A method and system for evaluating the safety of medication based on data analysis

Abstract

The present application relates to the technical field of medical data analysis, and more particularly to a medication safety evaluation method and system based on data analysis, comprising: collecting information such as patient medical orders, test indexes, diagnoses, courses of disease and allergic history, and combining with drug attribute data to obtain patient time series data according to timestamp alignment; eliminating drug interference effects to establish individualized dynamic baseline and obtain test index deviation information; combining with drug knowledge graph to perform correlation analysis on the deviation information and course of disease text to obtain patient dynamic state correlation graph; identifying risk factors such as drug interactions, organ compatibility, dose accumulation and suspected unnecessary medication according to the above and merging to obtain a comprehensive risk value; determining a risk level in combination with the risk value and its change trend to generate an evaluation report. The medication safety evaluation result generated by the technical scheme of the present application can comprehensively reflect the individualized pathophysiological state of the patient and its dynamic changes, and the accuracy and reliability of the evaluation result are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a method and system for drug safety assessment based on data analysis. Background Technology

[0002] Medication safety assessment is a crucial step in ensuring rational drug use and preventing drug-related adverse events. Its core task is to determine whether a patient's current medication regimen carries risks such as interactions, inappropriate dosage, or organ incompatibility. Currently, the common clinical approach is to check each patient's current medication against the fixed safe dosage range and contraindication list provided in the drug's instructions. If the dosage exceeds a uniform threshold specified in the instructions, an alert is triggered. This approach relies on a uniform standard applied to the entire population, failing to consider individualized dynamic information such as the patient's current organ function and indicator trends. This results in many alerts being triggered for patients whose physical condition is still within tolerable limits, creating an unnecessary alarm burden. Furthermore, because the assessment is based solely on a single, current value, it fails to detect early risk signals where indicators are showing a continuous deterioration trend but have not yet exceeded the fixed threshold, delaying the discovery of some truly concerning risks.

[0003] Furthermore, patients' laboratory test results are influenced by a variety of factors, with medication being only one of them. Current assessment methods often fail to adequately distinguish whether abnormal test results stem from the patient's condition itself or from normal fluctuations related to medication. This leads to some abnormal results being simply equated with disease or medication risk, exacerbating false positives. Simultaneously, unstructured texts such as patient medical records and nursing records contain a wealth of medication-related diagnostic and treatment information. However, current methods generally utilize this information only at a superficial level, such as keyword matching, failing to establish a substantial connection and integration with structured drug knowledge and patient indicator data. This results in incomplete information used for risk assessment.

[0004] Patent publication number CN118888078A discloses a drug combination risk management system based on a knowledge graph. This system uses a BERT-BiLSTM-CRF hybrid model to extract knowledge from real-world data such as medical literature, adverse event databases, and electronic medical records, constructing and iteratively updating a clinical drug interaction knowledge graph. Combining knowledge graph embedding and machine learning algorithms, it predicts the risks of drug combinations and generates medication recommendations based on rule-based reasoning steps, including excluding incompatible drugs, excluding contraindications, determining whether the total mass of components exceeds the dosage standard, and matching alternative drugs or neutralizers / inhibitors. It also establishes a management mode combining automatic intelligent review, pharmacist manual review, and joint decision-making by doctors and patients to generate risk alerts and decision suggestions. However, the risk assessment results provided by this technical solution are mainly general conclusions derived from the drug interaction knowledge graph and historical case samples. It does not differentiate between the varying risk levels exhibited by different patients under the same drug combination due to their individual pathophysiological states, which can easily lead to discrepancies between the risk level assessment and the patient's actual tolerance. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for drug safety assessment based on data analysis, in order to solve the technical problem that drug safety assessment in the prior art fails to fully incorporate individualized dynamic information of patients, resulting in inaccurate risk assessment results.

[0006] The technical solution of this invention is implemented as follows: On one hand, this invention provides a data analysis-based method for evaluating drug safety, comprising the following steps: S1. Collect the patient's medical order records, test result records, diagnosis change records, medical history text information, past allergy history and adverse reaction history, and combine them with drug attribute data. Standardize and align the data according to timestamps to obtain the patient's time-series data. S2. Based on patient time-series data, a baseline modeling method that eliminates drug interference effects is used to obtain individualized dynamic baselines and deviation information of test indicators relative to the baseline; a drug knowledge graph is constructed based on drug attribute data, and an evidence fusion method is used to perform correlation analysis on deviation information and disease course text information to obtain a dynamic patient status correlation graph. S3. Based on the patient dynamic status association graph and drug knowledge graph, identify the risk factors corresponding to drug interactions, organ compatibility, dose accumulation and suspected unnecessary drug use, and merge the risk factors that point to the same monitoring node to obtain the comprehensive risk value corresponding to each monitoring node. S4. Determine the current risk score based on the comprehensive risk value corresponding to the current time point, determine the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time series data, determine the final risk level based on the current risk score and risk change trend, and generate a medication safety assessment report.

[0007] Based on the above technical solutions, preferably, in step S1, the drug attribute data includes drug components, metabolic pathways, incompatibilities, safe dosage range, and the interference relationship between the drug and the test indicators. The interference relationship includes the interference amplitude coefficient and the onset time window. Each historical time node in the patient time series data includes the corresponding drug set, the start time of medication, dosage, route of administration, test indicator value, and diagnostic status of each drug in the drug set.

[0008] Based on the above technical solutions, preferably, step S2 specifically includes the following steps: S21. Based on the drug set corresponding to the current time node in the patient time series data, query whether there is a known test interference relationship between the drug and the test index in the drug knowledge graph. If so, calculate the corresponding expected interference amount according to the test interference relationship and deduct it from the measured value of the test index to obtain the correction value. Based on the correction value, recursively calculate the individualized dynamic baseline and the deviation information of the test index relative to the individualized dynamic baseline. S22. Construct a drug knowledge graph based on drug attribute data. The drug knowledge graph includes nodes and edges connecting the nodes. The nodes carry information on drug components, metabolic pathways, and incompatibilities. The edges include edges that characterize the strength of the adverse reaction association between the drug and organs or test indicators, and edges that characterize the test interference relationship between the drug and test indicators. S23. Based on the drug set corresponding to the current time node, candidate concept domains are determined in the drug knowledge graph. The entity linking method with dual domain constraints is used to link the disease course text information within the candidate concept domains to obtain structured clue evidence carrying link confidence. S24. Using the evidence fusion method, the confidence assignment of structured evidence corresponding to deviation information is combined with the link confidence of structured clue evidence to obtain a dynamic state association map of patients carrying a confirmed status.

