A drug safety signal mining method and system based on multi-source data

Abstract

The present application relates to the field of data processing, and more particularly to a drug safety signal mining method and system based on multi-source data. The method comprises the following steps: first, constructing a longitudinal event sequence containing medication and diagnosis events for each patient; then calculating the time sequence correlation factor of medication and subsequent adverse events, which is negatively correlated with the time interval between the two events and the number of mixed diagnosis events within the interval; at the same time, based on the matching degree of the event time interval and the known latency probability distribution, the latency rationality weight is obtained; then, combined with the medication dose, the time sequence correlation factor and the latency rationality weight, the influence of related medication events before the occurrence of adverse events is accumulated, and the cumulative correlation energy of specific drugs and adverse events is calculated; finally, the cumulative correlation energy of the patient group is analyzed in a non-proportional manner to realize drug safety signal mining and improve signal identification accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method and system for mining drug safety signals based on multi-source data. Background Technology

[0002] Post-marketing safety monitoring of drugs is a crucial link in safeguarding public health. Its core task is to promptly and accurately identify potential adverse drug event signals from massive amounts of data. Traditional signal mining methods, such as disproportionality analysis (DPA), mainly rely on static reporting data from spontaneous adverse drug event reporting systems, and analyze the co-occurrence frequency of drugs and events by constructing contingency tables.

[0003] However, existing technologies have significant limitations. First, traditional DPA methods typically treat data as static, cross-sectional event counts, constructing a simple two-dimensional contingency table of drug-adverse events for statistical analysis. This method completely ignores the temporal order of events, failing to distinguish whether a medication event occurred before or after an adverse event, and even less able to quantify the impact of time intervals on the strength of causal association. A medication record from a year prior to an adverse event and a record from a week prior are treated equally in traditional methods, which seriously violates the basic logic of pharmacology and epidemiology. Second, in the real world, patients' clinical conditions are extremely complex. An adverse event may not be caused by the target drug, but by the patient's underlying diseases, concomitant medications, or other complications. Traditional DPA methods lack effective mechanisms to identify and quantify the interference of these confounding factors, easily generating a large number of false positive signals and increasing the burden of subsequent medical evaluation. Third, different drugs trigger specific adverse reactions, and their latency periods often follow certain pharmacological patterns. Traditional methods cannot integrate this prior clinical knowledge. A drug-adverse event pair whose time interval does not conform to the known latency pattern may still be identified as a signal due to a coincidence in counting, which reduces the biological rationale and credibility of the signal.

[0004] Therefore, existing drug safety signal mining methods have significant shortcomings in processing fusion timing, controlling confounding bias information, and utilizing prior knowledge. Summary of the Invention

[0005] To address the problem of how to accurately analyze drug information that integrates time-series information and controls confounding biases, and to achieve accurate drug safety signal analysis, this invention provides a drug safety signal mining method and system based on multi-source data.

[0006] In a first aspect, the present invention provides a method for mining drug safety signals based on multi-source data, employing the following technical solution:

[0007] A method for mining drug safety signals based on multi-source data, comprising the following steps:

[0008] a. Construct a longitudinal event sequence for each patient that includes at least medication and diagnostic events;

[0009] b. For any medication event and subsequent adverse event in the longitudinal event sequence, calculate a time-series correlation factor. The time-series correlation factor is negatively correlated with the time interval between the two events and the number of confounding diagnostic events within the time interval.

[0010] c. For the medication event and adverse event, calculate a latency period rationality weight, which is obtained based on the degree of matching between the time interval between the two events and a known latency probability distribution characterizing the drug causing the adverse event;

[0011] d. Combining the drug dosage, the time-series correlation factor, and the reasonableness weight of the latency period, calculate a cumulative correlation energy between a specific drug and an adverse event. The cumulative correlation energy is the result of summing the effects of all relevant drug-related events prior to the occurrence of the adverse event.

[0012] e. Based on the cumulative correlation energy calculated in the patient population, a disproportionate analysis is performed to identify potential drug safety signals.

[0013] This invention constructs an index by comprehensively quantifying temporal proximity, confounding factors, latency rationality, and dose-effect. Ultimately, disproportionate analysis based on this more reliable and clinically significant index effectively suppresses false positive signals and more sensitively identifies potential risks with real clinical significance that are overlooked by traditional methods, achieving a significant leap forward in signal mining quality.

