A method and system for drug safety monitoring of a combination drug

Abstract

A pharmacovigilance method and system for combined drug use acquires multi-source information from electronic health records; estimates clinical symptom medication parameters based on a Bayesian engine; constructs a symptom network, calculates the betweenness centrality of each node, and identifies pivotal symptoms for adverse reaction propagation; calculates Bayesian factors for clinical symptom medication parameters, assesses the probability of adverse reaction occurrence and symptom severity, and generates a risk report. This invention captures the impact of drug interactions through a Bayesian hierarchical model, and utilizes the Horseshoe prior to effectively handle sparse data, reducing the false negative rate and improving signal detection sensitivity; secondly, it incorporates patient covariates and symptom network analysis to provide personalized risk assessment and key symptom alerts, improving the interpretability of conclusions and enhancing clinical relevance. Finally, it outputs an intuitive risk report to assist clinical decision-making, demonstrating strong practicality.

Description

Technical Field

[0001] This invention relates to the field of pharmacovigilance technology, and more specifically, to a pharmacovigilance method and system for combined drug use. Background Technology

[0002] With the increasing demand for treatment of complex diseases, combination therapy has become a common clinical approach, especially in the field of psychiatry, where the simultaneous use of antipsychotics and mood stabilizers is even more prevalent. Studies have indicated that while combination therapy improves efficacy, it also significantly increases the risk of adverse reactions.

[0003] However, traditional pharmacovigilance systems are typically based on single-drug-adverse-event association analysis methods (such as proportional-to-report ratio, reporting ratio, etc.), which do not adequately consider the synergistic effects between multiple drugs. These methods often ignore patient heterogeneity (such as covariates like age, gender, and comorbidities) and clinical scale information, easily leading to high missed detection rates and low clinical relevance of combined medication safety signals. In existing technologies, while Bayesian hierarchical models have been used for drug safety signal detection, they are limited to single-drug or simple hierarchical analysis and have not been extended to multi-drug interaction scenarios; graph theory methods have been applied in symptom network research, but have not yet been integrated with pharmacovigilance systems.

[0004] Therefore, there is an urgent need for a pharmacovigilance approach that can integrate multidrug interactions, patient heterogeneity, and clinical symptom associations. Summary of the Invention

[0005] (I) Purpose of the invention: In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a pharmacovigilance method and system for combined drug use, and to solve the problems of missed detection of combined drug use signals, false detection caused by patient heterogeneity, and lack of personalized risk assessment indicators.

[0006] (II) Technical Solution: To solve the above-mentioned technical problems, this technical solution provides a pharmacovigilance method for combined drug use, specifically including:

[0007] Step 1: Obtain multi-source information from electronic health records;

[0008] Step 2: Based on multi-source information, estimate clinical symptom medication parameters using a Bayesian engine;

[0009] Step 3: Construct a symptom network based on clinical symptom medication parameters, calculate the betweenness centrality of each node, and identify pivotal symptoms for the spread of adverse reactions;

[0010] Step 4: Calculate Bayesian factors for medication parameters for clinical symptoms, assess the probability of adverse reactions and the severity of symptoms, and generate a risk report.

[0011] Preferably, the data acquisition module is responsible for obtaining the patient's combined medication data, patient covariate data, and clinical symptom data from the electronic health record.

[0012] Preferably, step 2 includes:

[0013] Step 201: Construct a Bayesian hierarchical model;

[0014] Step 202: Perform Bayesian priors on the parameters of the Bayesian hierarchical model;

[0015] Step 203: Estimate the parameters of the Bayesian hierarchical model using the MCMC method;

[0016] Step 204: Verify the Bayesian hierarchical model.

[0017] Preferably, the Bayesian hierarchical model decomposes the overall risk into a single-drug baseline risk layer, a drug interaction layer, and a covariate regulation layer:

[0018] ,

[0019] in, The intercept term represents the baseline logarithmic probability of the occurrence of adverse reactions when all single drug indicator variables, drug interaction terms, and patient covariates are 0, reflecting the basic risk level without drug exposure and the influence of covariates. This represents the baseline risk parameter for the i-th drug alone; This represents the drug indicator variable for the i-th drug. =1 indicates that the patient is taking the i-th type of medication. =0 indicates no medication was taken; This represents the drug interaction parameter between drug i and drug j; Let represent the drug interaction term when the i-th drug is used in combination with the j-th drug; m represents the types of drugs; p represents the number of covariates; and k represents the order of the covariates. Indicates the parameters of the covariate effect; This represents the patient covariate.

