A signal management method for pharmacovigilance systems

Abstract

This invention belongs to the field of data processing technology, specifically relating to a signal management method for pharmacovigilance systems. The method includes: extracting the post text content, post timestamp, user account days, and user historical posting records from public posts; calculating an objective score; calculating a content value index; calculating domain focus based on the proportion of details in historical posting records; obtaining a user reputation index; acquiring the average user reputation index of all drug-event data pairs for the target drug, and calculating the time dispersion of the post timestamps of all public posts for the target drug, combining and weighting the two to obtain the signal strength of the target drug; obtaining a risk score for the target drug, and performing signal management and early warning. This invention solves the problem that traditional pharmacovigilance systems cannot effectively distinguish the authenticity and quality of internet information, and struggle to extract deep risk signals from unstructured text.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a signal management method for a pharmacovigilance system. Background Technology

[0002] In post-marketing drug safety supervision, pharmacovigilance is a key link in ensuring public drug safety and identifying and assessing the risk of adverse drug reactions. One of the core tasks of pharmacovigilance is to collect potential adverse event signals through various channels and analyze and judge them in order to discover unknown risks not mentioned in the drug instructions. Its main technical means include spontaneous reporting systems, literature retrieval and epidemiological studies.

[0003] Existing methods largely focus on statistical analysis of structured or semi-structured data, such as disproportionate analysis of reports from spontaneously reported systems. The core limitation of these methods lies in their heavy reliance on formatted reports voluntarily completed by healthcare professionals or patients, which suffer from severe underreporting and often lag behind the actual occurrence of adverse events. More importantly, with the widespread use of the internet, massive amounts of patient-generated medication experience texts, such as social media posts and patient forum threads, have become valuable sources of information for discovering unknown adverse reactions. However, traditional pharmacovigilance methods struggle to effectively utilize this type of unstructured data.

[0004] This inability to effectively utilize massive amounts of unstructured online information directly leads to inefficient signal detection and the omission of potential risks. A simple frequency count actually reflects a complex mix of information quality, publisher credibility, and event occurrence patterns. For example, a post concentrated in a short period by a competitor's hired online trolls describing dizziness caused by a certain drug, and a post shared by multiple long-term patients over several weeks detailing dizziness and specific symptoms after taking the same drug, may have the same number of dizziness mentions. However, from a pharmacovigilance perspective, the former is noise that needs to be filtered out, while the latter may be a genuine safety signal requiring further investigation. Traditional manual screening or simple methods based on keyword frequency, focusing only on the magnitude of mention counts, cannot effectively separate the authenticity of the signal from noise. This presents a double dilemma for signal management in pharmacovigilance systems:

[0005] First, a large number of emotional expressions with very little information, such as "this medicine is useless," are mixed with high-value posts containing detailed medication processes, resulting in low efficiency of signal analysis in the pharmacovigilance system.

[0006] Secondly, even if a post is rich in detail, it is difficult to determine its authenticity. For example, a carefully fabricated fake post published in a short period of time by a group of online trolls hired by a competitor may be indistinguishable from a post shared by multiple real patients over several weeks, describing real adverse reactions. This makes it difficult for existing technologies to effectively distinguish the quality of drug information, determine the authenticity of drug information sources, or identify the authenticity of group behavior patterns. As a result, pharmacovigilance experts are bogged down in reviewing massive amounts of low-quality information or miss early warning opportunities because they cannot identify weak but real signals. Summary of the Invention

[0007] To address the lack of comprehensive evaluation of drug information across three levels—content quality, user credibility, and group patterns—in order to effectively distinguish its authenticity and quality, this invention provides a signal management method for pharmacovigilance systems, comprising:

[0008] This process involves capturing publicly available posts related to the target drug from online data sources, extracting the post text content, post timestamps, user account days, and user historical posts; processing the post text content through multi-step natural language processing to obtain drug-event data pairs; obtaining the text length and number of details for individual publicly available posts, and calculating an objective score based on the objectivity of factual descriptive words in the post text content; calculating a content value index based on the interaction between the text length, number of details, and objective score of individual publicly available posts; obtaining the average content value for individual users, and calculating domain focus based on the proportion of details in historical posts; comprehensively evaluating user account days, domain focus, and average content value to obtain a user reputation index; obtaining the average user reputation index for all drug-event data pairs of the target drug, and calculating the time dispersion of the post timestamps of all publicly available posts for the target drug, combining and weighting these two factors to obtain the target drug signal strength; and weightedly fusing the content value index, objective score, and target drug signal strength of all publicly available posts for the target drug to obtain a target drug risk score for signal management and early warning.

