Adverse event retrieval method, system and device based on multi-factor fusion and medium

Abstract

The application provides an adverse event retrieval method, system, device and medium based on multi-factor fusion, which comprises the following steps: calculating the text similarity between a query event and each candidate event; calculating the category matching score, department matching score, grade proximity score and time decay score of the query event and the candidate event according to the event field representation; calculating the decoupled text vector by combining the structured label matrix and updating the text similarity to the decoupled text similarity; weighting and fusing the category matching score, department matching score, grade proximity score, time decay score and decoupled text similarity by using the factor adaptive weight to generate a comprehensive score; sorting each candidate event in the adverse event case set according to the comprehensive score, selecting the candidate events with high ranking to generate a similar event retrieval recommendation list of the query event, and generating retrieval recommendation information based on the comprehensive score of each candidate event. The application can improve the accuracy and scene adaptability of adverse event analysis.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to adverse event retrieval methods, systems, devices and media based on multi-factor fusion. Background Technology

[0002] Adverse medical events refer to events that occur and are discovered within a medical institution, excluding the natural progression of a patient's illness, and are caused by various human, equipment, process, and management factors that pose safety hazards or have resulted in negative consequences such as personal injury, medical disputes, or operational losses. These events encompass complete data including event type, department where they occurred, event details, root cause analysis, risk level, and handling and rectification measures. They are a core control target of the hospital's quality and safety management system. With the continuous improvement of adverse medical event reporting mechanisms, hospitals have accumulated massive amounts of historical event data. This data effectively supports the assessment of new events, the reuse of handling experience, the rectification of process loopholes, and risk prediction, serving as a core data resource for hospitals to continuously optimize their quality control systems.

[0003] Currently, adverse event similarity case retrieval and auxiliary analysis technologies are mainly divided into six categories, including keyword retrieval, full-text retrieval, single text similarity matching, rule-based filtering, manual experience comparison, and adaptive weight retrieval based on ranking learning. The first five categories rely on fixed rules and surface text matching logic, resulting in limited adaptability and intelligence. Adaptive weight retrieval based on ranking learning, which can adaptively adjust feature weights through machine learning, is currently the mainstream technology. While existing technologies can meet basic query needs, they have certain shortcomings in adapting to medical safety management scenarios. Traditional retrieval methods have weak domain semantic adaptability, failing to accurately match medical terminology and quality control business rules, easily leading to text similarity but mismatched business requirements. Furthermore, existing technologies struggle to achieve reasonable fusion of multi-dimensional features, generally exhibiting poor adaptability of fixed weights, redundancy and distortion in multi-feature fusion, and a lack of optimization based on event timeliness, failing to meet the differentiated similarity matching needs of different types of medical events. In addition, mainstream machine learning retrieval models suffer from cold-start challenges and domain adaptability shortcomings, resulting in poor model interpretability and retrieval results that are difficult to support manual review and management acceptance. The overall technical system lacks sufficient data mining depth, lacks an automated iterative optimization mechanism, cannot form a closed loop for experience reuse, and has weak retrieval and recall capabilities for small-sample, long-tail rare adverse events, making it difficult to fully utilize the value of historical event data and failing to meet the hospital's needs for refined and intelligent medical safety quality control improvement. Summary of the Invention

[0004] The present invention aims to provide a method, system, device and medium for adverse event retrieval based on multi-factor fusion, so as to solve the above-mentioned technical problems and improve the accuracy and scenario adaptability of adverse event analysis.

[0005] To address the aforementioned technical problems, this invention provides a method for retrieving adverse events based on multi-factor fusion, comprising: Calculate the text similarity between the query event and each candidate event in the adverse event case set; wherein, the adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level, and event department; The event field representation is used to calculate the category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate events; where the event field representation is constructed based on field information and event time. The decoupled text vectors of each candidate event in the query event and adverse event case set are calculated by combining the structured label matrix. The text similarity is then updated to the decoupled text similarity based on the decoupled text vectors. The structured label matrix is ​​generated based on the coding results of event category, event department, and event level. The system employs pre-defined adaptive weights to weight and fuse category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event. The types of factors include category factors, department factors, grade factors, time factors, and text factors. The candidate events in the adverse event case set are sorted according to the comprehensive score. A preset number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list for the query event. Retrieval recommendation information is generated based on the comprehensive score of each candidate event.