[0020] Preferably, baseline risk parameters for single-drug therapy Covariate effect parameters We take the conventional normal prior, but for key drug interaction parameters... The Horseshoe prior is used for sparsity contraction.

[0021] Preferably, step 3 includes:

[0022] Step 301: Construct a symptom network based on multi-source information;

[0023] Step 302: Calculate the betweenness centrality of each node to determine the pivotal symptoms.

[0024] Preferably, symptoms are used as nodes in the symptom network, and the propagation or association relationships between symptoms are used as edges. The edge weights are determined by calculating the co-occurrence probability of symptoms, the Pearson correlation coefficient, or the distance metric.

[0025] Preferably, graph theory analysis is used to quantify the betweenness centrality of each node, calculate the shortest path between all pairs of nodes in the symptom network, and count the frequency of the node as a mediator of the shortest path as the betweenness centrality of that node.

[0026] The more shortest paths a node has, the higher its betweenness centrality. The symptoms represented by that node are often located on the shortest paths between other symptoms, making it a hub for the spread of adverse drug reactions.

[0027] Preferably, step 4 includes:

[0028] Step 401: Calculate Bayesian factors to determine effective drug interaction parameters. To determine the probability of adverse reactions when taking multiple medications;

[0029] Step 402: Use a standardized rating scale to quantify the severity of symptoms and obtain the symptom severity score;

[0030] Step 403: Weight the probability of adverse reactions when taking multiple medications with the obtained symptom severity scores to generate a risk report.

[0031] Preferably, the calculation includes the drug interaction parameters. The model and models that do not include the drug interaction parameters The marginal likelihood ratio of the model:

[0032] ,

[0033] in, Indicating drug interaction parameters Bayesian factor; The inclusion factor represents the probability of the current observation data occurring when drug i and drug j interact. This represents the probability of the current observed data occurring when drug i and drug j do not interact.

[0034] when At that time, the drug interaction was determined to be significant.

[0035] Preferably, for each adverse reaction, the severity of the symptoms is assessed by converting it into a quantitative score using a standardized rating scale, resulting in a symptom severity score. The higher the score, the greater the correlation between the symptoms and the medication and the greater the severity.

[0036] The total severity score is obtained by multiplying the quantitative score of each individual symptom by its corresponding weighting coefficient and summing the results.

[0037] ,

[0038] Where 's' represents the symptom type; A represents the weighting coefficient of the symptom; A represents the quantitative score of the symptom.

[0039] A pharmacovigilance system for combined drug use, applicable to a pharmacovigilance method for combined drug use, includes a data acquisition module, a Bayesian inference module, a network analysis module, and a risk assessment module; the data acquisition module obtains multi-source information from electronic health records;

[0040] The Bayesian inference module estimates clinical symptom medication parameters based on multi-source information and a Bayesian engine.

[0041] The network analysis module constructs a symptom network based on clinical symptom medication parameters, calculates the betweenness centrality of each node, and identifies pivotal symptoms for the spread of adverse reactions.

[0042] The risk assessment module calculates Bayesian factors for medication parameters for clinical symptoms, assesses the probability of adverse reactions and the severity of symptoms, and generates a risk report.

[0043] (III) Beneficial Effects: This invention provides a pharmacovigilance method and system for combined drug use. First, it captures the influence of drug interactions using a Bayesian hierarchical model. The Horseshoe prior effectively handles sparse data, reducing the false negative rate and improving signal detection sensitivity. Second, it incorporates patient covariates and symptom network analysis to provide personalized risk assessment and key symptom alerts, improving the interpretability of conclusions and enhancing clinical relevance. Third, it uses a Bayesian method to validate signals, avoiding the limitations of traditional P-values, and utilizes multidimensional data fusion to provide reliable risk warnings, ensuring the mathematical rigor of statistical inference. Finally, it outputs an intuitive risk report to assist clinical decision-making, demonstrating strong practicality. Attached Figure Description

[0044] Figure 1 This is a flowchart of the steps of a pharmacovigilance method for combined drug use according to the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of a pharmacovigilance system for combined drug use according to the present invention;

[0046] Figure 3 This is a flowchart of step 2 of the present invention;

[0047] Figure 4 This is a flowchart of step 3 of the present invention;

[0048] Figure 5 This is a flowchart of step 4 of the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to preferred embodiments. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0050] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that these drawings are for illustrative purposes only and are not drawn to scale, and should not be construed as limiting the actual scope of protection of the present invention.