[0009] Based on the above technical solution, preferably, S21 specifically includes: calculating the expected interference amount of the test index according to the interference amplitude coefficient and onset time window of each drug in the drug set; subtracting the expected interference amount from the measured value of the test index to obtain a correction value; using a recursive smoothing method, calculating the individualized dynamic baseline and individualized fluctuation amplitude based on the correction value; calculating the standardized deviation according to the correction value, the individualized dynamic baseline, and the individualized fluctuation amplitude; calculating the deviation change rate according to the deviation corresponding to two adjacent observations and the observation time interval; and using the standardized deviation and deviation change rate as deviation information. ; ; in, To verify the index, As a time node, To test indicators exist Correction value after deducting the expected interference amount at any time. To test indicators exist Individualized dynamic baseline at any given moment, This is the individualized dynamic baseline from the previous moment. The preset smoothing coefficient, To test indicators exist Individual fluctuation range at any given moment To prevent the use of pre-defined minimal positive numbers with a denominator of zero, To test indicators exist The degree of deviation at any given moment.

[0010] Based on the above technical solutions, preferably, S23 specifically includes: determining candidate concept domains based on the drug set corresponding to the current time node and the test indicators in the drug knowledge graph that have known associations with each drug; calculating the semantic similarity between the disease course text fragment and each candidate concept within the candidate concept domain, taking the candidate concept with the highest similarity as the link target, and obtaining a semantic matching score; If multiple candidate concepts within a candidate concept domain have a semantic similarity exceeding a preset matching threshold, the semantic matching score is discounted based on the number of candidate concepts that meet the threshold to obtain the link confidence score. ; in, As a clue index, as clues The semantic matching score of the corresponding text segment. This represents the number of candidate concepts whose semantic similarity within the candidate concept domain exceeds a preset matching threshold. as clues Link confidence.

[0011] Based on the above technical solution, preferably, S24 specifically includes: calculating the reliability allocation of structured evidence using a saturated response function according to the deviation degree corresponding to the deviation information; determining the reliability allocation of textual evidence based on the link confidence of structured clue evidence; fusing the reliability allocation of structured evidence with the reliability allocation of textual evidence using evidence combination rules to obtain fused reliability and conflict quality; within a preset verification window, if the fused reliability is not less than a preset reliability threshold, the corresponding clue status is determined to be verified; if the conflict quality is not less than a preset conflict threshold, it is determined to be evidence conflict and marked for manual verification; if the verification window expires and neither of the above two conditions is met, it is determined to be invalid, and the corresponding clue is removed from the patient dynamic status association map. ; ; in, As a clue index, Indicates clues The proposition that the indicated changes in the condition have indeed occurred is true. This indicates a change in the condition in the opposite direction. This represents the uncertain mass that has not yet been assigned to either direction. Assigning reliability to structured evidence. Assigning reliability to textual evidence. To integrate reliability, The quality of the conflict between the two sets of evidence.

[0012] Based on the above technical solutions, preferably, step S3 specifically includes the following steps: S31. Based on the patient dynamic status association graph and drug knowledge graph, risk factors are identified from the dimensions of drug interaction, organ compatibility, dose accumulation and suspected unnecessary medication, and the preliminary risk quantification value corresponding to each risk factor and its source path in the drug knowledge graph are obtained. S32. For two or more risk factors that point to the same monitoring node, extract the terminal edge of the monitoring node in the source path of each risk factor. If the terminal edges are the same, it is determined to be multiple evidences of the same cause. If the terminal edges are different, it is determined to be a single cause with multiple sources. S33. For risk factors with multiple symptoms due to the same cause, the initial risk quantification value is merged using an incremental information approach; for risk factors with multiple clusters of single causes, the initial risk quantification value is merged using an overlay truncation approach; thus, the comprehensive risk value and merging type labeling for each monitoring node are obtained.

[0013] Based on the above technical solutions, preferably, in S33, for risk factors with multiple identifiable causes, the one with the highest preliminary risk quantification value is taken as the base value, and the remaining risk factors are marginally incrementally compensated against the base value according to a preset confirmation increment coefficient to obtain the first comprehensive risk value of the monitoring node; for risk factors with multiple identifiable causes, their preliminary risk quantification values ​​are directly added together and an upper limit cutoff value is set to obtain the second comprehensive risk value of the monitoring node. ; ; in, For a set of unique and convergent risk factors, For risk factor index, Risk factors The corresponding preliminary risk quantification value, for The highest preliminary risk quantification value within the set, A set of risk factors with the same cause and multiple evidences. for The number of risk factors in the set For the preset confirmation increment coefficient, , The first comprehensive risk value corresponding to the monitoring node. This is the second comprehensive risk value corresponding to the monitoring node.

[0014] Based on the above technical solution, preferably, in S4, different confidence weights are assigned to the corresponding comprehensive risk values ​​according to the merging type of each monitoring node, and the weighted risk values ​​of each monitoring node are summed to obtain the current risk score; the current risk score and the historical risk scores corresponding to each historical time node in the patient's time series data are used to construct a score sequence, and the slope of the score sequence changing with time is estimated using a linear regression method to obtain the risk change trend; the current risk score is compared with the preset score interval to obtain the preliminary risk level, and the preliminary risk level is corrected according to the risk change trend to obtain the final risk level. ; in, As a time node, Label the merge type corresponding to the monitoring node. For the pre-set weights of the corresponding merge types, For monitoring nodes in The comprehensive risk value corresponding to each moment. for The current risk score at any given moment.

[0015] In addition, the present invention also provides a data analysis-based medication safety assessment system to achieve the above-described method, comprising: The data integration module is used to collect patients' medical order records, test result records, diagnosis change records, medical history text information, past allergy history and adverse reaction history, and combine them with drug attribute data to standardize and align them according to timestamps to obtain patients' time-series data. The drug knowledge graph construction module is used to build a drug knowledge graph based on drug attribute data. The drug knowledge graph includes information on drug components, metabolic pathways and incompatibilities, as well as the strength of adverse reaction associations between drugs and organs or test indicators, and the test interference relationships between drugs and test indicators. The baseline analysis module is used to process test indicators based on patient time-series data and test interference relationships in the drug knowledge graph. It employs a baseline modeling method that eliminates drug interference effects to obtain individualized dynamic baselines and deviation information of test indicators relative to individualized dynamic baselines. The status association module is used to perform association analysis based on deviation information, disease course text information and drug knowledge graph, and to obtain a dynamic status association graph of patients using evidence fusion methods. The risk merging module is used to identify risk factors corresponding to drug interactions, organ compatibility, dose accumulation, and suspected unnecessary medication based on the patient dynamic status association graph and drug knowledge graph, and to merge risk factors that point to the same monitoring node to obtain the comprehensive risk value corresponding to each monitoring node. The risk assessment module is used to determine the current risk score based on the comprehensive risk value corresponding to the current time point, determine the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time series data, and determine the final risk level based on the current risk score and risk change trend, and generate a medication safety assessment report.

