drug selection features (1 for taking the target drug, 0 for not taking it, including the identifier of similar drug substitution), side effect risk features (whether the side effect corresponding to the candidate signal occurs is a binary variable, and the severity of occurrence is quantified as a continuous feature of 0-3 points), and covariate features (demographic feature encoding values, number of combined drug types, normalized indicators of liver and kidney function, quantitative values ​​of lifestyle habits, etc.). According to the rule of "1 patient corresponds to 1 row of features", the three types of features are horizontally concatenated to construct a joint input matrix with the number of columns equal to the total feature dimension (e.g., 80 columns) and the number of rows equal to the total number of patients.

[0084] S222. Based on the pre-acquired medical knowledge, the latent factors in the joint input matrix are divided into dimensions. Based on the division results, the total dimension of the latent factors and the proportion of each type of latent factor in the total dimension are assigned, and a corresponding prior space is defined for each latent factor.

[0085] In this optional embodiment, the dimensional division of the potential factors in the joint input matrix includes: treatment factors related to drug selection, outcome factors related to side effect risk, and confounding factors that associate drug selection with side effect risk.

[0086] It should be further explained that, based on pre-acquired medical knowledge (such as drug mechanisms of action, pathological pathways of side effects, and a list of common clinical confounding factors), the potential factors in the joint input matrix are clearly dimensionally divided. Processing factors focus on drug selection-related characteristics (such as drug type, dosage level, and duration of use), outcome factors revolve around side effect risk-related characteristics (such as whether side effects occur, their severity, and duration), and confounding factors cover covariate characteristics that link the two (such as age group, underlying disease type, number of concurrent medications, and liver and kidney function levels). This is combined with clinical data distribution and model building. The requirements are set at a total of 60 dimensions for potential factors, with the proportions allocated as follows: 10% for treatment factors (6 dimensions), 30% for outcome factors (18 dimensions), and 60% for confounding factors (36 dimensions) to ensure sufficient dimensions for core related features. At the same time, a targeted prior space is defined for each potential factor. The treatment factor adopts a Bernoulli prior (parameter p=0.3, which is consistent with the average usage rate of the target drug in clinical practice), the outcome factor adopts a Gaussian prior (mean μ=0, variance σ²=1, which is consistent with the quantitative distribution of side effect risk), and the confounding factor adopts a Dirichlet prior (concentration parameter α=1.0, which ensures the initial balance of the weights of each confounding factor).

[0087] S223. Based on the defined latent factor prior space, design a three-level network model including encoder, decoupler and decoder, and train the three-level network model using the separation representation algorithm.

[0088] It should be noted that the encoder adopts a 3-layer MLP structure (with 256, 128, and 64 hidden layer neurons respectively), which is responsible for mapping the joint input matrix into an initial high-dimensional feature vector. The decoupler embeds a mutual information constraint module, which separates the feature representations of the processing factor, outcome factor, and confounding factor through a separation representation algorithm (β-VAE combined with an adversarial decoupling strategy, with β set to 1.5) to avoid information redundancy between factors. The decoder adopts a 2-layer MLP (with 64 and 128 hidden layer neurons), which reconstructs the decoupled features into a vector with the same dimension as the original input. Then, the model is trained for 2000 rounds with the AdamW optimizer (learning rate 0.0005, weight decay 0.001), and validation is performed every 50 rounds to ensure model convergence (reconstruction error on the validation set ≤ 0.03).

[0089] S224. Based on the preset risk constraints, construct a multi-objective loss function and use the multi-objective loss function to optimize the trained three-level network model, and output the latent factor matrix.

[0090] S225. Calculate the correlation strength score of each latent factor in the latent factor matrix, and filter the calculation results. Based on the filtering results, use the correlation analysis algorithm to analyze the correlation strength between the drug and side effects, and output the confounding factor.