[0006] The above solution abandons the single text matching mode and fully explores the multi-dimensional structured core fields of adverse event cases, such as category, department, safety level, and occurrence time. It combines these with unstructured event text features to achieve multi-factor collaborative retrieval, comprehensively covering the attribute and content features of adverse events. By constructing a structured tag matrix to generate decoupled text vectors, it optimizes and updates text similarity, effectively avoiding the matching bias problems caused by semantic redundancy and feature coupling in traditional text retrieval, making the text matching results more closely aligned with the professional scenario characteristics of medical adverse events. Simultaneously, it adds four structured scoring dimensions: category, department, level, and time, combined with time decay scores to suit the time-sensitive application characteristics of adverse events, balancing the matching of inherent event attributes with dynamic adaptation of the time dimension. Furthermore, it employs factor adaptive weighting to complete the weighted fusion of multi-dimensional scores, which can autonomously adapt to the retrieval needs of different query events, avoiding the retrieval rigidity caused by fixed weights and ensuring the scientific and reasonable nature of the comprehensive score. By generating a search recommendation list based on accurate comprehensive scoring and ranking, it not only achieves efficient and accurate screening of similar adverse events, but also ensures the hierarchy and reference value of search results. It can provide accurate and effective case support for medical safety review, risk prevention and control, and event tracing, significantly improving the intelligence level and clinical application value of medical adverse event retrieval.

[0007] In one implementation, the adverse event case set contains several structured adverse event samples. Each adverse event sample retains at least the following fields: event category, event process, event analysis, adverse event level for medical safety, and the department involved. Specifically, these include: Adverse event samples are collected through preset dimensions; these dimensions include event time, event category, and event department. For each adverse event sample, retain at least the following fields: event category, event details, event analysis, adverse event level for medical safety, and the department involved. One-hot encoding is performed on the field information of event category, event department, and event level to generate one-hot label vectors for each field, resulting in a structured label matrix; Based on a pre-trained entity recognition model in the medical field, several core entities are extracted from the field information of event process and event analysis and mapped to a unified standard code to obtain entity normalization results and generate an entity code set. The event process and the field information of the event analysis are concatenated to generate the text representation basis of the adverse event, and the event text sequence is generated by combining the entity normalization results. The event time is extracted from the field information, and the event field representation is constructed based on the three-level classification results of the event time and event category, the event process, the event analysis, the event level, the event department, and the entity code list; among them, the entity code list is determined according to the extracted core entities and the mapped standard codes; The event labeling results of adverse event samples are generated based on the comparison between the total number of adverse event samples under each event subclass and the preset subclass sample threshold; among which, the event labeling results include long-tail events and rare events; Structured bad event samples are generated based on structured label matrices, entity encoding sets, event text sequences, event field representations, and event tagging results.

[0008] The above scheme defines standardized data collection dimensions and five core mandatory fields to establish a complete and unified data foundation. It generates a structured label matrix through one-hot encoding, quantifying classification fields and providing standard input for feature decoupling. It normalizes terminology based on medical entity recognition, eliminating semantic bias in synonymous texts while isolating structured information and generating clean event text. Furthermore, it divides long-tail tags according to sample size, providing a basis for optimizing rare event retrieval. Finally, it integrates all standardized products to generate unified structured samples, solving underlying problems such as messy raw data, inconsistent terminology, and feature duplication, while also distinguishing between regular and small-sample events, effectively improving the matching accuracy of similar cases and the recall effect of long-tail events.

[0009] In one implementation, the text similarity between the query event and each candidate event in the adverse event case set is calculated, specifically including: The field information of the event process and event analysis is preprocessed, and the preprocessed text is input into a pre-trained medical domain model for encoding to generate the first global semantic vector of the query event and the second global semantic vector of each candidate event. The entity overlap degree between the first global semantic vector and each of the second global semantic vectors is calculated based on the entity encoding set of adverse events; whereby the expression for the entity overlap degree is: ; In the formula, To query events; For the first One candidate event; For the degree of overlap between entities; The set of entity codes for the query event; For the first A set of entity codes for candidate events; The text similarity between the query event and each candidate event is calculated by combining entity overlap; the expression for text similarity is: ; In the formula, For text similarity, Custom weighting coefficients; Cosine similarity of global semantic vectors; This is the first global semantic vector for the query event; For the first The second global semantic vector of the candidate events.

[0010] In the above solution, professional semantic vectors are generated based on a medical pre-trained model to adapt to the context of medical terminology and avoid semantic understanding biases of general models. The entity matching degree is calculated by using the Jaccard coefficient to accurately capture core business elements such as diagnosis and symptoms, making up for the lack of details in the global vector. The weighted fusion logic is used to balance the weight of the full-text context and entity matching. The λ value can be adjusted for long-tail events to increase the proportion of entity matching. It takes into account the overall process matching of regular events and the core element retrieval needs of rare events. It can effectively solve the problem of traditional keyword matching where the words are similar but the diagnosis and treatment logic is inconsistent, and obtain the original text similarity that fits the medical business.