[0016] The present invention has the following advantages over the prior art: (1) This invention performs time-series standardization and alignment of multi-source heterogeneous data such as patients’ medical orders, test indicators, diagnostic changes, and medical history texts, and combines drug knowledge graphs to complete the construction of individualized dynamic baselines, dynamic state association analysis and risk factor merging in sequence, forming a complete assessment link from data integration to risk level determination. The drug safety assessment results can comprehensively reflect the patient’s individualized pathophysiological state and its dynamic changes. Compared with assessment methods that rely on fixed thresholds, the accuracy and reliability of the assessment results are significantly improved.

[0017] (2) This invention removes interference effects from test indicators by processing the test interference relationships recorded in the drug knowledge graph, so that the obtained baseline can distinguish whether the abnormal test indicators are due to changes in the patient's condition or normal fluctuations caused by related medications, thus avoiding misjudging normal physiological fluctuations caused by medications as abnormal risk signals and helping to reduce misjudgments caused by the fluctuations of the indicators themselves.

[0018] (3) This invention associates the disease course text information with the deviation information of test indicators and the drug knowledge graph through evidence fusion, so that the unstructured disease course description information can form a mutually corroborating or complementary relationship with the structured medication and indicator data, which makes up for the limitation of insufficient use of disease description information when relying solely on structured data to judge risk, and improves the rationality of risk attribution.

[0019] (4) This invention not only bases the comprehensive risk value at the current time point on the risk change trend determined by combining the comprehensive risk value corresponding to the patient's historical time points, but also makes a comprehensive judgment on the risk change trend, so that the risk level can reflect the dynamic evolution process of the risk, rather than making a static judgment based solely on the value at a single time point, which is conducive to timely identification of potential risks in the changing trend. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0021] Figure 1 This is a flowchart of the drug safety assessment method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of individualized dynamic baseline construction in this invention; Figure 3 This is a schematic diagram of the drug knowledge graph structure of the present invention; Figure 4 This is a schematic diagram illustrating the risk factor merging method of the present invention; Figure 5 This is a schematic diagram illustrating the risk level determination and intervention feedback of the present invention; Figure 6 This is a block diagram of the drug safety assessment system of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention provides a data analysis-based method for assessing medication safety, comprising the following steps: S1. Collect the patient's medical order records, test result records, diagnosis change records, medical history text information, past allergy history and adverse reaction history, and combine them with drug attribute data. Standardize and align the data according to timestamps to obtain the patient's time-series data. S2. Based on patient time-series data, a baseline modeling method that eliminates drug interference effects is used to obtain individualized dynamic baselines and deviation information of test indicators relative to the baseline; a drug knowledge graph is constructed based on drug attribute data, and an evidence fusion method is used to perform correlation analysis on deviation information and disease course text information to obtain a dynamic patient status correlation graph. S3. Based on the patient dynamic status association graph and drug knowledge graph, identify the risk factors corresponding to drug interactions, organ compatibility, dose accumulation and suspected unnecessary drug use, and merge the risk factors that point to the same monitoring node to obtain the comprehensive risk value corresponding to each monitoring node. S4. Determine the current risk score based on the comprehensive risk value corresponding to the current time point, determine the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time series data, determine the final risk level based on the current risk score and risk change trend, and generate a medication safety assessment report.

[0024] In one embodiment of the present invention, in step S1, the drug attribute data includes drug components, metabolic pathways, incompatibilities, safe dosage range, and the interference relationship between the drug and the test indicators. The interference relationship includes the interference amplitude coefficient and the onset time window. Each historical time node in the patient time series data includes the corresponding drug set, the start time of medication, dosage, route of administration, test indicator value, and diagnostic status of each drug in the drug set.

[0025] Specifically, the medical order records include drug name, dosage, route of administration, and start and end times; the laboratory indicator records include historical numerical sequences of various indicators such as liver function, kidney function, and electrolytes; the diagnosis change records reflect changes in the patient's diagnosis at different time points; the medical record text information comes from unstructured medical records, nursing records, and other medical documents; the safe dosage range and drug incompatibilities in the drug attribute data are taken from the drug instructions; the interference amplitude coefficient and onset time window in the drug-laboratory interference relationship are taken from the instructions or authoritative drug-laboratory interference literature, used to characterize the known, predictable direction and magnitude of interference effect of the drug on specific laboratory indicators. After aligning the above data with a unified timestamp, a patient time-series data covering multiple historical time points is formed. Each historical time point retains the corresponding drug set, the start time of each drug, the dosage, the route of administration, the values ​​of various laboratory indicators, and the current diagnostic status; retaining all patient data rather than just the latest data provides a data foundation for subsequent individualized baseline modeling and risk trend assessment, and retaining the start time of each drug provides a basis for determining whether the drug interference effect has reached its onset time.

[0026] In one embodiment of the present invention, step S2 specifically includes the following steps: S21. Based on the drug set corresponding to the current time node in the patient time series data, query whether there is a known interference relationship between the drug and the test index in the drug knowledge graph. If so, calculate the corresponding expected interference amount according to the interference relationship and deduct it from the measured value of the test index to obtain the correction value. Based on the correction value, recursively calculate the individualized dynamic baseline and the deviation information of the test index relative to the individualized dynamic baseline.

[0027] like Figure 2 As shown, based on the interference amplitude coefficient and onset time window of each drug in the drug set for the test indicator, the expected interference amount of the test indicator is calculated; the expected interference amount is subtracted from the measured value of the test indicator to obtain the correction value; a recursive smoothing method is used to calculate the individualized dynamic baseline and individualized fluctuation amplitude based on the correction value; the standardized deviation is calculated based on the correction value, the individualized dynamic baseline, and the individualized fluctuation amplitude; the deviation rate is calculated based on the deviation corresponding to two adjacent observations and the observation time interval; the standardized deviation and the deviation rate are used as deviation information. Specifically, for the drug set corresponding to the current time node t... For each drug in the test index i, if drug j has a known interference relationship with the test index i, the expected interference caused by the drug is calculated using the following formula: ; in, The interference amplitude coefficient of drug j on test index i is oriented and is taken from the drug instructions or authoritative drug-test interference literature. It serves as an attribute of the test interference relationship edge in the drug knowledge graph. The starting time for this medication administration is taken from the doctor's order record; The onset time window of drug j against test index i is used as another attribute of the test interference relationship edge in the drug knowledge graph; indicator function This is used to ensure that drugs that have not yet reached their onset of action are not included in the expected interference. The measured value of the test index i at time t is used. After deducting the expected interference, the correction value is obtained. Individualized dynamic baselines are calculated recursively for corrected values ​​rather than original measured values: ; Where i is the index of the test index, and t is the time node. To verify the correction value of index i after deducting the expected disturbance at time t, To test the individualized dynamic baseline of index i at time t, This is the individualized dynamic baseline from the previous moment. A preset smoothing coefficient controls the update weight of the latest observation relative to the historical baseline. The individualized volatility of this indicator is estimated synchronously and recursively. ; in, The preset smoothing coefficient for the fluctuation amplitude. To examine the individualized fluctuation range of index i at time t, the standardized deviation is calculated accordingly: ; in, To prevent the use of pre-defined minimal positive numbers with a denominator of zero, To examine the deviation of index i at time t, which characterizes the standardized deviation of the current correction value relative to the individualized baseline, the rate of deviation is further calculated based on the deviation between two adjacent observations and the observation time interval. ; in, This refers to the time point of the last observation. To examine the rate of change of indicator i at time t, reflecting the evolution trend of deviation over time, even if the absolute deviation of a certain observed value is still small, a continuous increase in the deviation in the same positive direction can indicate that the risk is in an early stage of evolution. The deviation... and the rate of change of deviation Together with the measured values, they form a triplet, serving as deviation information corresponding to the test index. , as an attribute of the structured indicator node in the patient dynamic state correlation graph.