[0009] Traditional pharmacovigilance methods are ineffective at distinguishing the authenticity and quality of online information, and struggle to extract deep-seated risk signals from unstructured text. This invention addresses this by employing a multi-step process to extract and process post and user information, removing interference, accurately acquiring key data, and differentiating between high- and low-value content. In user assessment, it comprehensively calculates user reputation across multiple dimensions, identifying reliable users and filtering out noise from low-reputation users. For signal analysis, it combines user reputation and time distribution to determine signal authenticity, then integrates multiple factors to calculate a risk score, accurately identifying genuine risk signals. Furthermore, through reasonable management and early warning systems, experts can efficiently obtain key information and quickly respond to risks, overcoming the inefficiency and risk-missing nature of traditional pharmacovigilance methods and improving the accuracy and effectiveness of pharmacovigilance.

[0010] Preferably, the post text content, post timestamp, user account days, and user's historical posting records of public posts are extracted, including:

[0011] The system associates each newly captured public post within the current monitoring period with its user identifier to the backend user database. After parsing the public posts and the user's public homepage information, it extracts the data and calculates and generates the post text content, post timestamp, user account days, and user's historical posting records.

[0012] Preferably, obtaining drug-event data pairs includes:

[0013] A data cleaning process is performed on the text content of a single public post to remove noise such as web page tags and advertising scripts. Natural language processing models are used for word segmentation and part-of-speech tagging to obtain tagged text. Named entity recognition technology is used to identify and extract the target drug name, adverse reaction symptoms or event description from the tagged text and combine them to generate drug-event data pairs.

[0014] Preferably, the objective score is calculated, including:

[0015] Using natural language processing technology, the number of medication information elements contained in a single public post is identified and counted, and their total number is recorded as the number of details; the total number of characters in the post text content is calculated and recorded as the text length; the ratio of factual descriptive words to subjective sentiment words in the post text content is calculated to obtain an objective score.

[0016] Preferably, the content value index satisfies the following expression:

[0017] ;

[0018] In the formula, V represents the content value index, which is a dimensionless value; L represents the text length, which is a unit; and D represents the number of details, which is an integer. Indicates objective scoring; Represents a logarithmic function.

[0019] This invention provides a method for calculating content value indicators, comprehensively considering text length, the amount of detail, and objective scoring. It appropriately manages text length to prevent excessively long texts from negatively impacting the score, while ensuring that details and objectivity contribute more reasonably to the value. By combining these three factors in the calculation, the value of post content can be comprehensively evaluated, accurately distinguishing between high-value and low-value posts. This allows for prioritizing high-value content in signal management, reducing interference from low-value information, and improving analytical efficiency and accuracy.

[0020] Preferably, computational domain focus includes:

[0021] The arithmetic mean of the content value metrics of all public posts in a user's historical posting record is recorded as the average content value; the proportion of posts with a detail count greater than 0 in the user's historical posting record is calculated to obtain the domain focus.

[0022] Preferably, the user credit metric satisfies the following expression:

[0023] ;

[0024] In the formula, This represents a user reputation metric, which is a dimensionless numerical value. This represents the number of days a user has an account; it is a dimensionless numerical value. Indicates domain focus, dimensionless; Indicates the average value of the content; Represents a logarithmic function; This represents the maximum value function.

[0025] The user reputation index calculation of this invention comprehensively considers user account days, domain focus, and average content value. It reasonably handles account days, taking into account user experience while avoiding excessive influence from overly high experience levels. By combining domain focus and average content value, it comprehensively measures user credibility, and the calculated reputation index accurately distinguishes users with different levels of credibility. This allows information from high-credibility users to receive greater attention in signal management, while reducing the influence of information from low-credibility users, effectively filtering out noise and improving the accuracy of signal analysis.

[0026] Preferably, the signal strength of the target drug satisfies the following expression:

[0027] ;

[0028] In the formula, It represents the signal strength of the target drug and is a dimensionless numerical value; The number of users of the target drug is a dimensionless integer. The user reputation index of the i-th user is a dimensionless numerical value. It represents the time dispersion of the target drug and is a dimensionless numerical value; Represents a logarithmic function; It represents a very small positive number, and guarantees that the denominator is not 0.

[0029] This invention calculates the signal strength of a target drug by combining the average user reputation score and the temporal dispersion of the target drug. The average user reputation score reflects the overall credibility of the signal source, while the temporal dispersion of the target drug reflects the time distribution of publicly posted content. Combining these two factors effectively distinguishes between genuine and fake signals. Signals posted dispersedly by high-reputation users are given higher strength recognition; low-quality signals posted in a concentrated manner are given lower strength.