[0011] In one implementation, the category matching score, department matching score, grade proximity score, and time decay score of the query event and candidate events are calculated based on the event field representation. Specifically, this includes: Obtain the three-level classification results of the event categories, and calculate the category matching score based on the three-level classification results; the matching rule expression for the category matching score is as follows: ; In the formula, For category matching functions; Candidate events; For the first One candidate event; , These are event subclasses for query events and candidate events, respectively. This is a category main class mapping function used to map event subclasses to their respective intermediate level main classes; This is a category mapping function used to map the main class of an event to its corresponding top-level category; The department matching score is calculated based on the event department in the event field and the pre-defined clinical department matching rules; the expression for the clinical department matching rule is as follows: ; In the formula, For department matching functions; , These are the specific departments to which the query event and candidate events belong, respectively. For departmental first-level classification mapping function; The level proximity score is calculated based on the event level in the event field and a pre-defined level matching rule; the expression for the level matching rule is as follows: ; In the formula, It is a proximity function of the order of precedence; , These are the event levels for the query event and the candidate event, respectively. This is a level mapping function that maps event levels I-IV to 4, 3, 2, and 1 respectively, from highest to lowest severity. This is the attenuation coefficient for grade differences, with a default value of 0.35; The time decay score is calculated based on the event time indicated in the event field and a pre-defined time decay function; the expression for the time decay function is as follows: ; In the formula, It is a time decay function; The time interval between the query event and the candidate events; This is the time half-life parameter, with a default value of 12 months. This is the maximum effective time window, with a default value of 60 months.

[0012] In the above scheme, four types of hospitals' core quality control business dimensions are quantitatively matched with scores. A tiered, category-based scoring system balances precise retrieval with long-tail case recall across subcategories. Departmental grading matches the differences in specialty diagnosis and treatment. A linear decay mechanism differentiates the severity of risks, avoiding confusion between minor and major events. A time half-life decay model matches the hospital's annual policy update patterns, automatically eliminating expired cases. All four score categories align with the actual constraints of medical safety management, compensating for the shortcomings of pure text retrieval that ignores business scenarios, risk levels, and timeliness, thus enhancing the quality control reference value of search results.

[0013] In one implementation, the decoupled text vectors of each candidate event in the query event and the adverse event case set are calculated by combining the structured label matrix. The text similarity is then updated to decoupled text similarity based on these decoupled text vectors. Specifically, this includes: The event text sequences of each candidate event in the adverse event case set are concatenated to construct an event text vector matrix; The event text vector matrix and the structured tag matrix are standardized to obtain a standard event text vector matrix and a standard structured tag matrix; the expression for the standardization process is as follows: ; ; In the formula, In the standard event text vector matrix, the value of the i-th adverse event sample in the j-th text dimension is given. Let X be the original feature value of the i-th adverse event sample in the j-th dimension. Let be the sample mean of the j-th dimension in the event text vector matrix X; Let be the sample standard deviation of the j-th dimension in the event text vector matrix X; In a standard structured label matrix, the value of the i-th adverse event sample in the k-th label dimension is given. In the structured label matrix, the original value of the i-th adverse event sample in the k-th label dimension is given. This represents the sample mean of the k-th label dimension in the structured label matrix; is the sample standard deviation of the k-th label dimension in the structured label matrix; i is the sample index, ranging from 1 to n, where n is the total number of candidate event samples; j is the text dimension index, ranging from 1 to d, where d is the dimension size of the text vector; k is the structured label dimension index, ranging from 1 to m, where m is the sum of the dimensions of the three one-hot encodings for event category, department, and level. The semantic component matrix of the text, which can be interpreted by the structured label matrix, is obtained based on latent principal component analysis; the expression of the semantic component matrix is ​​as follows: ; In the formula, It is a semantic component matrix of text that can be interpreted by a structured tag matrix, with dimensions n×d; This is the text score matrix extracted using partial least squares, with dimensions n× ; The number of potential principal components is used to characterize the core information in the text related to the structured tag matrix; The text payload matrix is ​​extracted based on partial least squares and has a dimension of d×k. It is used to represent the weight mapping relationship of each text dimension on the potential principal components. This is the transpose of the text payload matrix; Subtracting the text semantic component matrix from the standard event text vector matrix yields the pure semantic residual matrix; the expression for the pure semantic residual matrix is: ; In the formula, It is a purely semantic residual matrix; The standard event text vector matrix has dimensions n×d; The semantic residual matrix is ​​denormalized to restore it to its original numerical scale, resulting in the decoupled text vector matrix; the expression for the decoupled text vector is: ; In the formula, To decouple the value of the i-th adverse event sample in the text vector matrix in the j-th text dimension; For decoupled text vector matrices; Let be the value of the i-th adverse event sample in the j-th text dimension of the standardized residual matrix; Text similarity is updated based on decoupled text vectors, generating decoupled text similarity between the query event and each candidate event; the expression for decoupled text similarity is: ; In the formula, To decouple text similarity; Custom weighting coefficients; Cosine similarity is used to decouple pure semantic vectors; This is the first global semantic vector after decoupling and correction; For the first The second global semantic vector of each candidate event after decoupling and correction corresponds to the decoupling text vector matrix. The i-th row vector; This refers to the degree of overlap between entities.