[0091] It should be further explained that, based on the preset risk constraints (absolute error of side effect risk prediction ≤ 0.05, factor decoupling mutual information value ≤ 0.1), a multi-objective loss function is constructed, consisting of reconstruction loss (MSE loss, weight 0.4), decoupling loss (mutual information loss, weight 0.3), and risk constraint loss (L1 loss, weight 0.3). The gradient descent method is used to iteratively optimize and train the model, improving the factor decoupling effect while satisfying the risk constraints. The output is a latent factor matrix with a dimension of (total number of patients × 60). Normalized mutual information (NMI) is used to calculate the correlation strength score between each latent factor. A screening threshold of 0.6 is set, and strongly correlated factor pairs with scores ≥ 0.6 are retained. Based on the screening results, a partial correlation analysis algorithm (controlling the significance level of α = 0.05) is used to analyze the correlation strength between the factor (drug selection) and the outcome factor (side effect risk). Non-interfering factors with an absolute value of partial correlation coefficient ≤ 0.2 are removed, and finally 36 core confounding factors are output.

[0092] S23. Based on the identified confounding factors, using the representational balance technique and the counterfactual reasoning framework, predict the side effects of each patient taking the medication and not taking the medication, and calculate the individual causal effect value for each patient based on the prediction results.

[0093] In this optional embodiment, based on the identified confounding factors, the side effects of each patient taking the medication and not taking the medication are predicted using representational equilibrium techniques and a counterfactual reasoning framework, respectively. The individual causal effect value for each patient is calculated based on the prediction results, including:

[0094] S231. Using the identified confounding factors as constraints, construct a balanced representation learning model, and optimize the balanced representation learning model using a two-branch adversarial training mechanism. Generate a balanced representation vector based on the optimization results.

[0095] In this optional embodiment, the identified confounding factors are used as constraints to construct a balanced representation learning model, and a two-branch adversarial training mechanism is used to optimize the balanced representation learning model. The balanced representation vector is generated based on the optimization results, including:

[0096] S2311. Using the identified confounding factors as constraints, design a neural network model containing an input layer, a hidden layer, and an output layer as a balanced representation learning model.

[0097] S2312. Construct a bipartite network architecture and iteratively optimize it using an adversarial training strategy until the bipartite network architecture converges to a preset threshold, thus completing the optimization of the balanced representation learning model.

[0098] In this optional embodiment, a bi-branch network architecture is constructed, and the bi-branch network architecture is iteratively optimized through an adversarial training strategy until the bi-branch network architecture converges to a preset threshold, thereby completing the optimization of the balanced representation learning model, including:

[0099] S23121. Based on the designed balanced representation learning model, construct a dual-branch network architecture that includes a representation optimization branch and a hybrid discrimination branch.

[0100] S23122. Using a course learning strategy, the dual-branch network architecture is divided into multiple stages, and adversarial training is performed in each stage.

[0101] S23123. Based on the adversarial training process, a federated alternating update protocol is designed to perform collaborative optimization in a multi-center environment. The convergence status of the representation optimization branch is monitored through Nash equilibrium. When the representation optimization branch converges to a preset threshold, the optimized representation optimization branch is output to complete the optimization of the balanced representation learning model.

[0102] S2313. Using the identified confounding factors as input and whether the patient takes medication as output, train a pre-built propensity score model, and use the trained propensity score model to calculate the propensity score for each patient, thereby obtaining the conditional probability of the patient receiving medication under the confounding factor conditions.

[0103] It should be further explained that an XGBoost-based propensity score model was constructed. This model consists of 10 decision trees with a depth of 5 and a maximum of 20 leaf nodes. The identified confounding factors (such as age group coding, one-hot vectors of underlying disease types, number of combined medications, etc.) are used as input features, and whether the patient takes the target drug (1 indicates taking, 0 indicates not taking) is used as the binary classification output label. The dataset is divided into a training set (for model fitting) and a test set (for performance validation) in a 7:3 ratio. The model is trained with a learning rate of 0.01 and 1000 iterations. The hyperparameters are optimized through five-fold cross-validation to ensure that the model's AUC value on the test set is not less than 0.85 to guarantee prediction reliability. After the model is trained, the confounding factor features of all patients are input into the model, and the conditional probability (propensity score, with a value range of [0,1]) of each patient receiving drug treatment under the current confounding factor conditions is output.