[0028] S22. Construct a drug knowledge graph based on drug attribute data. The drug knowledge graph includes nodes and edges connecting the nodes. The nodes carry information on drug components, metabolic pathways, and incompatibilities. The edges include edges that characterize the strength of the adverse reaction association between the drug and organs or test indicators, and edges that characterize the test interference relationship between the drug and test indicators.

[0029] Using drug instructions, authoritative drug databases, and authoritative drug-testing interference literature as sources, entities such as drugs, organs / targets, metabolic enzymes, and adverse reactions are extracted to construct a drug knowledge graph. Nodes in the graph carry information on drug components, metabolic pathways, and drug incompatibilities; edges in the graph include two types of relationships with different properties, each serving different stages: the first type represents edges characterizing the strength of the adverse reaction association between drugs and organs or test indicators, and its attribute values... The first category is obtained by extracting the frequency and severity descriptors corresponding to this relationship from the original drug instructions, converting them into numerical values ​​through a preset mapping terminology, and combining them. The value range is (0,1]. The second category is the edge characterizing the interference relationship between the drug and the test indicator, and its attributes include the aforementioned interference amplitude coefficient. and onset time window In addition, if there are incompatibilities or interactions between drug pairs, the severity level is marked on the corresponding side, divided into four levels: contraindicated, severe, moderate, and mild, for use in subsequent risk factor quantification.

[0030] like Figure 3 As shown, this knowledge graph centers on drug nodes, connecting outwards to nodes representing drug components, metabolic pathways, metabolic enzymes, adverse reactions, organs or systems, and laboratory indicators. Relationships between nodes are represented by semantically labeled edges: drugs involve corresponding metabolic enzymes via metabolic pathways, and these enzymes are primarily distributed in specific organs or systems; adverse reactions associated with drugs affect corresponding organs or systems, and these organs or systems correspond to specific laboratory indicators; drugs also exhibit interactions or incompatibilities, potentially causing cross-drug adverse reaction ripples and interference on the organ systems and laboratory indicators involved by another drug. This knowledge graph provides a structured knowledge foundation for subsequent risk factor identification and risk attribution.

[0031] S23. Based on the drug set corresponding to the current time node, candidate concept domains are determined in the drug knowledge graph. The entity linking method with dual domain constraints is used to link the disease course text information within the candidate concept domains to obtain structured clue evidence carrying link confidence.

[0032] Candidate concept domains are determined based on the drug set corresponding to the current time point and the test indicators of known associations between each drug in the drug knowledge graph. Within the candidate concept domains, the semantic similarity between the disease course text fragments and each candidate concept is calculated, and the candidate concept with the highest similarity is taken as the link target to obtain a semantic matching score. If there are multiple candidate concepts in the candidate concept domain whose semantic similarity exceeds a preset matching threshold, the semantic matching score is discounted according to the number of candidate concepts that meet the threshold to obtain the link confidence score.

[0033] Specifically, symptom descriptions in medical records and nursing records often use vague language and multiple meanings. If we use the common method of linking text fragments with semantic similarity to ontology concepts, searching for the most similar concept across the entire ontology, it becomes difficult to determine which indicator the description should be associated with in the current patient. Therefore, for a patient at time t, let the set of drugs corresponding to their current medication be denoted as . Combined with a set of structured indicator nodes Strength of association with adverse reactions Define candidate conceptual domains: ; In other words, candidate concepts are limited to indicators that are pharmacologically relevant to the patient's current medication and that the patient currently has a structured record of, thus avoiding ambiguity issues arising from searching across the entire ontology. Within the candidate concept domain... Within the framework, the semantic similarity between the disease course text fragment and each candidate concept is calculated, and the one with the highest similarity is selected as the link target to obtain a semantic matching score. If multiple candidate concepts within a candidate concept domain have semantic similarity exceeding a preset matching threshold, it indicates that there are still many candidate concepts that are difficult to distinguish after constraints. In this case, the candidate concept is selected based on the number of candidate concepts that meet the threshold. Discounting the semantic matching score: ; Where k is the clue index. The link confidence of clue k, A larger value indicates a higher degree of ambiguity, and the extraction confidence level is correspondingly lower, thus preventing unresolved ambiguities from being treated as clear clues. Simultaneously, the direction of change indicated by the extracted disease course text fragments is considered. and semantic strength score The extraction results are uniformly structured clue evidence carrying link confidence.

[0034] S24. Using the evidence fusion method, the confidence assignment of structured evidence corresponding to deviation information is combined with the link confidence of structured clue evidence to obtain a dynamic state association map of patients carrying a confirmed status.