[0014] In the above solution, text feature decoupling is achieved through PLS (Partial Least Squares). First, standardization is used to eliminate interference from the two types of matrix dimensions. Then, redundant semantics that can be interpreted by category, department, and grade labels are stripped from the text, leaving only the pure semantics of the handling process to generate decoupling vectors. The similarity is then recalculated. This solves the core defects of duplicate scoring and weight distortion between the original text and structured field information. The text dimension independently represents the operational details and does not interfere with the four types of structured scores. This ensures that the contribution of each dimension is objective and unbiased when performing multi-factor weighting in the future, which can improve the business rationality of the comprehensive score.

[0015] In one implementation, pre-defined adaptive weights are used to weight and fuse category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event, specifically including: Based on the major category of the queried event, the weights of each factor under the corresponding category are retrieved; the weights are configured based on Dirichlet prior weights and include categories such as diagnosis and treatment operations, nursing processes, medicines and equipment, and public facilities. Based on the preset multi-factor fusion scoring function, the system calls category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event. The expression for the multi-factor fusion scoring function is as follows: ; In the formula, For comprehensive scoring; The weights are respectively assigned to the text factor, category factor, department factor, grade factor, and time factor, and satisfy the following conditions: ; To decouple text similarity; Scoring based on category matching; Departmental matching score; For rating matching score; For time decay; When the event field of a query event indicates a missing value, the available factors are dynamically normalized to generate a corrected comprehensive score. The expression for dynamic normalization is: ; In the formula, This is the revised overall score; To query the set of factors currently available for the event; The weight of the m-th available factor; The matching score for the m-th available factor; This is the dimension index for the fusion factor, with values ​​ranging from 1 to 5, corresponding to the text factor, category factor, department factor, grade factor, and time factor, respectively.

[0016] The above scheme assigns differentiated normalized weights to the four categories of adverse events, which are in line with the quality control evaluation focus of different medical scenarios and avoid the poor adaptability problem caused by uniform fixed weights; it adopts linear weighting to integrate and decouple text and four types of structured scores, so that the contribution of each dimension is clear and controllable; at the same time, it adds a dynamic normalization mechanism for missing fields, which only relies on effective factors to recalculate scores, thus avoiding scoring distortion when fields are blank.

[0017] In one implementation, candidate events in the adverse event case set are sorted according to a comprehensive score. A predetermined number of top-ranked candidate events are selected to generate a similar event retrieval recommendation list for the query event. Retrieval recommendation information is then generated based on the comprehensive score of each candidate event, specifically including: Based on the comprehensive score, all candidate events in the adverse event case set are sorted in descending order, and a predetermined number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list; the expression for the similar event retrieval recommendation list is: ; In the formula, This outputs a list of similar event recommendations containing the top K similar candidate events for a given query event q. To retrieve the first K sorted results; The events are sorted in descending order of their overall score, meaning that events with higher matching scores are ranked higher. For comprehensive scoring; Based on a predefined recommendation explanation vector, a quantitative explanation for each factor dimension in the comprehensive score is generated; whereby the expression for the recommendation explanation vector is: ; In the formula, Recommendation explanation vector; Weighted contribution score to the semantic dimension of the text; Weighted contribution score for event category dimension; Weighted contribution score for the departmental dimension of the event; Weighted contribution score for the event level dimension; Weighted contribution over time; The core similarity segments between each candidate event and the query event in the similar event retrieval recommendation list are determined based on the weighted contribution score of the text semantic dimension. Natural language recommendation reasons for candidate events are generated based on the scores of each sub-item in the comprehensive score and the core similar segments. Retrieval recommendation information for candidate events is generated based on itemized quantitative explanations, core similar segments, and natural language recommendation reasons.

[0018] In the above scheme, Top-K candidate cases are extracted in descending order of comprehensive score, and the most similar events with the highest matching degree are pushed first, reducing the workload of manual review. By splitting the weighted contribution of each dimension through recommendation explanation vector, the impact of each indicator on the total score is intuitively displayed. Based on the text contribution, similar fragments in the original text are located, and the score is automatically used to generate popular business recommendation explanations. The scheme outputs complete quantitative data, original text evidence and readable explanations, breaking the black box drawback of traditional retrieval algorithms. This allows quality control and medical staff to clearly understand the basis of the recommendation, making it easier to review and accept. The output logic takes into account both retrieval efficiency and interpretability of results.