[0051] A pharmacovigilance method for combined drug use, such as Figure 1 As shown, the specific steps include:

[0052] Step 1: Obtain multi-source information from electronic health records.

[0053] Step 2: Based on multi-source information, estimate clinical symptom medication parameters using a Bayesian engine.

[0054] Step 3: Construct a symptom network based on clinical symptom medication parameters, calculate the betweenness centrality of each node, and identify pivotal symptoms for the spread of adverse reactions.

[0055] Step 4: Calculate Bayesian factors for medication parameters for clinical symptoms, assess the probability of adverse reactions and the severity of symptoms, and generate a risk report.

[0056] A pharmacovigilance system for combined drug administration, applicable to the aforementioned pharmacovigilance method for combined drug administration, such as... Figure 2 As shown, the system includes a data acquisition module, a Bayesian inference module, a network analysis module, and a risk assessment module. The data acquisition module obtains multi-source information from electronic health records. The Bayesian inference module estimates clinical symptom medication parameters based on the multi-source information using a Bayesian engine. The network analysis module constructs a symptom network based on the clinical symptom medication parameters, calculates the betweenness centrality of each node, and identifies pivotal symptoms in the spread of adverse reactions. The risk assessment module calculates Bayesian factors for the clinical symptom medication parameters, assesses the probability of adverse reactions and symptom severity, and generates a risk report.

[0057] Specifically, the data acquisition module in step 1 includes:

[0058] The data acquisition module is responsible for extracting multi-source information from electronic health records, such as patients' drug prescriptions, patient characteristics, and patients' adverse event reports, thereby obtaining patients' combined medication data, patients' covariate data, and clinical symptom data.

[0059] The combined medication data includes the use of at least two drugs. The patient covariate data includes the patient's age, gender, and disease type. The clinical symptom data includes the time of symptom onset and severity.

[0060] More specifically, the data acquisition module first obtains multi-source patient information by reading various adverse reaction databases and HIS (Hospital Information System) data through an API (Application Programming Interface). Then, this multi-source information is cleaned by uniform encoding, removing duplicates and irrelevant information, and handling missing values. Finally, the data is summarized according to a unified standard format. This standard format includes patient ID, basic information, symptoms, drug name, dosage, manufacturer, reporting country, adverse reaction symptoms, and outcome.

[0061] Specifically, such as Figure 3 As shown, the specific operations of the Bayesian inference module in step 2 include:

[0062] Step 201: Construct a Bayesian hierarchical model.

[0063] A three-layered Bayesian hierarchical model was constructed, which decomposes the overall risk into a single-drug baseline risk layer, a drug interaction layer, and a covariate regulation layer. The single-drug baseline risk layer includes the main effects of each drug; the drug interaction layer includes the synergistic effects of pairwise or multi-drug combinations; and the covariate regulation layer includes the moderating effect of individual patient characteristics on the risk of adverse reactions.

[0064] Assuming a single drug indicator variable is taken by a patient and patient covariates As shown in formula (1), this represents the probability of adverse reactions occurring, i.e., the Bayesian hierarchical model of overall risk:

[0065] (1),

[0066] in, The intercept term represents the baseline logarithmic probability of the occurrence of adverse reactions when all single drug indicator variables, drug interaction terms, and patient covariates are 0, reflecting the basic risk level without drug exposure and the influence of covariates. This represents the baseline risk parameter for the i-th drug alone; This represents the drug indicator variable for the i-th drug. =1 indicates that the patient is taking the i-th type of medication. =0 indicates no medication was taken; This represents the drug interaction parameter between drug i and drug j; Let represent the drug interaction term when the i-th drug is used in combination with the j-th drug; m represents the types of drugs; p represents the number of covariates; and k represents the order of the covariates. This represents the covariate effect parameter, i.e., the moderating effect of the k-th covariate on the baseline risk of a single drug; It represents patient covariates, including individual characteristics such as age, gender, disease type, and clinical scale data, to moderate for individual differences in adverse reaction risk.