[0030] Preferably, the risk score of the target drug satisfies the following expression:

[0031] ;

[0032] In the formula, Indicates the risk score of the target drug; This indicates the total number of public posts related to the target drug. This represents the content value index of the i-th public post among all public posts related to the target drug. This represents the arithmetic mean of the objective ratings of all publicly available posts about the target drug. It is the signal strength of the target drug; , , , All are dimensionless numerical values; Represents a logarithmic function.

[0033] The target drug risk score calculation of this invention integrates the sum of content value, the average objective score, and signal strength. It considers the sum of content value to reflect the range of signal influence, combines the average objective score to determine the subjective severity of the signal, and adds signal strength to ensure signal authenticity. This combination comprehensively assesses drug risk, accurately distinguishes signals at different risk levels, and makes high-risk signals more prominent, facilitating the priority handling of high-risk situations in signal management.

[0034] Preferably, signal management and early warning include:

[0035] All target drugs are sorted in descending order according to their target drug risk scores, and drugs with scores exceeding the average risk scores of all target drugs are highlighted for warning. When a pharmacovigilance expert selects any highlighted warning signal, the system automatically pops up a visualization interface containing multi-dimensional diagnostic information. The visualization interface includes: the risk score of the target drug, the signal strength of the target drug, the number of users corresponding to the target drug, the time dispersion of the target drug, and the text content or link of at least one representative original post with the highest content value index corresponding to the target drug.

[0036] The beneficial effects of this invention are as follows: This invention is of great significance for post-market drug safety supervision. It fully utilizes the massive amount of user medication experience information on the internet, transforming previously unusable unstructured data into valuable risk signal sources, thus broadening the information channels for pharmacovigilance. By scientifically screening authentic and high-value information, accurate assessment of drug risks can be achieved, enabling the earlier detection of potential drug safety issues and providing timely and reliable risk information for regulatory authorities and relevant agencies. This helps to take timely measures to control risks, reduce the impact of adverse events on public health, and better protect public medication safety. Simultaneously, it improves the efficiency of pharmacovigilance work, reduces labor costs, and promotes the development of pharmacovigilance work towards greater efficiency and precision, playing a positive role in improving and optimizing the entire drug safety supervision system. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a signal management method for a pharmacovigilance system according to the present invention. Detailed Implementation

[0038] This invention discloses a signal management method for a pharmacovigilance system, referring to... Figure 1 This includes steps S1-S4:

[0039] S1: Capture public posts related to the target drug from network data sources, extract the post text content, post timestamp, user account days, and user's historical posting records; process the post text content through multi-step natural language processing to obtain drug-event data pairs.

[0040] It should be noted that the signal management method of this invention is based on a vast amount of unstructured text about drug use spontaneously generated by users on the internet. This invention first needs to capture and standardize this raw information from the complex network environment to generate multiple key data sequences required for subsequent calculations. Taking a new antihypertensive drug as an example, this new antihypertensive drug is designated as the target drug. After the target drug is launched, many patients will share their experiences on online platforms such as hypertension forums and patient support groups. This information includes unstructured text such as "After taking the new antihypertensive drug for two weeks, my blood pressure has decreased, but I always have a dry cough at night, and I feel dizzy after taking this drug; I don't know if it's a side effect." This text is valuable information for discovering unknown adverse reactions, but it is diverse in form and contains a large amount of irrelevant information. Traditional manual screening methods are inefficient and prone to missing key signals. This invention uses technical means to automatically and standardize the capture of this raw information, establishing a clean and well-organized data foundation for subsequent intelligent analysis.

[0041] Specifically, publicly available posts related to the target drug are captured from online data sources, and the post text content, post timestamps, user account days, and user historical posting records are extracted, including:

[0042] The system associates each newly captured public post within the current monitoring period with its user identifier to the backend user database. After parsing the public post and the user's public homepage information, it extracts the data and calculates and generates the post text content, post timestamp, user account days, and user's historical posting records. The post text content is obtained by directly extracting the post's body. The post timestamp is obtained by parsing the public post's timestamp metadata. The user account days are obtained by calculating the time difference between the user's registration time and the current time, with the user registration time being used as the default post timestamp if the attempt to parse the user's public homepage fails. The user's historical posting records are obtained by retrieving and accessing the system's historical database using the user identifier, and are presented as a data set containing the text content of the user's previously posted posts.

[0043] At this point, we have obtained the post text content, post timestamp, user account days, and user's historical posts for the public posts.

[0044] It should be noted that since subsequent user reputation analysis relies on the user's historical behavior, the user's account days, historical posts, post text content, and post timestamp are bound together to obtain drug-event data pairs, thereby ensuring that each data unit contains complete information in three dimensions: content, time, and user background.