[0035] Based on the deviation degree corresponding to the deviation information, the reliability allocation of structured evidence is calculated using a saturation response function; the reliability allocation of textual evidence is determined based on the link confidence of structured clue evidence; the reliability allocation of structured evidence and textual evidence are fused using evidence combination rules to obtain fused reliability and conflict quality; within a preset verification window, if the fused reliability is not less than a preset reliability threshold, the corresponding clue status is determined to be verified; if the conflict quality is not less than a preset conflict threshold, it is determined to be evidence conflict and marked for manual verification; if neither of the above two conditions is met by the expiration of the verification window, it is determined to be invalid, and the corresponding clue is removed from the patient dynamic status association map. Specifically, instead of using a hard threshold to determine whether individualized deviation degree constitutes valid evidence, it uses an analogy to the concentration-effect saturation response relationship in pharmacology, mapping the degree of deviation to continuous evidence strength: ; in, This is a half-saturation constant, and its value is dynamically adjusted according to the certainty of the pharmacological association between the patient's current medication and the test indicator i: first, the value with the highest association strength with indicator i is extracted from the current medication set in the drug knowledge graph. Based on this calculation: ; in , These are preset upper and lower bound constants. When the patient is currently taking medications that are strongly pharmacologically correlated with indicator i, Taking a smaller value makes it easier to obtain a high confidence level for deviations of the same magnitude; conversely, taking a larger value makes it easier to obtain a high confidence level for deviations of the same magnitude. Taking a larger value requires a more significant deviation to achieve the same level of confidence. Combining directional consistency, a structured evidence confidence assignment is defined for clue k: ; ; ; Here, H indicates that the change in condition indicated by clue k has indeed occurred. This indicates a change in the condition in the opposite direction. This represents the uncertainty mass that has not yet been assigned to either direction, and the three satisfy... The textual clues themselves constitute another independent source of evidence, providing a confidence level allocation. , , ,in The link confidence level after ambiguity discounting. The two sets of evidence from independent sources are merged according to the evidence combination rules: ; ; in, Assigning reliability to structured evidence. Assigning reliability to textual evidence. To integrate reliability, The conflict quality between the two sets of evidence represents the degree to which the trends in the structured data contradict the direction of the textual descriptions. The denominator is... This is the result after normalizing and removing conflicting parts. (In the validation window) Within, set a reliability threshold. With conflict threshold :like If the clue status is determined to be confirmed, the fused confidence level replaces the original confidence level and is incorporated into the graph edge weight calculation; if If the evidence is deemed conflicting, it will not proceed to the next scoring stage, triggering a manual verification reminder. It will also be passed to subsequent steps as an independent marker along with the graph. If the verification window... If neither of the above two conditions is met by the deadline, the node is deemed invalid, and the corresponding clue node is removed from the atlas and will no longer participate in subsequent calculations. After fusion, a dynamic status association atlas of patients carrying a confirmed status is obtained, which includes those carrying... , The structured indicator nodes, drug nodes carrying the strength of adverse reaction associations and the relationship of test interference, and clue nodes carrying the confirmation status and fusion reliability are used as inputs for subsequent risk factor identification.

[0036] In one embodiment of the present invention, such as Figure 4 As shown, step S3 specifically includes the following steps: S31. Based on the patient dynamic status association graph and drug knowledge graph, risk factors are identified from the dimensions of drug interaction, organ compatibility, dose accumulation and suspected unnecessary medication, and the preliminary risk quantification value corresponding to each risk factor and its source path in the drug knowledge graph are obtained.

[0037] Risk factors are identified from four dimensions: Drug interaction dimension: For any two drug nodes in the current drug set, the interaction edge between them is retrieved in the drug knowledge graph. The severity level of this edge is read and mapped to fixed values ​​according to four levels: contraindication, severe, moderate, and mild. If a drug pair has multiple interaction records from different sources, the one with the highest severity level is selected. The mapped value is the preliminary risk quantification value of this risk factor. The corresponding associated monitoring node is determined by the target organ attribute marked by this interaction edge. Organ adaptability dimension: For each drug, along the metabolic organ node it points to in the graph, the standardized deviation of the corresponding indicator node of that organ is taken and truncated to the [0,1] interval as the preliminary risk quantification value. If the indicator node carries the aforementioned evidence conflict marker, a conflict marker will be added when outputting the risk factor record. It will not directly participate in subsequent automatic merging calculations but will be transferred to the manual verification channel along with the marker. Dosage accumulation dimension: Drugs currently in use that belong to the same active ingredient or the same metabolic pathway are accumulated using an equivalence conversion factor to obtain the cumulative dose. The safety threshold is not a fixed value from the instructions but is adjusted downwards based on the patient's current standardized deviation of the corresponding organ. The adjusted threshold is: ; in The fixed safety threshold is as indicated in the instruction manual. This is the preset maximum reduction ratio. The deviation saturation value represents the extent to which the cumulative dose exceeds the reduced threshold, serving as the initial risk quantification value for this risk factor. For the suspected unnecessary medication dimension: reusing the aforementioned candidate concept domains, verification is performed from two levels: textual verification (whether the drug appears in semantically supported medication-diagnosis association records extracted from the disease course text) and physiological monitoring verification (whether there are significant deviations in the drug's indication-related indicators). Only when there is no supporting evidence at either level is the drug determined to be a suspected unnecessary medication, and a corresponding label is output. During identification, each risk factor simultaneously records its corresponding source path in the drug knowledge graph, i.e., the directed path formed by a specific drug node pointing to the corresponding monitoring node via a specific edge sequence, for subsequent merging type discrimination.

[0038] S32. For two or more risk factors that point to the same monitoring node, extract the terminal edge of the monitoring node in the source path of each risk factor. If the terminal edges are the same, it is determined to be multiple evidences of the same cause. If the terminal edges are different, it is determined to be a single cause with multiple sources.

[0039] S33. For risk factors with multiple symptoms due to the same cause, the initial risk quantification value is merged using an incremental information approach; for risk factors with multiple clusters of single causes, the initial risk quantification value is merged using an overlay truncation approach; thus, the comprehensive risk value and merging type labeling for each monitoring node are obtained.

[0040] For two or more risk factors that both point to the same monitoring node and do not carry conflict of evidence markers, extract the last edge (terminal edge) immediately adjacent to that monitoring node in their respective source paths. If the terminal edges of two risk factors are the same edge in the graph, meaning the same drug node points to the same monitoring node via the same edge, then these two risk factors are determined to originate from the same drug and exert their effects through the same pharmacological mechanism. They are marked as multiple evidences from the same cause and included in the set. If the terminal edges of two risk factors are different, regardless of whether they are the same drug but pass through different edges, or different drugs pointing to the same monitoring node through different edges, they are determined to represent independent pathogenic pathways converging at the same monitoring node, marked as single-cause co-convergence, and included in the set. This discrimination process is achieved by comparing edge identifiers in the knowledge graph, without relying on a manually pre-labeled classification table. For risk factors with multiple identifiable causes, the one with the highest preliminary risk quantification value is taken as the base value. The remaining risk factors are then marginally incrementally compensated against the base value according to a preset confirmation increment coefficient to obtain the first comprehensive risk value of the monitoring node. For risk factors with multiple identifiable causes, their preliminary risk quantification values ​​are directly added together and an upper limit cutoff value is set to obtain the second comprehensive risk value of the monitoring node. ; ; in, A set of risk factors with the same cause and multiple evidences. Let q be the set of unique, aggregate risk factors, and q be the risk factor index. This represents the preliminary risk quantification value corresponding to risk factor q. for The highest preliminary risk quantification value within the set, for The number of risk factors in the set For the preset confirmation increment coefficient, The first comprehensive risk value corresponding to the monitoring node. This is the second comprehensive risk value corresponding to the monitoring node. The design considerations for the above two merging methods are as follows: In the case of co-occurrence of a single cause, each risk factor represents the superposition of exposures to independent mechanisms, and a superposition truncation method is adopted; in the case of multiple evidences of the same cause, each risk factor originates from repeated confirmation of the same pharmacological mechanism. If simply superimposed, the risk would be overestimated. Therefore, an information increment method is adopted, using the strongest evidence as the base value and compensating for the rest with a marginal decrease, to avoid double counting of the same mechanism. For monitoring nodes that are neither co-occurrence of multiple evidences of the same cause nor co-occurrence of a single cause, but are supported by only a single source, their comprehensive risk value is directly taken as the preliminary risk quantification value of that single risk factor, and labeled as a conventional single type. After the above merging, the comprehensive risk value and merging type label corresponding to each monitoring node are obtained.