[0104] S2314. Convert the obtained conditional probabilities into corresponding weighting coefficients, and use the weighting coefficients to adjust the optimized balanced representation learning model. Output the balanced representation vector based on the weighted balanced representation learning model.

[0105] It should be further explained that the propensity score of each patient is converted into a weighted coefficient, where the weighted coefficient for patients taking medication is 1 / propensity score, and for patients not taking medication is 1 / (1-propensity score). The coefficient is truncated in the range of [0.01, 0.99] to avoid interference from extreme values. An optimized balanced representation learning model is constructed using an MLP structure with two hidden layers (128 and 64 neurons respectively) and an attention mechanism. The above weighted coefficients are incorporated into the model training process through a weighted loss function (such as weighted MSE loss) to dynamically adjust the model feature weight allocation, making it more focused on the feature learning of patients with high confounding influence. Based on the adjusted model, the processing factor, outcome factor, and confounding factor features in the joint input matrix are input, and the output is a 32-dimensional balanced representation vector that has canceled out the interference of confounding factors.

[0106] S232. Based on the generated balanced representation vector, construct an outcome regression model to simulate the side effects of taking the medication and not taking the medication, respectively.

[0107] S233. For each patient, the counterfactual reasoning framework is used to call the outcome regression model to predict the side effects of taking the medication and not taking the medication, and the difference between the side effects of taking the medication and not taking the medication is calculated according to different side effect types to obtain the individual causal effect value of the patient.

[0108] It should be further explained that the counterfactual reasoning framework adopts a latent outcome framework based on weighted balanced representations, and selects a ridge regression model with a regularization parameter λ of 0.001 as the outcome regression model. Using the balanced representation vector as input, prediction branches are constructed for two scenarios: medication and no medication. For each patient, the framework calls the model to first input the balanced representation vector to predict the side effect results when taking the medication (quantifying the severity of the side effect into a continuous value of 0-5, with 0 points for no occurrence), and then predicts the corresponding side effect results when not taking the medication. According to the preset 8 types of side effects (such as gastrointestinal reactions, skin allergies, abnormal liver and kidney function, etc.), the difference in side effect results for the same patient in the two scenarios is calculated (result in the medication group - result in the no medication group). A positive difference indicates that the medication increases the risk of the side effect, and a difference ≤ 0 indicates that there is no increase or decrease in risk, thus obtaining the individual causal effect value for each patient on various side effects (value range [-2,3]).

[0109] S24. Statistically aggregate the individual causal effect values ​​of all patients to calculate the average causal effect value of the drug and side effects.

[0110] It should be further explained that by statistically aggregating the individual causal effect values ​​of all patients on various side effects (e.g., taking the mean), the average causal effect value of the drug and the corresponding side effect pair is calculated. Then, a Bayesian network is constructed that includes covariates such as patient age, comorbidities, and concomitant medications. This average causal effect value is then optimized and calibrated. For example, after controlling for the hypertension covariate, the calibrated ACE of a certain antihypertensive drug causing dry cough is 1.8 (i.e., the drug increases the risk of dry cough by 80%). In the bias correction stage, a propensity score matching method is used to balance the baseline characteristics of the case group and the control group, effectively eliminating the interference of confounding factors such as the severity of the condition upon admission, ensuring that the final average causal effect value can truly reflect the causal relationship between the drug and the side effects. Through Bayesian network and propensity score matching correction, the false positive rate can be reduced by 40%. Comparative experiments show that in 100 sets of simulated data, the causal determination accuracy of this invention reaches 92%, while the traditional method is only 68%.