[0045] Preferably, the post text content is processed through multi-step natural language processing to obtain drug-event data pairs, including:

[0046] A data cleaning process is performed on the text content of individual public posts to remove noise such as webpage tags and advertising scripts. Natural language processing models are then used for word segmentation and part-of-speech tagging to obtain tagged text. Named entity recognition technology is employed to identify and extract the target drug name, adverse reaction symptoms, or event descriptions from the tagged text and combine them to generate drug-event data pairs. For example, data pair 1: Drug: Antihypertensive drug X, Event: Dry cough, Details: One 50mg tablet daily, at bedtime; Data pair 2: Drug: Antihypertensive drug X, Event: Dizziness, Details: One 50mg tablet daily, during the day.

[0047] Thus, the corresponding drug-event data pair was obtained from a single public post.

[0048] S2: Obtain the text length and number of details of a single public post, and calculate an objective score based on the objectivity of factual descriptive words in the post text content; calculate the content value index based on the interaction between the text length, number of details, and objective score of a single public post; obtain the average content value of a single user, and calculate the domain focus based on the proportion of details in the historical post records; comprehensively evaluate the user's account days, domain focus, and average content value to obtain the user reputation index.

[0049] It's important to note that the core innovation of this invention lies in avoiding the simplistic and fragile signal aggregation logic of traditional pharmacovigilance, which relies solely on frequency of mentions and emotional intensity. When using internet data for pharmacovigilance, the main problem is the extremely low signal-to-noise ratio and a large amount of online noise. For example, posts with strong emotions but zero information, such as "It didn't work, don't buy it!", and false information from competitors like "My friend experienced severe diarrhea after taking this medication," can lead to frequent false alarms, overwhelming pharmacists. Meanwhile, truly valuable signals, such as "After taking two pills daily for a month, I experienced palpitations and insomnia," may be lost due to the limited number of publishers. To address this issue, this invention moves beyond simple counting. Instead, it analyzes the intrinsic clinical value of individual public posts. In real-world social media scenarios, a brief complaint, such as "This medication makes me feel worse," has extremely low clinical value, while a detailed description, such as "I took the medication as prescribed, starting at 20mg / day. Starting in the third week, I experienced persistent nausea and loss of appetite," contains crucial information necessary for assessing causality, including the medication, dosage, duration of use, and specific symptoms. Therefore, this invention constructs a content value index to enable the system to automatically identify and analyze this difference in information content, assigning a significantly higher initial weight to the latter, thus prioritizing high-quality information from the very beginning of the analysis.

[0050] Specifically, the text length and number of details of a single public post are obtained, and an objective score is calculated based on the objectivity of factual descriptive words in the post text, including:

[0051] Using natural language processing technology, the number of medication information elements contained in a single public post is identified and counted, and their total number is recorded as the number of details; the total number of characters in the post text content is calculated and recorded as the text length; the ratio of factual descriptive words to subjective sentiment words in the post text content is calculated to obtain an objective score.

[0052] It should be noted that, in order to solve the aforementioned technical problem of mixed quality and noise, this invention first evaluates the content value of a single public post. The number of details in this invention refers to the total number of specific information elements related to medication identified in a public post through keyword or regular expression matching. These information elements include not only adverse reaction events, such as "dosage: 50mg, duration of use: three weeks, frequency of use: once daily, specific physiological indicator changes: blood pressure dropped from 150 to 120", but also other information. The purpose of the number of details is to measure the specificity and information richness of a post's content.

[0053] Preferably, the content value index is calculated based on the interaction between the text length, number of details, and objective rating of a single public post, including:

[0054] Content value metrics satisfy the following expression:

[0055] ;

[0056] In the formula, V represents the content value index, which is a dimensionless value; L represents the text length, which is a unit; and D represents the number of details, which is an integer. Indicates objective scoring; Represents a logarithmic function.

[0057] In the formula, This indicates that a logarithmic transformation is applied to the text length, causing the contribution of text length to content value to exhibit diminishing marginal utility. and These represent linear enhancements to the number of details and the objective score, respectively, ensuring that the metrics do not drop directly to zero even when the scores for detail or objectivity are low. Using a parameterless product form, it is shown that text length, number of details, and objective score have a synergistic effect on content value. That is, a high-value data must perform well in all dimensions, and the absence of any one dimension will significantly lower the total score.

[0058] For example, in one scenario, a low-value public post states, "This medicine doesn't work, I feel terrible," with a text length L=7, no detailed information element D=0, and a low objective score O=0.1, resulting in a content value V=3.39. In another scenario, a high-value post details, "After taking 50mg of the medicine for three weeks, I experienced dizziness and nausea every night, with the symptoms disappearing during the day," with a text length L=35, containing detailed information element D=4, and an objective score of... If the content value is 0.9, then its content value V = 43.52. It is evident that this invention can effectively distinguish texts with different information contents and assign a significantly higher initial value to high-quality content. All calculation results are rounded to two decimal places.