[0041] In one embodiment of the present invention, such as Figure 5 As shown, in step S4, different confidence weights are assigned to the corresponding comprehensive risk values ​​according to the merging type of each monitoring node, and the weighted risk values ​​of each monitoring node are summed to obtain the current risk score: ; Where t is the time node. The clustering type is labeled for the monitoring nodes, and its value range includes multiple evidences of the same cause, multiple clusters of a single cause, and regular single cause. For the pre-set weights of the corresponding merge types, The comprehensive risk value of the monitoring node at time t. Let t be the current risk score.

[0042] The current risk score is combined with the historical risk scores corresponding to each historical time point in the patient's time-series data to form a score sequence. A linear regression method is used to estimate the slope of the score sequence over time to obtain the risk change trend. ; Where n represents the number of historical evaluation time points being traced back. For the l-th historical evaluation moment, , These represent the average score for each time point within the window. The risk change trend at time t is represented. Regression calculations are performed directly using actual timestamps instead of equally spaced serial numbers to accommodate the uneven sampling intervals in clinical laboratory testing.

[0043] A pre-defined scoring range is used to divide risk into three levels: low, medium, and high. The boundary values ​​must be such that the low-risk boundary value is less than the high-risk boundary value. Based on the current risk score's location within this range, an initial risk level is determined. This initial risk level is then adjusted based on the aforementioned risk trend: if the initial risk level is medium and the trend is upward, the risk level is raised one level to high risk, triggering an early warning before the score actually crosses the high-risk range; if the initial risk level is high and the trend is downward, the risk level remains unchanged, preventing a premature downgrade of the already determined high-risk level due to a short-term downward trend; otherwise, the initial risk level remains unchanged, resulting in the final risk level.

[0044] Based on the final risk level obtained, and considering the risk factor types (drug interactions, organ compatibility, dose accumulation, suspected unnecessary medication) and the specific drugs and monitoring nodes involved at each monitoring node, corresponding intervention measures are matched from a pre-set intervention suggestion library: for drug interaction risk factors, suggestions to change medication or adjust dosing intervals are matched; for organ compatibility risk factors, suggestions to adjust dosage or increase the frequency of monitoring corresponding organ function indicators are matched; for dose accumulation risk factors, suggestions to reduce dosage or extend dosing intervals are matched; and for suspected unnecessary medication, suggestions to verify the necessity of the corresponding drug prescription are matched. Clues marked as requiring manual verification due to contradictions between structured data and disease course text descriptions are listed separately in the assessment report with the triggering reason, indicating that the corresponding content requires manual judgment rather than directly providing automated intervention suggestions. The final medication safety assessment report includes the final risk level, a detailed list of risk factors corresponding to each monitoring node, the matched intervention suggestions, and a list of items requiring manual verification.

[0045] The study tracks whether the aforementioned intervention recommendations are actually adopted, including whether medication is changed, dosage is adjusted, and monitoring is increased. It also tracks the actual medication outcomes at subsequent time points after adoption or non-adoption, including whether the adverse reactions indicated by the warning actually occurred, or whether the test indicators indeed evolved in the direction of the warning. The adoption status of the intervention recommendations and the actual medication outcomes are re-entered into the patient's time-series data as new historical time-point data for use in the next round of evaluation. Simultaneously, relevant parameters are dynamically adjusted based on the accuracy of the warnings: if a certain type of risk factor repeatedly triggers a warning in the historical record but no corresponding adverse outcome subsequently occurs, it indicates that the current judgment is conservative, and the threshold reduction ratio in the cumulative dosage dimension is adjusted downwards by a preset step size. Or the confirmatory increment coefficient in the merging of multiple evidences due to the same cause This would lead to a more lenient assessment in subsequent judgments; if situations repeatedly occur where the warning level is not reached but adverse outcomes do occur subsequently, it indicates that the current assessment is too lenient, and the assessment should be adjusted upwards according to the preset step size. or Or increase the confidence weight of the corresponding merge type. Furthermore, if the interference amplitude coefficient of a certain drug on a certain test indicator... The strength of the adverse reaction association between a drug and an organ / test indicator. After multiple verifications of actual medication outcomes, a persistent deviation was found between the values ​​and the preset values ​​in the drug knowledge graph. Similarly, the values ​​of the corresponding edge attributes in the drug knowledge graph were synchronously corrected according to a preset step size, ensuring that the calculation basis for the individualized dynamic baseline and evidence fusion process could be continuously adjusted based on the actual patient observation results. The corrected parameters took effect in the next round of evaluation, thus forming a complete closed loop from data collection, through risk factor identification and merging, risk scoring and grading assessment, intervention recommendation matching, to intervention result feedback and parameter correction, and finally back to a new round of data collection. This allows the evaluation method to continuously adjust individualized judgment criteria based on the actual patient medication outcomes, rather than remaining fixed.

[0046] This invention, through its technical means of establishing an individualized dynamic baseline after eliminating the expected interference effect of drugs, and calculating the comprehensive risk value by distinguishing between two merging methods—one for multiple symptoms and one for a single cause—can reduce the risk judgment bias caused by failing to distinguish between expected drug response and actual disease abnormality, and failing to distinguish the source structure of risk factors. This provides a more relevant reference for clinical drug safety assessment based on the individual patient's condition.

[0047] like Figure 6 As shown, the present invention also provides a drug safety assessment system based on data analysis, including a data integration module, a drug knowledge graph construction module, a baseline analysis module, a state association module, a risk merging module, and a risk assessment module.

[0048] The data integration module collects patients' medical orders, laboratory test results, diagnostic change records, medical history, past allergies, and adverse reaction history. It then combines this data with drug attribute data, standardizing and aligning it according to timestamps to obtain the patient's time-series data, which serves as the unified data input for subsequent modules. This module can obtain these raw records through data interfaces with the hospital information system, laboratory information system, and electronic medical record system. It aligns and stitches data from different sources and sampling frequencies according to a unified timestamp. The specific data structure and standardization alignment methods are consistent with the corresponding content in the aforementioned methods section and will not be repeated here.