[0067] Step 202: Perform Bayesian priors on the parameters of the Bayesian hierarchical model.

[0068] Baseline risk parameters for single drugs Covariate effect parameters A conventional normal prior can be used, while key drug interaction parameters... The Horseshoe prior is used for sparsity contraction.

[0069] The shrinkage intensity is controlled by the hierarchical Cauchy distribution scale parameter λ~Cauchy(0,1), which enables adaptive shrinkage of high-dimensional interaction signals without the need for manual adjustment of the penalty parameter. The Bayesian hierarchical model can automatically learn important interaction signals based on the data.

[0070] The Horseshoe prior can adaptively compress most meaningless parameters to near zero when processing sparse high-dimensional data, thereby reducing overfitting and noise. At the same time, the Horseshoe can automatically filter variables, retaining only key interaction signals with statistical significance, and providing significant support for truly important interaction signals, thereby solving the problems of data sparsity and multicollinearity in combination drug scenarios.

[0071] In this step, the scale-mixed normal distribution is used. The core idea is to use the heavy-tailed property to assign non-zero weights to important model parameters and to strongly shrink noise parameters, thus balancing the sparsity and estimation accuracy of the Bayesian hierarchical model.

[0072] Step 203: Estimate the parameters of the Bayesian hierarchical model using the MCMC method.

[0073] The posterior distribution of the Bayesian hierarchical model parameters is estimated using the MCMC (Markov Chain Monte Carlo) method. This is used to compress meaningless drug interaction coefficients to near zero, retain significant interaction signals, and output the posterior means and confidence intervals of the single-drug baseline risk layer, drug interaction layer, and covariate regulation layer.

[0074] The MCMC method includes, but is not limited to, Gibbs sampling, Metropolis-Hastings method, etc.

[0075] Step 204: Verify the Bayesian hierarchical model.

[0076] The Bayesian hierarchical model was validated using existing datasets for known drug interaction parameters. The ability to identify drug interaction parameters ensures that... The estimated value is consistent with the actual value.

[0077] The Bayesian hierarchical model can simultaneously estimate single-drug effects and multi-drug interaction effects with limited sample size and numerous variables, and incorporates patient covariates into the model to account for individual heterogeneity.

[0078] Specifically, such as Figure 4 As shown, the specific operation steps of the network analysis unit in step 3 are as follows:

[0079] Step 301: Construct a symptom network based on multi-source information.

[0080] Based on the aforementioned multi-source information such as clinical time-series data or medical records, a symptom network is constructed, with symptoms as nodes and the propagation or correlation relationships between symptoms as edges. The edge weights are determined by calculating the co-occurrence probability of symptoms, Pearson correlation coefficient, or distance metric.

[0081] For example, in a set of data on psychiatric patients, it was found that "insomnia" and "anxiety" were often reported together. In this case, an edge from "insomnia" to "anxiety" would be formed in the symptom network diagram. The edge weight reflects the correlation between "insomnia" and "anxiety". The edge can be Pearson correlation or various distance definitions.

[0082] Step 302: Calculate the betweenness centrality of each node to determine the pivotal symptoms.

[0083] Graph theory analysis is used to quantify the betweenness centrality of each node. The shortest paths between all pairs of nodes in the symptom network are calculated, and the frequency with which each node acts as a "mediator" on these shortest paths is counted; this frequency is the betweenness centrality. The more shortest paths a node has, the higher its betweenness centrality; conversely, the fewer shortest paths a node has, the lower its betweenness centrality.

[0084] Betweenness centrality is defined as the frequency with which a node acts as a mediator for the shortest path between other nodes in the symptom network. It is an indicator that characterizes the importance of a node by the number of shortest paths passing through it, and can measure the criticality of the symptom in various propagation paths.

[0085] The Louvain method was used to perform community clustering on the symptom network to identify symptom groups with similar transmission patterns, improving the clinical relevance of hub nodes. By comparing the betweenness centrality of each node and combining it with the node density within the community, nodes with betweenness centrality greater than or equal to a first threshold and ranking in the top 10% of connectivity within the community were marked as risk transmission hubs. The node with the highest betweenness centrality was identified as a hub symptom for adverse reaction transmission, reflecting its key mediating role in multi-drug adverse reaction pathways. For example, if a node has the highest betweenness centrality, it indicates that the symptom represented by that node is frequently located on the shortest path between other symptoms, and may be a hub for the transmission of multi-drug adverse reactions.