[0059] It should be noted that the factual descriptive words and subjective sentiment words in the post text were obtained by searching and matching from publicly available sentiment dictionaries and factual keyword dictionaries. That is, sentiment dictionaries contain a large number of words expressing personal feelings, opinions and emotions, such as "uncomfortable", "painful", "useless", "recommended", etc. Sentiment dictionaries are a mature and publicly available resource in the NLP field. Factual keyword dictionaries contain words describing objective facts, data, behaviors and states, especially words commonly used in the medical field, such as dosage units "mg", time units "week" and "day", objective symptoms "dizziness", "nausea", "rash", etc. Factual keyword dictionaries are usually not publicly available or universal, but domain-specific. In this invention, it is a dictionary specifically built for the medical field, which is a standard practice in NLP applications.

[0060] It should be noted that when calculating the objective score in the example, the text of a single public post is first segmented with reference to a sentiment dictionary and a factual keyword dictionary, and then the number of words in the text that match the sentiment dictionary is counted. And the number of words in the factual keyword dictionary that hit the target words. Through formula Calculate an objective score; for the example post "This medicine is useless, I feel unwell", a Chinese word segmentation tool based on HMM is used to segment the post into independent words, namely "this, medicine, useless, , unwell", and string matching is performed. Among them, "useless, unwell" matches the sentiment dictionary, therefore... No factual keywords. Substitute into the formula =0.1; For the example post "After taking 50mg of medication for three weeks, I experienced dizziness and nausea every night, and the symptoms disappeared during the day," a Chinese word segmentation tool based on HMM was used to segment the post into independent words, namely "medication, 50mg, three weeks, after, every night, appeared, dizziness, and, nausea, daytime, symptoms, disappeared," and a string exact match was performed. Among them, "50mg, three weeks, every night, dizziness, nausea" matched the factual keyword dictionary. No subjective emotional words. Substitute into the formula =0.9. middle, , As a smoothing factor, regarding , The design is based on a refined model design, which follows the well-known additive smoothing principle in natural language processing. and These represent partial and total pseudo-counts, respectively. In online text, high-value factual content is relatively scarce, while sentimental content is very common. This invention sets different prior pseudo-counts for these two categories, assigning a smaller prior pseudo-count to the scarce and important factual category. For the prevalent emotional categories, a relatively large prior pseudo-count is set. Then the total pseudo count It is precisely the sum of these two asymmetric priors, namely .exist In this context, if after segmenting the text of a single public post, the number of words in both the sentiment dictionary and the factual keyword dictionary that match are zero, then... When the value is 0, the default objective rating for this post is set. .

[0061] It should be noted that after initially screening the quality of individual public posts using content value indicators, to further address the issue of determining the authenticity of information sources, this invention requires analyzing the credibility of the user who posted the information. In pharmacovigilance practice, a post with high content value, if from a newly registered account that has never posted health-related content before, should have a lower weight as a potential signal than the same post from a long-term active user in a health community with a generally high-quality history of posts. For example, when monitoring the safety of a children's cold medicine, if the system captures a post with high content value detailing "a 3-year-old baby developed a rash after drinking cold syrup," and if the user A who posted the post has an account registered for 5 years and has posted hundreds of posts on parenting and common childhood illness care, with generally objective and detailed content, then the authenticity of this information is very high. Conversely, if the user B who posted the same content has an account registered for only 1 day and this is their only post, then it is very likely a malicious troll account. In order to enable the system to distinguish the credibility of information sources, this invention analyzes the user's qualifications and expertise, assigns higher credibility weight to user A's posts, and suppresses the influence of user B's posts, thereby achieving the purpose of filtering network noise and improving signal accuracy.

[0062] Preferably, the average content value of a single user is obtained, and domain focus is calculated based on the proportion of details in historical posts, including:

[0063] The arithmetic mean of the content value metrics of all public posts in a user's historical posting record is recorded as the average content value; the proportion of posts with a detail count greater than 0 in the user's historical posting record is calculated to obtain the domain focus.

[0064] Preferably, a comprehensive evaluation is conducted based on user account duration, domain focus, and average content value to obtain user reputation metrics, including:

[0065] The user reputation metric satisfies the following expression:

[0066] ;

[0067] In the formula, This represents a user reputation metric, which is a dimensionless numerical value. This represents the number of days a user has an account; it is a dimensionless numerical value. Indicates domain focus, dimensionless; Indicates the average value of the content; Represents a logarithmic function; This represents the maximum value function.