[0019] Secondly, this application also provides an adverse event retrieval system based on multi-factor fusion, including: a similarity calculation module, a dimension matching calculation module, a similarity update module, a comprehensive score calculation module, and a retrieval result generation module; The similarity calculation module is used to calculate the text similarity between the query event and each candidate event in the adverse event case set. The adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level, and event department. The dimension matching calculation module is used to calculate the category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate events based on the event field representation; where the event field representation is constructed based on field information and event time; The similarity update module is used to calculate the decoupled text vectors of each candidate event in the query event and adverse event case set by combining the structured label matrix, and update the text similarity to the decoupled text similarity based on the decoupled text vectors; wherein, the structured label matrix is ​​generated based on the encoding results of event category, event department and event level; The comprehensive score calculation module is used to perform weighted fusion of category matching score, department matching score, grade proximity score, time decay score and decoupled text similarity using pre-set factor adaptive weights to generate a comprehensive score between the query event and each candidate event; among which, the types of factors include category factors, department factors, grade factors, time factors and text factors; The search results generation module is used to sort each candidate event in the adverse event case set according to the comprehensive score, select a preset number of top-ranked candidate events to generate a similar event search recommendation list for the query event, and generate search recommendation information based on the comprehensive score of each candidate event.

[0020] Thirdly, this application also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the above-described method for retrieving adverse events based on multi-factor fusion.

[0021] Fourthly, this application also provides a computer-readable storage medium, which includes a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the above-described multi-factor fusion-based adverse event retrieval method. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an adverse event retrieval method based on multi-factor fusion provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of a module of an adverse event retrieval system based on multi-factor fusion provided in one embodiment of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] Example 1 See Figure 1 , Figure 1 This is a flowchart illustrating a method for retrieving adverse events based on multi-factor fusion according to an embodiment of the present invention. The embodiment includes steps 101 to 105, each step being as follows: Step 101: Calculate the text similarity between the query event and each candidate event in the adverse event case set; wherein, the adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level and event department.

[0027] In one embodiment, the adverse event case set includes several structured adverse event samples. Each adverse event sample retains at least the following fields: event category, event process, event analysis, adverse event level for medical safety, and the department involved in the event. Specifically, these include: Adverse event samples are collected through preset dimensions; these dimensions include event time, event category, and event department. For each adverse event sample, retain at least the following fields: event category, event details, event analysis, adverse event level for medical safety, and the department involved. One-hot encoding is performed on the field information of event category, event department, and event level to generate one-hot label vectors for each field, resulting in a structured label matrix; Based on a pre-trained entity recognition model in the medical field, several core entities are extracted from the field information of event process and event analysis and mapped to a unified standard code to obtain entity normalization results and generate an entity code set. The event process and the field information of the event analysis are concatenated to generate the text representation basis of the adverse event, and the event text sequence is generated by combining the entity normalization results. The event time is extracted from the field information, and the event field representation is constructed based on the three-level classification results of the event time and event category, the event process, the event analysis, the event level, the event department, and the entity code list; among them, the entity code list is determined according to the extracted core entities and the mapped standard codes; The event labeling results of adverse event samples are generated based on the comparison between the total number of adverse event samples under each event subclass and the preset subclass sample threshold; among which, the event labeling results include long-tail events and rare events; Structured bad event samples are generated based on structured label matrices, entity encoding sets, event text sequences, event field representations, and event tagging results.

[0028] In this embodiment of the invention, valid original records are screened from multiple sources within the hospital, such as the adverse event reporting system, quality control platform, and electronic medical records, while invalid data is filtered out. Preset collection dimensions limit the screening to three core criteria: event occurrence time, event three-level classification, and the department to which the event belongs. Only reports containing these three basic identifiers are retained, while blank / incomplete reports without time, classification, or department are directly discarded. Basic cleaning is performed simultaneously during the collection process: completely duplicate reports, garbled characters, and test-filled samples are removed, retaining only genuine clinical adverse event records to provide a clean raw dataset for subsequent field standardization and feature calculation. Each adverse event sample retains at least the following fields: event category, event process, event analysis, adverse event level, and event department. If the hospital's original report text has missing fields or inconsistent field naming, samples with blank fields are not discarded directly; instead, missing fields are explicitly marked and filled with placeholders. The field completeness of each sample is recorded to avoid inaccurate sample matching due to incomplete fields. After execution, the system outputs complete original samples of the fields with missing field markers. Each sample stores the above five types of core business information.