[0014] Preferably, the method for obtaining the longitudinal event sequence includes:

[0015] Patient medication, diagnosis and test records are integrated from multiple databases, and data cleaning and conceptual standardization are performed. Finally, the longitudinal event sequence is obtained by sorting by timestamp.

[0016] Preferably, the heterogeneous diagnostic events are identified by matching in a pre-built reference library containing known strong correlation rules between diagnoses and adverse events.

[0017] This invention achieves objective identification of disease diagnoses that may themselves lead to adverse events by pre-mining and constructing a reference library containing strong correlation rules between diagnosis and adverse events. Including the number of these confounding events in the correlation factor calculation is equivalent to denoising the original correlation strength, making the final correlation assessment more purely reflect the impact of the drug itself.

[0018] Preferably, the latency probability distribution is a Gaussian function determined based on the expected value and standard deviation of the latency for a specific drug-adverse event pair in the clinical knowledge base.

[0019] This invention moves beyond blindly mining data; instead, it matches observed medication-adverse event time intervals with known pharmacological patterns. This assigns higher weight to event pairs that are more temporally consistent with pharmacological expectations, while suppressing event pairs that are temporally incompatible. This combination of data-driven and knowledge-guided approaches ensures that the ultimately selected signals are not only statistically significant but also more clinically reliable.

[0020] Preferably, the method for calculating the cumulative correlated energy is as follows:

[0021] For each relevant medication event prior to the occurrence of the adverse event, the medication dosage, the corresponding time-series correlation factor, and the reasonableness weight of the latency period are multiplied together, and all products are summed up.

[0022] This invention ingeniously integrates three core dimensions: drug dosage, time-series correlation factors, and the rationality weight of latency period. By applying such precise weighting to each dose before the occurrence of an adverse event and accumulating the results, it achieves accurate capture of the cumulative toxic effects caused by long-term, repeated medication.

[0023] Preferably, before performing the disproportionate analysis, the method further includes:

[0024] The cumulative correlation energy calculated for a specific drug-adverse event pair across all patients will be aggregated to construct a correlation energy matrix with drugs as rows and adverse events as columns.

[0025] Preferably, the disproportionate analysis includes calculating a time-series proportional reporting ratio, which is calculated by dividing the proportion of the energy of a specific drug-adverse event pair in the total energy generated by the drug by the proportion of the total energy of the adverse event in the total energy of the entire matrix.

[0026] This invention takes into account that each cell of the input matrix is ​​not an uncorrelated discrete signal, but an energy value containing rich clinical information. Therefore, the calculated disproportionate signal strength can more accurately reflect the true risk association strength and effectively distinguish between real strong signals and pseudo signals caused by various biases.

[0027] Preferably, the steps for identifying potential drug safety signals include:

[0028] The calculated time-series proportion reporting ratio is compared with a preset threshold, and drug-adverse event pairs that exceed the threshold are screened as potential safety signals.

[0029] Preferably, the time interval effect in the time-series correlation factor is modeled by a function that decays exponentially with the time interval, the decay rate of which is set based on the pharmacological half-life information of the drug.

[0030] Secondly, this invention provides a drug safety signal mining system based on multi-source data, employing the following technical solution:

[0031] A drug safety signal mining system based on multi-source data includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned drug safety signal mining method based on multi-source data is implemented.

[0032] By adopting the above technical solution, a computer program is generated from the above-mentioned method for mining drug safety signals based on multi-source data and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0033] The present invention has the following technical effects:

[0034] This invention can systematically integrate temporal proximity, confounding factors, and matching degree with prior clinical knowledge. This makes the discovered correlation signals no longer simple statistical coincidences, but potential causal associations with higher evidence strength after being tested by quasi-causal logical chains, thereby significantly suppressing the generation of false positive signals from the source.

[0035] Furthermore, by accumulating correlation energy, this method cleverly integrates multiple key clinical dimensions, such as drug dosage, temporal effects, confounding interference, and latency patterns, into a single, continuous quantitative indicator through a mathematical model. This enables the method not only to identify the presence or absence of a signal but also to accurately measure its strength, and is particularly effective in identifying complex signals with cumulative toxic effects caused by long-term, repeated drug use.