[0111] S3. Combining individual causal effect values, average causal effect values, and related covariates, draw a spatiotemporal risk heat map. Based on the risk distribution characteristics in the spatiotemporal risk heat map, classify the risks in the entire life cycle of the drug and trigger corresponding early warning strategies.

[0112] In this optional embodiment, a spatiotemporal risk heatmap is drawn by combining individual causal effect values, average causal effect values, and related covariates. Based on the risk distribution characteristics in the spatiotemporal risk heatmap, the risks throughout the entire life cycle of the drug are classified and corresponding early warning strategies are triggered, including:

[0113] S31. Collect relevant spatiotemporal data based on individual causal effect values, average causal effect values, and related covariates, and construct a spatiotemporal dataset.

[0114] S32. Based on the constructed spatiotemporal dataset, calculate the comprehensive risk value of the average causal effect value and the individual causal effect value, and use the preset spatiotemporal diffusion model to simulate the propagation and evolution of the comprehensive risk value in the time and space dimensions, and draw a spatiotemporal risk heat map based on the simulation results.

[0115] S33. Use a spatiotemporal convolutional network to extract risk distribution features from the drawn spatiotemporal risk heat map, and use an attention mechanism to weight the extracted risk distribution features in order to predict future risk change trends.

[0116] S34. Based on the predicted risk change trend, design a dynamic threshold optimization strategy based on reinforcement learning, and use the dynamic threshold optimization strategy to classify the risk level in the entire life cycle of the drug, and trigger the corresponding early warning strategy based on the classification result.

[0117] It should be further explained that the spatiotemporal data of individual causal effect values ​​for each patient, the mean causal effect value set at 0.8, and related covariates such as age group, geographic coding, and medication start time were collected. Time was divided into 52 nodes by week, and space was divided into 30 units by administrative region. These three types of core data were integrated with spatiotemporal labels to construct a spatiotemporal dataset with dimensions of 52×30×20. The comprehensive risk value (range [0,2]) was calculated using the weighted formula "Comprehensive Risk Value = Individual Causal Effect Value × 0.6 + Mean Causal Effect Value × 0.4". A preset spatiotemporal random walk model (diffusion coefficient 0.3) was then used to model the data. The transmission and diffusion patterns of adverse reactions were simulated in terms of time (e.g., weekly risk transmission efficiency of 0.7) and space (e.g., diffusion weight of adjacent regions of 0.5). Based on the simulation results, a spatiotemporal risk heatmap was generated using GeoPandas. Darker colors indicate a higher incidence of adverse reactions. Drill-down was supported by population dimensions (age <18 years / 18-65 years / 65 years, gender, allergy history) and medication dimensions (high / medium / low dosage, route of administration: oral / injection, number of combined medications: 0 / 1-2 / ≥3). A spatiotemporal convolutional network with 3 convolutional layers (3×3 kernel, stride 1) and 2 pooling layers was used to extract spatiotemporal correlation features from the heatmap. A four-head attention mechanism is used to highlight the feature weights of high-risk spatiotemporal nodes. These weighted features are then input into a pre-defined Prophet model to predict risk trends over the next three months. A reinforcement learning dynamic threshold optimization strategy based on DQN is designed, using the predicted trend as the state input, the threshold adjustment range as the action space, and the risk warning accuracy (target ≥ 0.9) as the reward function. This optimization yields dynamic thresholds for different risk levels, which are then used to classify the risk levels throughout the drug's lifecycle (orange warning: ≥ 2 independent data source confirmation signals, such as hospital EMR and medical insurance data both showing abnormal rash incidence, or causal inference p < 0.05; red warning: ...). A positive or likely signal that meets WHO-UMC standards (such as a reasonable time relationship between medication and response, improvement after discontinuation of medication and no explanation for comorbidities) is triggered when the predicted value exceeds the 95th percentile of historical data (for example, an anomaly in a single data source, such as a 200% week-on-week surge in mentions of a certain drug on social media about palpitations, but not yet meeting the MedDRA coding standard); the spatiotemporal risk heatmap has dynamic drill-down analysis capabilities, which can improve decision-making efficiency by 50% compared to traditional tabular reports. After a provincial drug regulatory authority applied it, the response time for risk events was significantly reduced from 72 hours to 24 hours.