[0049] The drug knowledge graph construction module is used to build a drug knowledge graph based on drug attribute data. The constructed graph carries information on drug components, metabolic pathways, and incompatibilities, as well as two types of edge attributes: the strength of adverse reaction associations between drugs and organs or test indicators, and the test interference relationships between drugs and test indicators. These attributes are available for querying and use by the baseline analysis module, state association module, and risk merging module during their respective processing. This module can extract knowledge and construct the graph based on drug instructions, authoritative drug databases, and authoritative drug-test interference literature. The specific node and edge composition is consistent with the corresponding content in the aforementioned methods section.

[0050] The baseline analysis module is connected to the data integration module and the drug knowledge graph construction module. It processes laboratory indicators based on patient time-series data using a baseline modeling method that eliminates drug interference effects, obtaining an individualized dynamic baseline and information on the deviation of laboratory indicators from this baseline. During processing, this module calls upon the interference relationships provided by the drug knowledge graph construction module, deducts the expected interference of the current medication on the laboratory indicators, and then recursively calculates the individualized baseline. The specific calculation method is consistent with the corresponding content in the aforementioned methods section.

[0051] The state association module is connected to the baseline analysis module, data integration module, and drug knowledge graph construction module. It is used to perform association analysis based on deviation information, disease course text information, and the drug knowledge graph, employing evidence fusion methods to obtain a dynamic state association graph of the patient. During the association analysis process, this module determines candidate concept domains in the drug knowledge graph based on the current medication set, links entities to the disease course text information, and fuses the structured clue evidence obtained from the links with the structured evidence corresponding to the deviation information to obtain a dynamic state association graph of the patient carrying a confirmed state. The specific entity linking and evidence fusion methods are consistent with the corresponding content in the aforementioned methods section.

[0052] The risk merging module is connected to the state association module and the drug knowledge graph construction module. Based on the patient dynamic state association graph and the drug knowledge graph, it identifies risk factors corresponding to drug interactions, organ compatibility, dose accumulation, and suspected unnecessary medication. It then merges risk factors that commonly point to the same monitoring node to obtain a comprehensive risk value for each monitoring node. During the merging process, this module distinguishes whether risk factors originate from the same pharmacological mechanism, employing different merging methods to calculate the comprehensive risk value. The specific risk factor identification and merging methods are consistent with the corresponding content in the aforementioned methods section.

[0053] The risk assessment module is connected to the risk merging module and the data integration module. It determines the current risk score based on the comprehensive risk value corresponding to the current time point, determines the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time-series data, and determines the final risk level based on the current risk score and risk change trend, generating a medication safety assessment report. This module can further match corresponding intervention recommendations and track the adoption of intervention recommendations and subsequent medication outcomes. The tracking results are fed back to correct relevant parameters involved in the assessment process. The specific methods for risk score calculation, level correction, intervention recommendation matching, and parameter correction are consistent with the corresponding content in the aforementioned methods section.

[0054] The modules mentioned above are connected sequentially according to the data flow relationship. The outputs of the data integration module and the drug knowledge graph construction module serve as the inputs of the baseline analysis module, the state association module, and the risk merging module, respectively. The processing results of each module are passed to the risk assessment module at each level. Finally, the risk assessment module outputs a drug safety assessment report, realizing a complete processing flow from multi-source data collection to risk assessment result output.

[0055] 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 assessing medication safety based on data analysis, characterized in that, Includes the following steps: S1. Collect the patient's medical order records, test result records, diagnosis change records, medical history text information, past allergy history and adverse reaction history, and combine them with drug attribute data. Standardize and align the data according to timestamps to obtain the patient's time-series data. S2. Based on patient time-series data, a baseline modeling method that eliminates drug interference effects is used to obtain individualized dynamic baselines and deviations of test indicators from the baseline. A drug knowledge graph was constructed based on drug attribute data. Evidence fusion method was used to perform correlation analysis on deviation information and disease course text information to obtain a dynamic status correlation graph of patients. S3. Based on the patient dynamic status association graph and drug knowledge graph, identify the risk factors corresponding to drug interactions, organ compatibility, dose accumulation and suspected unnecessary drug use, and merge the risk factors that point to the same monitoring node to obtain the comprehensive risk value corresponding to each monitoring node. S4. Determine the current risk score based on the comprehensive risk value corresponding to the current time point, determine the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time series data, determine the final risk level based on the current risk score and risk change trend, and generate a medication safety assessment report.

2. The drug safety assessment method based on data analysis as described in claim 1, characterized in that, In step S1, the drug attribute data includes drug components, metabolic pathways, incompatibilities, safe dosage range, and the interference relationship between the drug and the test indicators. The interference relationship includes the interference amplitude coefficient and the onset time window. Each historical time node in the patient time series data includes the corresponding drug set, the start time of medication, dosage, route of administration, test indicator value, and diagnostic status of each drug in the drug set.

3. The method for drug safety assessment based on data analysis as described in claim 2, characterized in that, Step S2 specifically includes the following steps: S21. Based on the drug set corresponding to the current time node in the patient time series data, query whether there is a known test interference relationship between the drug and the test index in the drug knowledge graph. If so, calculate the corresponding expected interference amount according to the test interference relationship and deduct it from the measured value of the test index to obtain the correction value. Based on the correction value, recursively calculate the individualized dynamic baseline and the deviation information of the test index relative to the individualized dynamic baseline. S22. Construct a drug knowledge graph based on drug attribute data. The drug knowledge graph includes nodes and edges connecting the nodes. The nodes carry information on drug components, metabolic pathways, and incompatibilities. The edges include edges that characterize the strength of the adverse reaction association between the drug and organs or test indicators, and edges that characterize the test interference relationship between the drug and test indicators. S23. Based on the drug set corresponding to the current time node, candidate concept domains are determined in the drug knowledge graph. The entity linking method with dual domain constraints is used to link the disease course text information within the candidate concept domains to obtain structured clue evidence carrying link confidence. S24. Using the evidence fusion method, the confidence assignment of structured evidence corresponding to deviation information is combined with the link confidence of structured clue evidence to obtain a dynamic state association map of patients carrying a confirmed status.