[0036] Furthermore, this invention replaces the original event count matrix with a correlation energy matrix containing rich clinical information, ensuring that subsequent time-series proportional reporting ratio calculations are based on higher-quality data that has undergone deep information purification and weighting. This allows genuine safety signals to stand out with a higher signal-to-noise ratio, significantly improving the sensitivity and specificity of signal mining. Attached Figure Description

[0037] Figure 1 This is a flowchart of a method for mining drug safety signals based on multi-source data provided in an embodiment of the present invention;

[0038] Figure 2 The cumulative correlation energy matrix provided in the embodiments of the present invention;

[0039] Figure 3 The above is a comparison chart of detection performance provided in an embodiment of the present invention. Detailed Implementation

[0040] This invention discloses a method for mining drug safety signals based on multi-source data, referring to... Figure 1 This includes steps S1-S5:

[0041] S1: Obtain the patient's longitudinal event sequence.

[0042] It should be noted that in order to conduct effective temporal causal analysis, patient data scattered across different systems and in different formats must first be integrated into a unified data sequence organized according to time flow.

[0043] Specifically, patient-related data of each type is collected from multiple source databases, including but not limited to electronic health records (EHRs), hospital information systems (HIS), adverse drug event reporting systems (AERS), and health insurance claims databases. The types of data acquired primarily include patient demographic information, medication records, diagnostic records, laboratory test results, and vital sign measurement records.

[0044] The collected raw data undergoes preprocessing, including data cleaning, format standardization, and concept standardization. For example, different brand names of drugs are uniformly mapped to their active ingredients, and different disease diagnosis codes are unified into a standard terminology set.

[0045] Based on the preprocessed data, a longitudinal event sequence strictly ordered by timestamps is constructed for each patient. The sequence consists of a series of time event tuples, in the form of... ,in It refers to a specific clinical event (such as medication, diagnosis, or testing). The precise timestamp of the event, and satisfying .

[0046] Understandably, by constructing a longitudinal event sequence, static, isolated data points can be transformed into a dynamic, coherent flow of events, providing a data foundation for all subsequent time-series analysis, causal relationships, and correlations.

[0047] S2: Perform time-series correlation analysis on medication events and adverse events in the longitudinal event sequence to obtain time-series correlation factors.

[0048] It is important to note that analyzing drug safety requires examining the correlation between adverse events and medication-related events. However, simply because a medication-related event occurred before an adverse event does not automatically imply a link between the two. The length of the time interval between them and the presence of other interfering factors will affect the correlation analysis. Therefore, these two factors must be eliminated when conducting correlation analyses of medication-related events and adverse events.

[0049] Preferably, as an example, time-series correlation analysis is performed on medication events and adverse events in a longitudinal event sequence to obtain time-series correlation factors, including:

[0050] Let any adverse event be obtained from the vertical event sequence and denoted as . In this adverse event The previous acquisition of the i-th medication event is denoted as .

[0051] Calculate the time-series association factor between adverse events and medication events:

[0052]

[0053] in, Indicating a medication incident and adverse events The temporal correlation factors between them. and They represent and The timestamp of the event. Indicates within the time window Internal factors may lead to adverse events. The number of other confounding diagnostic events, This represents a time decay constant, the value of which can be set based on statistical information about the half-life of drugs in the pharmacology knowledge base. exp() represents an exponential function with the natural constant as the base. This is an exponentially decaying term, which is calculated by considering the medication events. and adverse events Substituting the time interval into a function that decays exponentially with the time interval. Calculated in This is the time interval variable in the function. Used to reflect the decay rate of the function.

[0054] Understandably, the exponentially decaying term It is used to exclude the influence of time span between adverse events and medication events in association analysis, and to count confounding events. This is used to exclude other diagnostic interferences in the association analysis between adverse events and medication events. Therefore The higher the value, the greater the potential for a single medication event to have a direct, undisturbed impact on a subsequent adverse event, thus providing a more accurate and undisturbed data foundation for subsequent calculations.

[0055] It should be added that, within the time window Internal factors may lead to adverse events. Other methods for obtaining the number of provocative diagnostic events include:

[0056] A longitudinal event sequence of several patients is obtained. A backtracking analysis window of a preset length is set for the temporal association rule mining algorithm. Based on the backtracking analysis window, the longitudinal event sequence of each patient is analyzed using the temporal association rule mining algorithm to obtain frequent itemsets that are greater than the preset support lower limit. Association rules that are greater than both the preset lift lower limit and the preset confidence lower limit are obtained from the frequent itemsets and recorded as reference association rules. All reference association rules are used to form a reference association rule library.