[0068] In the formula, This indicates a logarithmic transformation of the number of days a user's account has been active, essentially performing a non-linear analysis of the user's credit history. It suggests that longer account registration times generally correlate with a stronger credit history, but this growth effect weakens over time. The +1 in the formula ensures that even newly registered users with A=0 can receive the base score; As a user's domain focus, it directly measures the degree of focus that user has on the relevant field; This represents the user's historical performance, that is, the average value of the content that the user has published in the past. This indicates that by multiplying these three dimensions, a reputation assessment model is constructed. It means that a high-reputation user must simultaneously possess the characteristics of deep experience, high focus, and high average content quality. A deficiency in any dimension, such as a user with deep experience but extremely low quality of past posts, will significantly lower the final user reputation index, thereby effectively identifying and suppressing the influence of low-quality or spam accounts.

[0069] For example, consider a newly registered account suspected of being a paid troll account. Its account duration is A=2 days, domain focus is F=0.1, and the average value of its historical content is low. If the user's reputation index U=0.55, then there is another scenario: a senior patient user with an account duration of A=1825 days (5 years), domain focus F=0.9, and high-quality historical posts. If the user reputation index U=202.77, it can be seen that the user reputation index can greatly amplify the influence of senior and trustworthy users, while suppressing the weight of low-reputation accounts. All the above calculation results are rounded to two decimal places.

[0070] S3: Obtain the average user reputation index of all drug-event data pairs for the target drug, and calculate the time dispersion of the timestamps of all public posts for the target drug. Combine and weight the two to obtain the signal strength of the target drug.

[0071] It should be noted that after evaluating the quality of individual content and the authenticity of individual users, to address the problem of identifying the authenticity of group behavior patterns—that is, distinguishing between genuine group phenomena and fake online attacks—this invention needs to construct a target drug signal strength that can integrate the strength of group evidence and temporal evolution characteristics. A genuine drug safety issue is usually not an isolated event, but rather manifests as a group phenomenon reported by multiple independent patients with different backgrounds within a certain time window. To capture the emergence pattern of such signals and distinguish their authenticity, this invention needs to construct a target drug signal strength that can integrate the strength of group evidence and temporal evolution characteristics. For example, regarding a certain hypoglycemic drug Y, if the system finds that 10 different high-reputation users have posted within a month, all reporting "ankle edema after taking the drug," although these posts are scattered across different times and forums, they all point to a potential, group-wide adverse reaction signal. This contrasts sharply with another scenario: if 50 newly registered accounts suddenly appear within two hours on a certain afternoon, posting similar content such as "hypoglycemic drug Y causes edema," this is highly likely a commercial attack or organized online attack. This invention distinguishes between two distinct situations by constructing a target drug signal strength and analyzing the average user reputation and time distribution pattern (i.e., time dispersion) of the target drug signal. It assigns a very high target drug signal strength to the naturally fermenting group signal, thereby effectively suppressing short-term bursts of low-reputation group behavior.

[0072] Specifically, the mean user reputation index of all drug-event data pairs for the target drug is obtained, and the time dispersion of the posting timestamps of all public posts for the target drug is calculated. These two values ​​are then combined and weighted to obtain the signal strength of the target drug, including:

[0073] The signal strength of the target drug satisfies the following expression:

[0074] ;

[0075] In the formula, It represents the signal strength of the target drug and is a dimensionless numerical value; The number of users of the target drug is a dimensionless integer. The user reputation index of the i-th user is a dimensionless numerical value. It represents the time dispersion of the target drug and is a dimensionless numerical value; Represents a logarithmic function; It represents a very small positive number, and guarantees that the denominator is not 0.

[0076] In the formula, It calculates the average of the reputation metrics of all independent users of the target drug, which represents the overall credibility basis of the signal source group; This represents the temporal distribution pattern of the signal, where the standard deviation of the publication timestamps is... The larger the value, the more dispersed the posting times are, and the more consistent they are with the natural occurrence of real adverse reaction events. The smaller the size, the more likely it is to be a short-term, concentrated attack, which is consistent with the characteristics of a naval attack. Turning the time distribution pattern into a moderating factor: a group of high-reputation users whose posting times are naturally dispersed will have a stronger signal; conversely, even a seemingly large group will have a stronger signal if their posting behavior is highly concentrated in time. It will also approach 0, thereby effectively suppressing the intensity of the target drug signal and achieving precise filtering of false group signals.

[0077] For example, in scenario one, a suspected online troll attack targets a drug-related signal, involving N=20 users with a low average reputation score of 5. The posts explode within half a day, exhibiting a high degree of temporal dispersion. If the target drug signal strength is C=0.91, then there is a second scenario: a naturally developing target drug signal, containing N=10 independent users with a relatively high average reputation score of 80, with posts appearing gradually over a month, and a time dispersion. If the target drug signal strength is C=221.81, then this indicator successfully suppressed short-term bursts of low-quality signals and significantly amplified persistent, genuine signals composed of high-reputation groups. The calculation results of C=0.91 and C=221.81 are both rounded to two decimal places.