[0029] Because machine learning models cannot recognize text classification, it is necessary to count the total number of event categories (C), the total number of unique departments (D), and the total number of risk levels (G). For a single sample, generate binary one-hot vectors for each of the three fields: set the position corresponding to the sample's category to 1, and set the other category positions to 0. The same applies to departments and risk levels. Then, concatenate the three vectors into a total one-hot label vector with dimensions C+D+G. Finally, stack the one-hot vectors generated for all samples row by row to form a structured label matrix. Here, n represents the total number of samples, with each row corresponding to one adverse event sample. The medical NER named entity model is then fine-tuned, extracting five fixed core entities from the two long text segments: event description and event analysis. These entities are: disease name, drug, treatment procedure, medical device, and target patient population. For colloquialisms, abbreviations, and aliases, such as "infusion leakage / drug extravasation" and "myocardial infarction / heart attack," they are mapped to a unified hospital terminology coding library. Synonyms are converted into unique numerical codes to eliminate differences in synonymous terminology, generating a normalized list of entities for each sample. All unique standard codes from the entire sample are aggregated to form a global entity code set. The original event description and event analysis segments are then concatenated as the base text, and the normalized standard entity codes corresponding to the standardized terms are appended to the end of the text. It should be noted that event category, department, and level are not included in the event text sequence. If the category and department are directly included in the text, subsequent text vectors will repeatedly contain structured information, creating a double scoring with the one-hot coding matrix, amplifying the weight of structured factors, and causing the sorting to deviate from the quality control business logic. Therefore, only the pure plot content of the event process, symptoms, and diagnosis / treatment entities is retained to generate a clean event text sequence without structured interference. Further, standardized dates (uniformly formatted as year and month) are parsed from standardized fields, and then seven categories of fixed information are summarized to assemble a data packet representing the event: ① standardized event time; ② three-level hierarchical event classification; ③ original event process; ④ original event analysis; ⑤ standardized risk level values; ⑥ standardized department codes; ⑦ entity code list. This field representation serves as the data carrier for the entire process: subsequent structured factor matching, text generation, and long-tail determination can directly read the information within this data packet. Furthermore, the total number of samples under each event subclass is globally counted. If the preset judgment threshold is 5: if the total number of samples in a subclass is ≤5, all samples under that subclass are uniformly marked as long-tail / rare events; if the number of samples in a subclass is >5, they are marked as regular samples. Each sample is bound to a Boolean event labeling result and stored in the field's associated metadata. When a retrieval query shows a sample labeled as long-tail, the medical text representation module will activate a category vector concatenation semantic transfer algorithm to expand the semantic features of small samples and improve the recall capability of similar cases across subclasses.

[0030] For example, an original report record is as follows: Patient Li experienced drug leakage during intravenous infusion in the cardiology department on May 10, 2023, with local skin redness, swelling and pain. The incident was not treated in time and was classified as a Level II adverse nursing event. Root cause: The nurse's rounds were too long and the infusion site was not observed in time. Subcategory: Nursing - Drug Administration and Infusion - Drug Extravasation. No fields are missing.

[0031] The reported record was processed as follows: (1) Collection and screening: Collection dimensions: time 2023-05, category nursing medication infusion, department cardiology. No missing data, retained and entered into the sample pool.

[0032] (2) Core fields retained: ① Event category: Nursing - Medication administration and infusion - Extravasation; ② Event details: On May 10, 2023, Li, a patient in the Cardiology Department, experienced extravasation of medication during an intravenous infusion, resulting in local redness, swelling, and pain; ③ Event analysis: The nurse's rounds were too long, and the infusion puncture site was not checked regularly; ④ Event level: Level II; ⑤ Event department: Cardiology Department. ⑥ No blank fields, completeness marked as 100%.

[0033] One-hot encoding generates a single row of the label matrix: Global statistics: 42 event categories, 36 clinical departments, and 4 risk levels; One sample one-hot vector: Nursing medication / infusion category position 1, Cardiology department position 1, Level II risk position 1, all other dimensions are 0; This vector serves as the structured label matrix. One of the lines.

[0034] (4) Medical entity extraction + normalization, generating entity code set: NER model extracts entities: intravenous infusion, drug extravasation, local skin redness and swelling, nurse rounds; Mapping the hospital's standard codes: Intravenous infusion = ENT012, extravasation of medication = ENT047, local skin redness and swelling = ENT091, nurse rounds = ENT116; This entity code list = [ENT012,ENT047,ENT091,ENT116]; Four new codes have been added to the global entity code set.