[0118] S4. Use the preset report template to convert the early warning strategy into a standard format risk assessment report, adjust the format of the risk assessment report according to the drug regulatory requirements, and output the adjusted risk assessment report according to the preset format.

[0119] It should be further explained that the standard risk assessment report template, which includes 10 fixed modules such as basic drug information and individual / average causal effect values, is used. The obtained risk level (low / medium / high / extremely high), corresponding early warning strategy (routine monitoring / enhanced monitoring, etc.), and relevant core data (such as a comprehensive risk value of 0.8, a risk growth rate of 30% in the next 3 months, and 5 high-risk regions) are accurately filled into the corresponding fields of the template (data filling accuracy ≥99%). The report format is adjusted according to drug regulatory requirements (such as compliance with the "Regulations on the Reporting and Monitoring of Adverse Drug Reactions"), and new sections on adverse reaction incidence statistics (retaining 3 decimal places), regional risk proportion analysis presented in pie charts, and regulatory compliance explanations are added, transforming the original tabular spatiotemporal format. The data is transformed into line charts to display time trends and heatmaps to embed spatial distribution visualizations. The main text is adjusted to SimSun font size 12, and the titles to Heiti font size 12, with key conclusions (such as extremely high risk warnings and recommendations to suspend use) highlighted in bold. The adjusted risk assessment report is output in both preset PDF (A4 paper, encryption enabled) and editable Word formats, along with appendices explaining data sources and calculation methods. Furthermore, the warning system incorporates built-in XML templates for ICSR and PSUR reports, mapping analysis results to standard fields via XSLT. For example, the "Global Cumulative Exposure" field in PSUR is automatically calculated and populated from sales data, and the format is automatically adjusted according to the regulatory requirements of the target region; for instance, the FDA requires the inclusion of MedWatch data. The 3500A form and narratives section require the use of eCTD format and the inclusion of a Risk Management Plan (RMP). One-click export supports PDF / XML / Excel formats and automatically pushes data to the head of the drug safety department via the company's email system. Through template and regulatory rule engines, reports are automatically generated, improving efficiency by 90%. For example, after a pharmaceutical company tried it, the PSUR report generation time was reduced from 15 days to 1.5 days, and compliance costs were reduced by 60%.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring and early warning of drug risks throughout their entire lifecycle based on multi-source data fusion, characterized in that, The method includes: S1. Preprocess the collected multi-source data, construct a time-series heterogeneous graph based on the preprocessing results, and use a contrastive learning algorithm to calculate the similarity between drug and side effect pairs in the time-series heterogeneous graph to generate a candidate signal list. S2. Based on the candidate signal list and related covariates, construct a dataset and use the separation representation learning algorithm to identify confounding factors in the dataset that are associated with drug selection and side effect risk. Calculate the individual causal effect value for each patient based on the confounding factors to obtain the average causal effect value. S2 includes: S21. Based on the generated candidate signal list, divide the patients into a medication group and a non-medication group, and construct a dataset by combining covariates related to medication use and side effects. S22. Decouple the constructed dataset to the prior space of potential factors, and combine it with the preset risk constraints to identify confounding factors that are associated with drug selection and side effect risk. S22 includes: S221. The dataset is processed using a generative adversarial interpolation algorithm, and a joint input matrix containing drug selection, side effect risk, and covariates is constructed based on the processing results. S222. Based on the pre-acquired medical knowledge, the latent factors in the joint input matrix are divided into dimensions. Based on the division results, the total dimension of the latent factors and the proportion of each type of latent factor in the total dimension are allocated, and a corresponding prior space is defined for each latent factor. S223. Based on the defined latent factor prior space, design a three-level network model including encoder, decoupler and decoder, and train the three-level network model using the separation representation algorithm. S224. Based on the preset risk constraints, construct a multi-objective loss function and use the multi-objective loss function to optimize the trained three-level network model, and output the latent factor matrix. S225. Calculate the correlation strength score of each latent factor in the latent factor matrix, and filter the calculation results. Based on the filtering results, use the correlation analysis algorithm to analyze the correlation strength between the drug and side effects, and output the confounding factor. S23. Based on the identified confounding factors, using the representational balance technique and the counterfactual reasoning framework, predict the side effects of each patient taking the medication and not taking the medication, and calculate the individual causal effect value for each patient based on the prediction results. S24. Statistically aggregate the individual causal effect values ​​of all patients to calculate the average causal effect value of the drug-side effect pair; S3. Combining individual causal effect values, average causal effect values, and related covariates, draw a spatiotemporal risk heat map. Based on the risk distribution characteristics in the spatiotemporal risk heat map, classify the risks in the entire life cycle of the drug and trigger corresponding early warning strategies. S4. Use the preset report template to convert the early warning strategy into a standard format risk assessment report, adjust the format of the risk assessment report according to the drug regulatory requirements, and output the adjusted risk assessment report according to the preset format.

2. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 1, characterized in that, The process involves preprocessing the collected multi-source data, constructing a temporal heterogeneity graph based on the preprocessing results, and using a contrastive learning algorithm to calculate the similarity between drug and side effect pairs in the temporal heterogeneity graph, generating a candidate signal list including: S11. Collect multi-source data of patients from multi-source heterogeneous platforms and map the multi-source data into a unified standardized format; S12. Based on the standardized multi-source data, the distance algorithm is used to remove duplicate data and clean outliers. Differential privacy technology is used to perform privacy protection processing on the cleaned multi-source data. Based on the privacy protection processing results, a time-series heterogeneous graph containing drugs, side effects and time-series relationships is constructed. S13. Use a contrastive learning algorithm to learn the temporal heterogeneous graph, calculate the similarity between drug and side effect pairs in the temporal heterogeneous graph, and filter the drug and side effect pairs based on the calculation results to generate a candidate signal list.

3. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 2, characterized in that, The process involves using a contrastive learning algorithm to learn the temporal heterogeneity graph, calculating the similarity between drug and side effect pairs in the graph, and then filtering the drug and side effect pairs based on the calculation results to generate a candidate signal list, including: S131. Hierarchical attention mechanism is used to capture the multidimensional dynamic correlation features between drugs, side effects and covariates over time, and contrastive learning algorithm is used to optimize the discriminativeness of graph structure embedding vectors. S132. Using a cross-modal semantic fusion algorithm, semantically align the optimized graph structure embedding vector with the corresponding text embedding vector to generate a cross-modal embedding vector. S133. Based on the generated cross-modal embedding vectors, calculate the Gaussian kernel embedding similarity of drug and side effect pairs in the graph structure space and the semantic similarity in the text semantic space. Based on the calculation results, use an adaptive threshold learning strategy to dynamically screen drug and side effect pairs and construct a candidate set. S134. Analyze the candidate set using an association rule mining algorithm, calculate the support and confidence of each drug-side effect pair, and combine the preset support and confidence thresholds to select drug-side effect pairs that simultaneously meet the support and confidence thresholds, generating a candidate signal list.

4. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 1, characterized in that, The dimensional segmentation of the latent factors in the joint input matrix includes: treatment factors related to drug selection, outcome factors related to side effect risk, and confounding factors that link drug selection and side effect risk.

5. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 4, characterized in that, Based on the identified confounding factors, the representativeness balance technique and counterfactual reasoning framework are used to predict the side effects of each patient taking the medication and not taking the medication, respectively. Based on the prediction results, the individual causal effect value for each patient is calculated, including: S231. Using the identified confounding factors as constraints, construct a balanced representation learning model, and optimize the balanced representation learning model using a two-branch adversarial training mechanism. Generate a balanced representation vector based on the optimization results. S232. Based on the generated balanced representation vector, construct an outcome regression model to simulate the side effects of taking the medication and not taking the medication, respectively. S233. For each patient, the counterfactual reasoning framework is used to call the outcome regression model to predict the side effects of taking the medication and not taking the medication, and the difference between the side effects of taking the medication and not taking the medication is calculated according to different side effect types to obtain the individual causal effect value of the patient.

6. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 5, characterized in that, The process of constructing a balanced representation learning model using the identified confounding factors as constraints, optimizing the balanced representation learning model using a two-branch adversarial training mechanism, and generating a balanced representation vector based on the optimization results includes: S2311. Using the identified confounding factors as constraints, design a neural network model containing an input layer, a hidden layer, and an output layer as a balanced representation learning model. S2312. Construct a bipartite network architecture and iteratively optimize the bipartite network architecture through an adversarial training strategy until the bipartite network architecture converges to a preset threshold, thus completing the optimization of the balanced representation learning model. S2313. Using the identified confounding factors as input and whether the patient takes medication as output, train a pre-built propensity score model, and use the trained propensity score model to calculate the propensity score for each patient, thereby obtaining the conditional probability of the patient receiving medication under the confounding factor conditions. S2314. Convert the obtained conditional probabilities into corresponding weighting coefficients, and use the weighting coefficients to adjust the optimized balanced representation learning model. Output the balanced representation vector based on the weighted balanced representation learning model.

7. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 6, characterized in that, The process of constructing a dual-branch network architecture and iteratively optimizing it using an adversarial training strategy until the dual-branch network architecture converges to a preset threshold, thereby completing the optimization of the balanced representation learning model, includes: S23121. Based on the designed balanced representation learning model, construct a dual-branch network architecture that includes a representation optimization branch and a hybrid discrimination branch. S23122. Using a course learning strategy, the dual-branch network architecture is divided into multiple stages, and adversarial training is performed in each stage. S23123. Based on the adversarial training process, a federated alternating update protocol is designed to perform collaborative optimization in a multi-center environment. The convergence status of the representation optimization branch is monitored through Nash equilibrium. When the representation optimization branch converges to a preset threshold, the optimized representation optimization branch is output to complete the optimization of the balanced representation learning model.

8. The method for monitoring and early warning of drug life-cycle risks based on multi-source data fusion according to claim 1, characterized in that, The process of combining individual causal effect values, average causal effect values, and related covariates to create a spatiotemporal risk heat map, and classifying the risks throughout the entire life cycle of a drug based on the risk distribution characteristics in the spatiotemporal risk heat map and triggering corresponding early warning strategies, includes: S31. Based on individual causal effect values, average causal effect values, and related covariates, collect relevant spatiotemporal data and construct a spatiotemporal dataset. S32. Based on the constructed spatiotemporal dataset, calculate the comprehensive risk value of the average causal effect value and the individual causal effect value, and use the preset spatiotemporal diffusion model to simulate the propagation and evolution of the comprehensive risk value in the time and space dimensions, and draw a spatiotemporal risk heat map based on the simulation results. S33. Use a spatiotemporal convolutional network to extract risk distribution features from the drawn spatiotemporal risk heat map, and use an attention mechanism to weight the extracted risk distribution features in order to predict future risk change trends. S34. Based on the predicted risk change trend, design a dynamic threshold optimization strategy based on reinforcement learning, and use the dynamic threshold optimization strategy to classify the risk level in the entire life cycle of the drug, and trigger the corresponding early warning strategy based on the classification result.

Images