[0078] It should be noted that in the calculation When determining the temporal dispersion of a target drug, a time window is first preset. The length of the time window is from the detection of the first public post about the target drug until the acquisition of the current target drug signal. The timestamps of all public posts within the time window are obtained and converted into uniform, continuous values. Then, the standard deviation of all timestamps within the time window is calculated and denoted as . The unit is days; a constant reference unit of time is introduced. and order Heavens, then If the standard deviation of all timestamps for the target drug is 0.2, then , If the standard deviation of all timestamps for the target drug is 15, then , .

[0079] S4: Weight and integrate the value indicators, objective scores, and signal strength of all public posts related to the target drug to obtain a risk score for the target drug, and then manage and warn of the signals.

[0080] It should be noted that although this invention has obtained the intensity of target drug signals that characterize the authenticity of the signals, in actual pharmacovigilance work, the overall number of mentions of the signal and the intensity of negative emotions are still important reference dimensions for assessing the breadth and urgency of its public health impact. To generate a target drug risk score that can assist vigilance specialists in risk ranking and decision-making, the high-level features extracted in this invention must be effectively integrated with traditional statistical measures. Continuing with the example of the signal of "ankle edema" caused by hypoglycemic drugs, assuming that the preceding steps have confirmed that this is a genuine signal with high intensity, naturally reported by a high-reputation user group over a period of time, vigilance specialists still need to consider the severity of the problem at the decision-making level. If this target drug signal has only 10 posts in a month, and the content is mostly "mild edema, not affecting daily life," i.e., the intensity of negative emotions is weak, then its urgency may not be high. However, if the target drug signal includes 200 posts in a month, and the content is mostly "severe edema, unable to wear shoes, very painful," then this constitutes a major public health risk requiring immediate intervention. This invention effectively combines the authenticity and severity of signals by constructing a target drug risk score, generating a comprehensive target drug risk score that ensures that signals that are both authentic and critically urgent can stand out from all signals to be evaluated.

[0081] Specifically, the risk score for the target drug is obtained by weighting and fusing the content value indicators, objective scores, and signal strength of all publicly available posts about the target drug. This includes:

[0082] The target drug risk score satisfies the following expression:

[0083] ;

[0084] In the formula, Indicates the risk score of the target drug; This indicates the total number of public posts related to the target drug. This represents the content value index of the i-th public post among all public posts related to the target drug. This represents the arithmetic mean of the objective ratings of all publicly available posts about the target drug. It is the signal strength of the target drug; , , , All are dimensionless numerical values; Represents a logarithmic function.

[0085] In the formula, This indicates the subjective severity of all public posts about the target drug. When the arithmetic mean of the objective ratings of all public posts about the target drug is low, that is, when there are more subjective and emotional words, the value of this item will increase, thereby amplifying the risk level of the target drug signal. The breadth of influence of the target drug signal is indicated by the sum of the content value of all public posts about the target drug, rather than the simple number of posts. This allows high-quality posts to receive higher weight. The highest target drug risk scores are those that are authentic, contain a wealth of high-quality details, and strongly express group sentiment.

[0086] For example, if a target drug signal has a low risk score, the total content value of all its public posts is... The arithmetic mean of the objective ratings for all publicly available posts about the target drug is Leaning towards objectivity Then its target drug risk score If a target drug has a high risk score, the total content value of all its publicly posted content is considered a target drug signal. The arithmetic mean of the objective ratings for all publicly available posts about the target drug is Subjective and emotional. Then its target drug risk score As can be seen, this score can effectively distinguish between true signals of varying severity. All calculation results are rounded to two decimal places.

[0087] It is important to note that the purpose of this invention is not simply to output a target drug risk score, but rather to present this risk score to pharmacovigilance experts in an intuitive, traceable, and interactive manner, thereby achieving efficient signal management. In other words, an isolated target drug risk score has limited meaning. When a pharmacovigilance expert sees a high-scoring target drug signal, their primary need is to quickly understand the composition of the score and download the original data or report for verification. Therefore, this invention requires an efficient signal presentation and management mechanism that transforms the calculation results into decision support information, allowing experts to quickly penetrate from a macro-level risk ranking to the signal strength, group characteristics, time patterns, and ultimately the most representative individual original reports.