[0086] The first threshold value ranges from 0.6 to 0.8, and can be dynamically adjusted according to the clinical scenario. In this invention, the first threshold value is 0.7.

[0087] A virtual damping coefficient λ is applied to the risk propagation hub to reduce the basic propagation weight. The initial value of the virtual damping coefficient is 0.3. When the week-on-week growth rate of the adverse reaction incidence rate associated with the node corresponding to the risk propagation hub is greater than or equal to 20%, the virtual damping coefficient λ is automatically increased to 0.6 to reduce the propagation weight of the node of the risk propagation hub in the symptom network and inhibit the excessive spread of risk signals. When the adverse reaction incidence rate associated with the node corresponding to the risk propagation hub decreases for two consecutive weeks, the virtual damping coefficient λ is reduced back to 0.4.

[0088] By dynamically attenuating the propagation weight of highly volatile hub nodes, the interference of short-term data fluctuations on risk assessment is reduced, thereby improving the stability of signal detection.

[0089] The network analysis module quantifies the "bridging" role of symptoms in the transmission path, which can locate the core early warning symptoms of adverse reactions to combined medications. It can help identify high-risk transmission paths and key target symptoms, providing effective clues for signal understanding and intervention, and providing targeted targets for clinical intervention.

[0090] Specifically, such as Figure 5 As shown, the specific operation of the risk assessment module in step 4 is as follows:

[0091] Step 401: Calculate Bayesian factors to determine effective drug interaction parameters. This determines the probability of adverse reactions when taking multiple medications.

[0092] Quantifying drug interaction parameters using Bayes factor (BF) Statistical reliability, assessing drug interaction parameters To determine the clinical significance of drug interactions and thus identify effective drug interaction parameters. More specifically, as shown in formula (2), for each drug interaction parameter Calculate parameters including the drug interaction parameters The model and models that do not include the drug interaction parameters The marginal likelihood ratio of the model:

[0093] (2),

[0094] in, Indicating drug interaction parameters Bayesian factor; This indicates parameters representing drug interactions. The value is not zero, which represents the probability of the current observed data occurring under the assumption that drug i and drug j interact. Indicating drug interaction parameters The value equals 0, which represents the probability of the current observation data occurring under the assumption that there is no interaction between drug i and drug j.

[0095] According to Jeffreys's criteria, when When the signal is supported by strong evidence, it can significantly reduce the risk of false positives. The drug interaction is significant, ensuring that only truly clinically significant drug interactions are included in subsequent risk assessments, providing a rigorous statistical basis for the generation of personalized risk indices.

[0096] Based on effective drug interaction parameters The probability of adverse reactions for each patient when taking multiple drugs at the same time is calculated according to formula (1), and an individualized risk index is constructed by combining patient covariate information.

[0097] Step 402: Use a standardized rating scale to quantify the severity of symptoms and obtain the symptom severity score.

[0098] For each adverse reaction, the severity of symptoms is assessed by converting it into a quantitative score using a standardized rating scale. A higher score indicates a greater correlation between the symptoms and the medication, and a higher severity; conversely, a lower score indicates a lower correlation between the symptoms and the medication, and a lower severity.

[0099] The severity of symptoms includes the severity of the symptoms themselves, the duration of the symptoms, the reversibility of the symptoms, and the impact on quality of life. The standardized rating scale considers not only the severity of the symptoms themselves but also their impact on the patient's quality of life.

[0100] The standardized scoring scales include, but are not limited to, the Naranjo method or the Horn's Drug Interaction Probability Scale. The Naranjo method is a standardized tool for assessing the causal relationship between adverse drug reactions and drugs. Through a structured scoring system, it helps clinicians, pharmacists, and other professionals objectively determine whether adverse reactions are caused by a specific drug, rather than other factors. The Horn's Drug Interaction Probability Scale is a standardized tool for assessing the likelihood of drug-drug interactions. Through a structured scoring system, it helps clinicians objectively determine whether a patient's clinical event is caused by the interaction between two or more drugs, rather than other factors.

[0101] Step 403: Weight the probability of adverse reactions when taking multiple medications with the obtained symptom severity scores to generate a risk report.