[0057] In the time window Each diagnostic event and adverse event within the event pair constitutes an event pair. If the event pair exists in the reference association rule base, then the diagnostic event in that event pair is considered the cause of the adverse event. Other confounding diagnostic events, statistically obtained time windows Internal factors leading to adverse events The number of other confounding diagnostic events.

[0058] S3: Analyze the factors influencing the latency period of adverse drug reactions to obtain the weight of the rationality of the latency period.

[0059] It should be noted that different drugs have a latency period for adverse reactions. If the influence of the latency period is not considered, it cannot accurately reflect the correlation between the adverse event and the medication event.

[0060] It should be further explained that the latency period of medication generally follows a certain distribution pattern. The closer the time interval between medication and adverse events is to this distribution pattern, the greater the likelihood that the adverse reaction is caused by the medication and the higher its probability of occurrence. The further away from the expected value of the drug latency period, the lower its probability of occurrence.

[0061] Preferably, as an example, the calculation of the latency period reasonableness weight includes:

[0062] First, obtain the latency distribution parameters of drug D causing specific adverse events A obtained from clinical trials, i.e., the expected latency. and the standard deviation of the incubation period A Gaussian function with mean equal to the expected latency period and standard deviation equal to the standard deviation of the latency period is constructed as the latency probability distribution, which reflects the time distribution pattern of adverse drug reaction event A.

[0063] Then, the medication event and adverse events The time interval is input into the latency probability distribution to calculate the latency rationality weight:

[0064]

[0065] in, Indicating a medication incident and adverse events Reasonable weighting of incubation period between them express and The time interval between occurrences and This represents the expected latency period and the standard deviation of the latency period.

[0066] Understandable, Reflects and When the time interval between occurrences follows a pattern consistent with the incubation period distribution, the larger this value, the better. and The greater the degree to which the time interval between occurrences conforms to the distribution pattern of the incubation period, therefore... and The greater the correlation between them, the more indicative of the medication event. Leading to adverse events The greater the likelihood, the higher the probability.

[0067] S4: Based on the reasonable weight of the latency period and the time-series correlation factor, the correlation between drugs and adverse reactions under the continuous accumulation of drug usage is analyzed to obtain the cumulative correlation energy, and a correlation energy matrix is ​​constructed based on the cumulative correlation energy.

[0068] S40: Based on the reasonable weight of the latency period and the time-series correlation factor, the cumulative correlation energy is obtained by analyzing the correlation between drugs and adverse reactions under the continuous accumulation of drug usage.

[0069] It should be noted that the above steps only analyze the association between a single drug use event and adverse reaction, and do not consider the cumulative association between the drug and adverse reactions.

[0070] Preferably, as an example, based on the reasonableness weight of the latency period and the time-series correlation factor, the cumulative correlation energy is obtained by analyzing the association between drugs and adverse reactions under continuous drug usage accumulation, including:

[0071]

[0072] in, This indicates the number of times drug D was used before the adverse reaction occurred. This is the dosage of drug D used for the i-th time. These represent medication events. and adverse events The temporal correlation factors and the reasonable weight of the latency period between them Indicates drug D and adverse reactions The cumulative energy of the relationship between them.

[0073] Understandable, This reflects the association between the i-th medication event and the adverse reaction event. Since the dosage is also related to the intensity of the adverse reaction, the cumulative dosage of the drug is incorporated into the analysis of the relationship between the drug and the adverse reaction to obtain the association under cumulative drug dosage.

[0074] S41: Construct a correlation energy matrix based on cumulative correlation energy.

[0075] Preferably, as an example, a correlation energy matrix is ​​constructed based on the cumulative correlation energy, including:

[0076] Drug D and adverse reactions were calculated based on each patient's longitudinal event sequence. The cumulative correlation energy, which relates drug D received by all patients to adverse reactions The cumulative correlation energy is added together to obtain the relationship between drug D and adverse reactions. The total cumulative correlation energy between them.

[0077] Construct a correlation energy matrix The rows of the matrix represent drugs, the columns represent adverse events, and the cells represent... The value is related to drug D and adverse reactions. The total cumulative correlation energy between them.

[0078] Figure 2 To accumulate the correlation energy matrix, the image shows that this solution can accurately locate the most noteworthy risk hotspots from a large, messy, and sparse database.

[0079] S5: Analyze the correlation energy matrix to obtain the proportional report ratio.