[0035] (5) Concatenate to generate a pure event text sequence: Basic text = event description + event analysis; add standardized entity terminology, but do not include department or rank text; Final text sequence: 2023-05-10 Cardiology patient Li experienced drug extravasation during intravenous infusion, resulting in local redness, swelling, and pain; the nurse's rounds were too long, and the infusion puncture site was not checked regularly; Standard entities: intravenous infusion, drug extravasation, local skin redness and swelling, nurse rounds.

[0036] (6) Constructing event field representations: 1. Event Time: May 2023; 2. Three-level classification: Major category: Nursing-related; Main category: Drug administration procedures; Subcategory: Extravasation of medication. 3. Original text of the event's events and original text of the event analysis; 4. Level value: 2 (Level II mapping value); 5. Standardized departmental coding: Internal Medicine - Cardiology; 6. Entity code list: [ENT012,ENT047,ENT091,ENT116].

[0037] (7) Long-tail event labeling: The total number of samples in the subclass "Nursing - Drug administration and infusion - Drug extravasation" is 8, which is greater than the preset threshold of 5; this sample is labeled as: non-long-tail / routine event.

[0038] (8) Package and generate standardized samples.

[0039] In one embodiment, calculating the text similarity between the queried event and each candidate event in the adverse event case set specifically includes: The field information of the event process and event analysis is preprocessed, and the preprocessed text is input into a pre-trained medical domain model for encoding to generate the first global semantic vector of the query event and the second global semantic vector of each candidate event. The entity overlap degree between the first global semantic vector and each of the second global semantic vectors is calculated based on the entity encoding set of adverse events; whereby the expression for the entity overlap degree is: ; In the formula, To query events; For the first One candidate event; For the degree of overlap between entities; The set of entity codes for the query event; For the first A set of entity codes for candidate events; The text similarity between the query event and each candidate event is calculated by combining entity overlap; the expression for text similarity is: ; In the formula, For text similarity, Custom weighting coefficients; Cosine similarity of global semantic vectors; This is the first global semantic vector for the query event; For the first The second global semantic vector of the candidate events.

[0040] In this embodiment of the invention, a unified preprocessing operation is performed on the query event, the event process of all candidate events in the case set, and the event analysis text. This includes removing punctuation marks, filtering meaningless common stop words, and completing Chinese word segmentation. Throughout the process, medical professional terms and standardized entity words such as intravenous infusion, drug names, surgical procedures, and symptom descriptions are preserved to avoid the loss of key business semantics. After preprocessing, the query event text is input separately into a pre-trained language model in the medical field, outputting a fixed-dimensional vector, defined as the first global semantic vector. Then, the preprocessed text of each candidate event in the case library is fed into the same model, generating a second global semantic vector for each candidate event. The second vectors of all candidate events are stacked row-wise to form a global text matrix, providing a unified numerical input for subsequent batch similarity calculations. This vector can capture the overall semantic logic of the entire text. For example, if two texts have different word orders but both describe drug extravasation, the vector space distance will be significantly closer. However, relying solely on the global vector can easily lose details of key local diagnostic and treatment entities; therefore, subsequent supplementation and correction using entity overlap are necessary. The entity overlap ratio is expressed using Jaccard similarity coefficient logic: the numerator is the number of entity codes simultaneously contained in both types of events, and the denominator is the total number of all unique entity codes in both types of events. The calculated result ranges from [0,1]. The closer the value is to 1, the higher the overlap in the diseases, drugs, procedures, and devices involved in the two adverse events, and the consistent core risk factors. A value close to 0 indicates that the two events are completely unrelated in terms of diagnostic and treatment elements. For example, if two infusion-related events share three core entities and only one unique entity exists, the entity overlap ratio will reach a high level; if an infusion event and a surgical event have no overlap in entities, the overlap ratio will be 0. Entity overlap ratio, as a separate one-dimensional matching indicator, can compensate for the problem that global vectors easily overlook local key operations and symptom entities, avoiding erroneous matches where "the overall semantics are similar but the core risk operations are completely different." Then, the text similarity between the query event and each candidate event is calculated in combination with the entity overlap ratio. It should be noted that this text similarity is a basic text similarity score without PLS decoupling correction, which can balance the two dimensions of overall text semantics and core diagnostic and treatment element matching. In the expression for text similarity The weight is a manually preset, adaptively adjustable balanced weight, generally fixed at 0.7, which means that the overall semantic flow of the whole text is given priority by default. If the current query event is marked as a long-tail rare event, it can be reduced to 0.5~0.6 to increase the weight ratio of entity matching, strengthen the matching role of core elements in small samples, and improve the recall effect of rare events.

[0041] Step 102: Calculate the category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate events based on the event field representation; where the event field representation is constructed based on field information and event time.