[0102] The weighting coefficients are preset based on the probability and severity of adverse reactions to symptoms in clinical practice; the weighting coefficients are dynamically adjusted in combination with individual patient characteristics. As shown in formula (3), the quantitative score of a single symptom is multiplied by the corresponding weighting coefficient, and the sum is obtained to obtain the total severity score:

[0103] (3),

[0104] Where 's' represents the symptom type; A represents the weighting coefficient of the symptom; A represents the quantitative score of the symptom.

[0105] A three-dimensional risk envelope was constructed with medication duration as the X-axis, patient age as the Y-axis, and number of comorbidities as the Z-axis. The range of each axis was determined by historical data quantiles. For example, the X-axis ranged from 0 to 180 days, the Y-axis from 18 to 90 years old, and the Z-axis from 0 to 8 comorbidities.

[0106] Nonparametric kernel density estimation was used to calculate the 95% confidence intervals for the incidence of adverse reactions under each dimension combination. The boundaries of the intervals were connected to form the surface of the risk envelope, and the internal region of the risk envelope was defined as the risk exposure space. As shown in formula (4), sampling points were allocated according to risk density within the risk envelope.

[0107] Risk density = base density × (1 + regional risk value / global risk mean) (4).

[0108] When the risk density is greater than or equal to the first risk threshold, the area is identified as a high-risk region, and the sampling density of the high-risk region is automatically increased by 2 times. When the risk density is less than the first risk threshold, the area is identified as a low-risk region, and the sampling density of the low-risk region is reduced by 50%. With the same sample size, the parameter estimation accuracy of the high-risk region is greatly improved, reducing the result bias caused by data sparsity.

[0109] The first risk threshold ranges from 1.2 to 2.0, and in this invention, the first risk threshold is set to 1.5. For example, in the scenario of combined antibiotic use, when the risk envelope is a region with a duration of medication > 30 days, an age > 65 years, and ≥ 3 comorbidities, and the regional risk density is 2.3 times the global average, the risk density = baseline density × (1 + 2.3) = 3.3 × baseline density, which far exceeds the first risk threshold of 1.5. Therefore, the sampling density of this region is increased by 2 times.

[0110] By combining the overall severity score with the probability of adverse reactions, a final personalized risk report is generated, achieving a two-dimensional risk quantification of "probability of occurrence - severity of consequences".

[0111] The risk report includes basic patient information, a detailed list of major drug interaction risks detected, symptom nodes with high betweenness centrality, personalized medication adjustment recommendations, and specific intervention guidelines for different risk levels.

[0112] The proposed medication adjustment recommendations include dosing time, discontinuation recommendations, a list of alternative medications, and monitoring indicators. High-risk drug combinations that need to be discontinued and their discontinuation priorities should be clearly identified; for example, immediately discontinuing medications A and B and switching to medication C as monotherapy. Three to five alternative medications with similar efficacy and cross-risks significantly below the first risk threshold should be provided, along with the level of evidence-based medicine.

[0113] The following is a description with reference to specific embodiments:

[0114] A dataset simulating combined medication use in a psychiatric setting was constructed: assuming 1000 patients, 30% taking drug A and 20% taking drug B, with some patients also receiving combination therapy of A and B. The model parameters were set as intercept -3.5, and the effect of single-drug A was considered. Single drug B effect Interaction effect between the two drugs In addition to the age covariate effect This generates information on the occurrence of adverse reactions for each patient.

[0115] After fitting the data using the Bayesian hierarchical model, the estimated drug interaction parameters are approximately... The 95% confidence interval is approximately [0.85, 3.04], which is significantly greater than 0, corresponding to a Bayesian factor well exceeding 100, indicating that the drug interaction parameters are highly significant. The single-drug main effect estimates also match the set values. , The age effect was not significant (95% of the interval contains 0).

[0116] Simultaneously, a simplified symptom network was constructed, using common adverse symptoms such as "insomnia," "dizziness," "depressed mood," and "hallucinations" as network nodes. Co-occurrence relationships between nodes were calculated based on simulated data to determine edge weights. Betweenness centrality analysis showed that the "insomnia" node had the highest centrality, confirming it as a hub for the transmission of adverse reactions among multiple symptoms.