4. The drug safety assessment method based on data analysis as described in claim 3, characterized in that, S21 specifically includes: calculating the expected interference amount of the test indicator based on the interference amplitude coefficient and onset time window of each drug in the drug set; subtracting the expected interference amount from the measured value of the test indicator to obtain the correction value; using a recursive smoothing method, calculating the individualized dynamic baseline and individualized fluctuation amplitude based on the correction value; calculating the standardized deviation based on the correction value, individualized dynamic baseline, and individualized fluctuation amplitude; calculating the deviation rate based on the deviation and observation time interval between two adjacent observations; and using the standardized deviation and deviation rate as deviation information. ； ； in, To verify the index, As a time node, To test indicators exist Correction value after deducting the expected interference amount at any time. To test indicators exist Individualized dynamic baseline at any given moment, This is the individualized dynamic baseline from the previous moment. The preset smoothing coefficient, To test indicators exist Individual fluctuation range at any given moment To prevent the use of pre-defined minimal positive numbers with a denominator of zero, To test indicators exist The degree of deviation at any given moment.

5. The drug safety assessment method based on data analysis as described in claim 3, characterized in that, S23 specifically includes: determining candidate concept domains based on the drug set corresponding to the current time node and the test indicators of known associations between each drug in the drug knowledge graph; calculating the semantic similarity between the disease course text fragment and each candidate concept within the candidate concept domain, taking the candidate concept with the highest similarity as the link target, and obtaining the semantic matching score; If multiple candidate concepts within a candidate concept domain have a semantic similarity exceeding a preset matching threshold, the semantic matching score is discounted based on the number of candidate concepts that meet the threshold to obtain the link confidence score. ； in, As a clue index, as clues The semantic matching score of the corresponding text segment. This represents the number of candidate concepts whose semantic similarity within the candidate concept domain exceeds a preset matching threshold. as clues Link confidence.

6. The drug safety assessment method based on data analysis as described in claim 3, characterized in that, S24 specifically includes: calculating the reliability allocation of structured evidence using a saturated response function based on the deviation degree corresponding to the deviation information; determining the reliability allocation of textual evidence based on the link confidence of structured clue evidence; fusing the reliability allocation of structured evidence with the reliability allocation of textual evidence using evidence combination rules to obtain fused reliability and conflict quality; within a preset verification window, if the fused reliability is not less than a preset reliability threshold, the corresponding clue status is determined to be verified; if the conflict quality is not less than a preset conflict threshold, it is determined to be evidence conflict and marked for manual verification; if the verification window expires and neither of the above two conditions is met, it is determined to be invalid, and the corresponding clue is removed from the patient dynamic status association map. ； ； in, As a clue index, Indicates clues The proposition that the indicated changes in the condition have indeed occurred is true. This indicates a change in the condition in the opposite direction. This represents the uncertain mass that has not yet been assigned to either direction. Assigning reliability to structured evidence. Assigning reliability to textual evidence. To integrate reliability, The quality of the conflict between the two sets of evidence.

7. The drug safety assessment method based on data analysis as described in claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Based on the patient dynamic status association graph and drug knowledge graph, risk factors are identified from the dimensions of drug interaction, organ compatibility, dose accumulation and suspected unnecessary medication, and the preliminary risk quantification value corresponding to each risk factor and its source path in the drug knowledge graph are obtained. S32. For two or more risk factors that point to the same monitoring node, extract the terminal edge of the monitoring node in the source path of each risk factor. If the terminal edges are the same, it is determined to be multiple evidences of the same cause. If the terminal edges are different, it is determined to be a single cause with multiple sources. S33. For risk factors with multiple symptoms due to the same cause, the initial risk quantification value is merged using an incremental information approach; for risk factors with multiple clusters of single causes, the initial risk quantification value is merged using an overlay truncation approach; thus, the comprehensive risk value and merging type labeling for each monitoring node are obtained.

8. The drug safety assessment method based on data analysis as described in claim 7, characterized in that, In S33, for risk factors with multiple evidences for the same cause, the one with the highest preliminary risk quantification value is taken as the base value, and the remaining risk factors are marginally incrementally compensated for the base value according to the preset confirmation increment coefficient to obtain the first comprehensive risk value of the monitoring node. For risk factors that are both singular and concurrent, their initial risk quantification values ​​are directly added together and an upper limit cutoff value is set to obtain the second comprehensive risk value of the monitoring node: ； ； in, For a set of unique and convergent risk factors, For risk factor indexing, Risk factors The corresponding preliminary risk quantification value, for The highest preliminary risk quantification value within the set, A set of risk factors with the same cause and multiple evidences. for The number of risk factors in the set For the preset confirmation increment coefficient, , The first comprehensive risk value corresponding to the monitoring node. This is the second comprehensive risk value corresponding to the monitoring node.

9. The drug safety assessment method based on data analysis as described in claim 7, characterized in that, In S4, different confidence weights are assigned to the corresponding comprehensive risk values ​​based on the merging type of each monitoring node, and the weighted risk values ​​of each monitoring node are summed to obtain the current risk score. The current risk score is then combined with the historical risk scores corresponding to each historical time point in the patient's time-series data to form a score sequence. A linear regression method is used to estimate the slope of the score sequence over time to obtain the risk change trend. The current risk score is compared with a preset score interval to obtain the initial risk level, and the initial risk level is revised based on the risk change trend to obtain the final risk level. ； in, As a time node, Label the merge type corresponding to the monitoring node. For the pre-set weights of the corresponding merge types, For monitoring nodes in The comprehensive risk value corresponding to each moment, for The current risk score at any given moment.

10. A medication safety assessment system based on data analysis, characterized in that, The system is used to implement the method as described in any one of claims 1-9, comprising: The data integration module is used to collect patients' medical order records, test result records, diagnosis change records, medical history text information, past allergy history and adverse reaction history, and combine them with drug attribute data to standardize and align them according to timestamps to obtain patients' time-series data. The drug knowledge graph construction module is used to build a drug knowledge graph based on drug attribute data. The drug knowledge graph includes information on drug components, metabolic pathways and incompatibilities, as well as the strength of adverse reaction associations between drugs and organs or test indicators, and the test interference relationships between drugs and test indicators. The baseline analysis module is used to process test indicators based on patient time-series data and test interference relationships in the drug knowledge graph. It employs a baseline modeling method that eliminates drug interference effects to obtain individualized dynamic baselines and deviation information of test indicators relative to individualized dynamic baselines. The status association module is used to perform association analysis based on deviation information, disease course text information and drug knowledge graph, and to obtain a dynamic status association graph of patients using evidence fusion methods. The risk merging module is used to identify risk factors corresponding to drug interactions, organ compatibility, dose accumulation, and suspected unnecessary medication based on the patient dynamic status association graph and drug knowledge graph, and to merge risk factors that point to the same monitoring node to obtain the comprehensive risk value corresponding to each monitoring node. The risk assessment module is used to determine the current risk score based on the comprehensive risk value corresponding to the current time point, determine the risk change trend by combining the comprehensive risk values ​​corresponding to each historical time point in the patient's time series data, and determine the final risk level based on the current risk score and risk change trend, and generate a medication safety assessment report.

Images