[0088] Preferably, signal management and early warning include:

[0089] All target drugs are sorted in descending order according to their target drug risk scores, and drugs with scores exceeding the average risk scores of all target drugs are highlighted for warning. When a pharmacovigilance expert selects any highlighted warning signal, the system automatically pops up a visualization interface containing multi-dimensional diagnostic information. The visualization interface includes: the risk score of the target drug, the signal strength of the target drug, the number of users corresponding to the target drug, the time dispersion of the target drug, and the text content or link of at least one representative original post with the highest content value index corresponding to the target drug.

[0090] It should be noted that the signal management system of the present invention can periodically perform risk score calculation methods on multiple target drugs in the preset list, and sort and compare the calculated risk scores of each target drug to achieve global early warning.

[0091] This completes the signal management of the pharmacovigilance system.

[0092] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A signal management method for a pharmacovigilance system, characterized by, The method comprises the following steps: Capture public posts related to the target drug from network data sources, extract post text content, post timestamp, user account days, and user historical speech records; process the post text content through multiple natural language processing steps to obtain drug-event data pairs; The text length and the number of details of a single public post are obtained, and an objective score is calculated in combination with the objective degree of factual description words in the post text content, including: through natural language processing technology, the medication information elements contained in a single public post are identified and counted, and the total number is recorded as the number of details; the total number of characters in the post text content is calculated, recorded as the text length; the ratio of factual description words to subjective emotional words in the post text content is calculated to obtain the objective score; the content value index is calculated according to the interaction relationship of the text length, the number of details and the objective score of a single public post, including: ; In the formula, V represents the content value index, which is a dimensionless value; L represents the text length, and the dimension is individual; D represents the number of details, which is an integer; represents the objective score; represents the logarithmic function; the content average value of a single user is obtained, and the field concentration is calculated according to the proportion of the number of details in the historical speech records, including: the arithmetic mean of the content value indexes of all public posts in the historical speech records of a single user is recorded as the content average value; the proportion of the posts with the number of details greater than 0 in the historical speech records of the user is calculated to obtain the field concentration; the user account days, the field concentration and the content average value are comprehensively evaluated to obtain the user reputation index, including: ; In the formula, represents the user reputation index, which is a dimensionless value; represents the user account days, which is a dimensionless value; represents the field concentration, which is dimensionless; represents the content average value; represents the logarithmic function; represents the maximum function; Obtain the average of the user credibility indicators of all drug-event data pairs of the target drug, and calculate the time dispersion of the post timestamps of all public posts of the target drug, and combine and weight the two to obtain the target drug signal strength; Weight and fuse the content value indicators, objective scores, and target drug signal strength of all public posts of the target drug to obtain the target drug risk score, and perform signal management and early warning.

2. The signal management method for a pharmacovigilance system according to claim 1, wherein, The extraction of post text content, post timestamp, user account days, and user historical speech records comprises: The system associates each newly captured public post in the current monitoring period with the background user database through its user identifier, and through the analysis of public post and user public homepage information, the data extraction is completed, and the post text content, post timestamp, user account days, and user historical speech records are calculated and generated.

3. The signal management method for a pharmacovigilance system according to claim 1, wherein, The drug-event data pair is obtained by: Performing data cleaning process on the post text content of a single public post to remove web tags and advertising script noise, and using natural language processing model for word segmentation and part-of-speech tagging to obtain tagged text; using named entity recognition technology to identify and extract target drug name, adverse reaction symptoms or event description from the tagged text and combine them to generate drug-event data pairs.

4. The signal management method for a pharmacovigilance system according to claim 1, wherein, The target drug signal strength satisfies the following expression: ； In the formula, represents the target drug signal intensity, which is a dimensionless value; represents the number of users of the target drug, which is a dimensionless integer; represents the user reputation index of the i th user, which is a dimensionless value; represents the time dispersion degree of the target drug, which is a dimensionless value; represents a logarithmic function; represents a very small positive number, which ensures that the denominator is not 0.

5. The signal management method for a pharmacovigilance system according to claim 1, wherein, The target drug risk score satisfies the following expression: ； In the formula, represents the target drug risk score; represents the number of all public posts of the target drug; represents the content value index of the i-th public post among all public posts of the target drug; represents the arithmetic mean of the objective scores of all public posts of the target drug; is the target drug signal strength; , , , are all dimensionless values; represents the logarithmic function.

6. The signal management method for a pharmacovigilance system according to claim 1, wherein, The signal management and early warning comprises: All target drugs are arranged in descending order according to their target drug risk scores, and the drug categories with scores exceeding the average of all target drug risk scores are highlighted for early warning; when a drug safety expert selects any highlighted early warning signal, the system automatically pops up a visual interface containing multi-dimensional diagnostic information, which includes the risk score of the target drug, the target drug signal strength, the number of users corresponding to the target drug, the time dispersion corresponding to the target drug, and the text content or link of at least one representative original post with the highest content value indicator corresponding to the target drug.

Images