[0080] Preferably, as an example, the proportional reporting ratio is obtained by analyzing the correlation energy matrix, including:

[0081] Calculate the time series proportional reporting ratio:

[0082]

[0083] in, Represents a cell in the correlation energy matrix The value, This represents the sum of all energy values ​​in the row containing drug D. This represents the sum of all energy values ​​in the column containing adverse event A. This represents the sum of all energy values ​​in the entire associated energy matrix.

[0084] Understandable, This reflects the percentage of the total energy generated by drug D that is attributable to adverse event A. This value reflects the percentage of the total energy generated by all drugs in the entire database that is attributable to adverse event A; it serves as a reference percentage. This reflects that, compared to the reference ratio, the larger this value is, the greater the proportion of energy from drug D causing adverse event A is higher than the reference ratio. Therefore, the greater the disproportionate concentration of drug D's energy on adverse event A, and consequently the greater the likelihood that drug D will trigger adverse reaction A.

[0085] A lower limit threshold is set, and drug-adverse reaction pairs that exceed the lower limit threshold are screened out. The drug-adverse reaction pairs are then sorted in descending order of their time-series reporting ratios and provided to pharmacovigilance experts for subsequent medical evaluation.

[0086] Figure 3 The image shows a performance comparison. As can be seen from the image, this solution has a higher true positive rate under the same standard. In other words, this solution has higher detection accuracy under the same standard.

[0087] This invention also discloses a drug safety signal mining system based on multi-source data, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a drug safety signal mining method based on multi-source data according to this invention is implemented.

[0088] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0089] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (DRAM), high-bandwidth memory, hybrid memory cube, etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

Claims

1. A method for drug safety signal mining based on multi-source data, characterized in that, The method comprises the steps of: a. constructing a longitudinal event sequence for each patient, which at least contains medication, diagnostic events; b. calculating a time correlation factor for any medication event and subsequent adverse event in the longitudinal event sequence, which is negatively correlated with the time interval between the two events and the number of confounding diagnostic events existing in the time interval; c. calculating a latency reasonableness weight for the medication event and adverse event, which is based on the matching degree between the time interval and a known latency probability distribution representing the drug-induced adverse event; d. combining the medication dose, the time correlation factor and the latency reasonableness weight to calculate a cumulative correlation energy for a specific drug-adverse event pair, which is the result of accumulating the effects of all relevant medication events before the occurrence of the adverse event; e. constructing a correlation energy matrix with drugs as rows and adverse events as columns based on the cumulative correlation energy calculated in the patient population, and performing disproportionality analysis to identify potential drug safety signals; The disproportionality analysis comprises calculating a time proportion reporting ratio, which is calculated by dividing the proportion of the energy of a specific drug-adverse event pair in the total energy of the drug by the proportion of the total energy of the adverse event in the total energy of the whole matrix. 2.The method of claim 1, wherein, The method for obtaining the longitudinal event sequence comprises: Integrating the medication, diagnosis and test records of patients from multiple source databases, performing data cleaning and concept standardization, and finally sorting by timestamp to obtain the longitudinal event sequence. 3.The drug safety signal mining method based on multi-source data according to claim 1, characterized in that, The confounding diagnostic events are identified by matching in a pre-constructed reference library containing known diagnostic-adverse event strong association rules. 4.The method of claim 1, wherein, The latency probability distribution is a Gaussian function determined based on the expected value and standard deviation of the latency of a specific drug-adverse event pair in the clinical knowledge base.

5. The method of claim 1, wherein the method is characterized by, The calculation method of the cumulative correlation energy is: For each relevant medication event before the occurrence of the adverse event, multiply the medication dose, the corresponding time correlation factor and the latency reasonableness weight, and accumulate all the products.

6. The method of claim 1, wherein the method is characterized by, The step of identifying potential drug safety signals comprises: Comparing the calculated time proportion reporting ratio with a preset threshold to screen out drug-adverse event pairs exceeding the threshold as potential safety signals.

7. The method of claim 1, wherein the method further comprises: The time interval effect in the time correlation factor is modeled by a function that decays exponentially with time interval, and the decay rate of the function is set based on the pharmacological half-life information of the drug. 8.A system for drug safety signal mining based on multi-source data, characterized in that, It comprises: A processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, a drug safety signal mining method based on multi-source data according to any one of claims 1-7 is realized.

Images