[0117] Based on the above results, a personalized risk report is generated for each patient. This example verifies that the method of the present invention can correctly identify known interaction signals in a multi-drug scenario and locate potential symptom propagation paths through network analysis, demonstrating strong reliability and clinical interpretability.

[0118] This invention discloses a pharmacovigilance method and system for combined drug use. First, it captures the impact of drug interactions using a Bayesian hierarchical model. The Horseshoe prior effectively handles sparse data, reducing the false negative rate and improving signal detection sensitivity. Second, it incorporates patient covariates and symptom network analysis to provide personalized risk assessment and key symptom alerts, improving the interpretability of conclusions and enhancing clinical relevance. Third, it employs Bayesian methods to validate signals, avoiding the limitations of traditional P-values, and utilizes multidimensional data fusion to provide reliable risk warnings, ensuring the mathematical rigor of statistical inference. Finally, it outputs an intuitive risk report to assist clinical decision-making, demonstrating strong practicality.

[0119] The above description illustrates preferred embodiments of the present invention and helps those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative and should not be construed as limiting the specific implementation of the present invention to these embodiments. For those skilled in the art, several simple deductions and modifications can be made without departing from the inventive concept, and all such modifications should be considered within the protection scope of the present invention.

Claims

1. A pharmacovigilance method for combined drug use, characterized in that, Specifically, it includes: Step 1: Obtain multi-source information from electronic health records; Step 2: Based on multi-source information, construct a three-layer Bayesian hierarchical model using a Bayesian engine to estimate clinical symptom medication parameters; the Bayesian hierarchical model decomposes the overall risk into a single-drug baseline risk layer, a drug interaction layer, and a covariate regulation layer. ,in, Represents the intercept term; This represents the baseline risk parameter for the i-th drug alone; The drug indicator variable represents the i-th drug; This represents the drug interaction parameter between drug i and drug j; Let represent the drug interaction term when the i-th drug is used in combination with the j-th drug; m represents the types of drugs; p represents the number of covariates; and k represents the order of the covariates. Indicates the parameters of the covariate effect; Represents patient covariates; baseline risk parameters for single-drug therapy. Covariate effect parameters We take the conventional normal prior, but for key drug interaction parameters... Horseshoe prior is used for sparsity contraction to simultaneously estimate single-drug effects and multi-drug interaction effects under conditions of limited sample size and numerous variables, and patient covariates are included in the model to consider individual heterogeneity. Step 3: Construct a symptom network based on clinical symptom medication parameters, calculate the betweenness centrality of each node, and identify pivotal symptoms for the spread of adverse reactions; Step 4: Calculate Bayesian factors for clinical symptom medication parameters to assess the probability of adverse reactions and symptom severity. Construct a three-dimensional risk envelope with medication duration as the X-axis, patient age as the Y-axis, and number of comorbidities as the Z-axis. Use non-parametric kernel density estimation to calculate the 95% confidence intervals for the incidence of adverse reactions under each dimension combination. Connect the interval boundaries to form the surface of the risk envelope. Distribute sampling points within the risk envelope according to risk density, where risk density = baseline density × (1 + regional risk value / global risk mean). When the risk density is greater than or equal to the first risk threshold, it is a high-risk area, and the sampling density is increased; when the risk density is less than the first risk threshold, it is a low-risk area, and the sampling density is decreased. Combine the total symptom severity score with the probability of adverse reactions to generate a risk report.

2. The pharmacovigilance method for combined drug use according to claim 1, characterized in that, Data on patients' combined medication use, covariates, and clinical symptoms can be obtained from electronic health records.

3. The pharmacovigilance method for combined drug use according to claim 1, characterized in that, Step 2 includes: Step 201: Construct a Bayesian hierarchical model; Step 202: Perform Bayesian priors on the parameters of the Bayesian hierarchical model; Step 203: Estimate the parameters of the Bayesian hierarchical model using the MCMC method; Step 204: Verify the Bayesian hierarchical model.

4. The pharmacovigilance method for combined drug use according to claim 1, characterized in that, Step 3 includes: Step 301: Construct a symptom network based on multi-source information; Step 302: Calculate the betweenness centrality of each node to determine the pivotal symptoms.

5. The pharmacovigilance method for combined drug use according to claim 4, characterized in that, Symptoms are used as nodes in the symptom network, and the propagation or correlation relationships between symptoms are used as edges. The edge weights are determined by calculating the co-occurrence probability of symptoms, the Pearson correlation coefficient, or the distance metric.

6. The pharmacovigilance method for combined drug use according to claim 4, characterized in that, Graph theory analysis is used to quantify the betweenness centrality of each node. The shortest path between all pairs of nodes in the symptom network is calculated, and the frequency of the node acting as a mediator in the shortest path is used as the betweenness centrality of that node. The more shortest paths a node has, the higher its betweenness centrality. The symptoms represented by that node are often located on the shortest paths between other symptoms, making it a hub for the spread of adverse drug reactions.

7. The pharmacovigilance method for combined drug use according to claim 1, characterized in that, Step 4 includes: Step 401: Calculate Bayesian factors to determine effective drug interaction parameters. To determine the probability of adverse reactions when taking multiple medications; Step 402: Use a standardized rating scale to quantify the severity of symptoms and obtain the symptom severity score; Step 403: Weight the probability of adverse reactions when taking multiple medications with the obtained symptom severity scores to generate a risk report.

8. The pharmacovigilance method for combined drug use according to claim 7, characterized in that, Calculate including the drug interaction parameters The model and models that do not include the drug interaction parameters The marginal likelihood ratio of the model: ， in, Indicating drug interaction parameters Bayesian factor; This represents the probability of observed data occurring when drug i and drug j interact. This represents the probability of the observed data occurring when drug i and drug j do not interact. when At that time, the drug interaction was determined to be significant.

9. A pharmacovigilance method for combined drug use according to claim 7, characterized in that, For each adverse reaction, the severity of symptoms is assessed by converting it into a quantitative score using a standardized rating scale. The higher the score, the greater the correlation between the symptoms and the medication and the greater the severity. The total severity score is obtained by multiplying the quantitative score of each individual symptom by its corresponding weighting coefficient and summing the results. ， Where 's' represents the symptom type; A represents the weighting coefficient of the symptom; A represents the quantitative score of the symptom.

10. A pharmacovigilance system for combined drug administration, applicable to a pharmacovigilance method for combined drug administration, characterized in that, It includes a data acquisition module, a Bayesian inference module, a network analysis module, and a risk assessment module; The data acquisition module obtains information from multiple sources from electronic health records; The Bayesian inference module constructs a three-layered Bayesian hierarchical model based on multi-source information and a Bayesian engine to estimate clinical symptom medication parameters. This Bayesian hierarchical model decomposes overall risk into a single-drug baseline risk layer, a drug interaction layer, and a covariate regulation layer. ,in, Represents the intercept term; This represents the baseline risk parameter for the i-th drug alone; The drug indicator variable represents the i-th drug; This represents the drug interaction parameter between drug i and drug j; Let represent the drug interaction term when the i-th drug is used in combination with the j-th drug; m represents the types of drugs; p represents the number of covariates; and k represents the order of the covariates. Indicates the parameters of the covariate effect; Represents patient covariates; baseline risk parameters for single-drug therapy. Covariate effect parameters We take the conventional normal prior, but for key drug interaction parameters... Horseshoe prior is used for sparsity contraction to simultaneously estimate single-drug effects and multi-drug interaction effects under conditions of limited sample size and numerous variables, and patient covariates are included in the model to consider individual heterogeneity. The network analysis module constructs a symptom network based on clinical symptom medication parameters, calculates the betweenness centrality of each node, and identifies pivotal symptoms for the spread of adverse reactions. The risk assessment module calculates Bayesian factors for clinical symptom medication parameters to assess the probability of adverse reactions and symptom severity. It constructs a three-dimensional risk envelope with medication duration as the X-axis, patient age as the Y-axis, and the number of comorbidities as the Z-axis. Using non-parametric kernel density estimation, it calculates 95% confidence intervals for adverse reaction rates under each dimension combination. Connecting the interval boundaries forms the surface of the risk envelope. Sampling points are allocated within the risk envelope according to risk density: Risk density = Baseline density × (1 + Regional risk value / Global risk mean). When the risk density is greater than or equal to a first risk threshold, it is considered a high-risk area, and the sampling density is increased; when the risk density is less than the first risk threshold, it is considered a low-risk area, and the sampling density is decreased. The total symptom severity score is combined with the probability of adverse reaction occurrence to generate a risk report.

Images