[0042] In one embodiment, the category matching score, department matching score, grade proximity score, and time decay score of the query event and candidate events are calculated based on the event field representation, specifically including: Obtain the three-level classification results of the event categories, and calculate the category matching score based on the three-level classification results; the matching rule expression for the category matching score is as follows: ; In the formula, For category matching functions; To query events; for One candidate event; , These are event subclasses for query events and candidate events, respectively. This is a category main class mapping function used to map event subclasses to their respective intermediate level main classes; This is a category mapping function used to map the main class of an event to its corresponding top-level category; The department matching score is calculated based on the event department in the event field and the pre-defined clinical department matching rules; the expression for the clinical department matching rule is as follows: ; In the formula, For department matching functions; , These are the specific departments to which the query event and candidate events belong, respectively. For departmental first-level classification mapping function; The level proximity score is calculated based on the event level in the event field and a pre-defined level matching rule; the expression for the level matching rule is as follows: ; In the formula, It is a proximity function of the order of precedence; , These are the event levels for the query event and the candidate event, respectively. This is a level mapping function that maps event levels I-IV to 4, 3, 2, and 1 respectively, from highest to lowest severity. This is the attenuation coefficient for grade differences, with a default value of 0.35; The time decay score is calculated based on the event time indicated in the event field and a pre-defined time decay function; the expression for the time decay function is as follows: ; In the formula, It is a time decay function; The time interval between the query event and the candidate events; This is the time half-life parameter, with a default value of 12 months. This is the maximum effective time window, with a default value of 60 months.

[0043] In this embodiment of the invention, category matching is used to measure the degree of similarity between two adverse events under the national unified three-level classification system. Cases within the same subclass have the highest reference value, while cases across major categories have only a weak reference value. It should be noted that the "major category-subclass" three-level classification system corresponds to the National Medical Quality and Safety Event Classification Standard. The top-level major categories include categories such as diagnosis and treatment related, nursing related, drug related, and facility related. If the query result is completely identical to the candidate subclass (e.g., both are nursing-administration-extravasation),... The highest reference value is achieved when the event business type is completely consistent; if the subclasses are different but the intermediate main class after mapping is the same (e.g., extravasation of medication, wrong administration, both belonging to the main class of medication operation), the score is 0.7, retaining a relatively high reference value; if the main classes are different but the top-level category is the same (e.g., extravasation of medication, patient fall, both belonging to the nursing category), the score is 0.3, possessing only broad reference value; if the top-level categories are completely different (nursing event vs. surgical treatment event), the score is only 0.1, with almost no reuse value. Layered scoring avoids the drawbacks of Boolean filtering, preventing the direct elimination of cross-category reference cases due to incomplete consistency in the factor classes, thus balancing precise retrieval and recall of rare long-tail events. Department matching score is used to quantify the similarity of scenarios in the departments where the incident occurred. Different specialties have vastly different treatment processes and risk control priorities, making rectification plans for cases within the same department the most reusable. The query department is extracted from the field representation. Candidate Departments Based on the departmental first-level classification mapping function Calculation: Department names match exactly (all are cardiology departments). For cases where the ward procedures and population risks are completely identical, even if the departments are different, the mapped primary specialty system is the same (cardiology and respiratory medicine belong to the internal medicine system; hepatobiliary and general surgery belong to the surgical system), resulting in a score of 0.7, indicating similar specialty management logic. For cases belonging to completely unrelated specialty systems (e.g., cardiology and the operating room), the score is 0.2, indicating significant differences in departmental scenarios. Surgical events will have a specialty bonus added on top of this, further increasing the ranking weight of cases within the same operating room, aligning with surgical safety management needs. The proximity score is used to measure the similarity in severity risk between two events, avoiding the use of a fatal Level I event to reference a minor Level IV hazard event, and preventing misleading quality control assessments. Level mapping rules: The levels I / II / III / IV are mapped to 4, 3, 2, and 1 respectively, with higher values ​​indicating more severe damage; attenuation coefficient. The larger the grade difference, the linearly decreasing the matching score. When the difference exceeds 3 points, it is directly truncated to 0: Same grade (difference 0): score 1; Difference of 1 grade (Ⅰ&Ⅱ, Ⅱ&Ⅲ): It still serves as a strong reference; there is a difference of 2 levels (Ⅰ&Ⅲ, Ⅱ&Ⅳ): Only slightly applicable; differences of ≥3 levels (Ⅰ & Ⅳ): score 0, completely lacking risk reference value. Through linear decay and truncation mechanisms, events of different severity can be strictly distinguished, avoiding the mixing of high-risk and potential cases. The time decay score measures the timeliness of cases. Hospital systems and nursing procedures are updated annually, and the reference value of outdated cases decays exponentially. A half-life exponential decay model is used, with the decay logic as follows: at 0 months, it is considered a contemporary event and given the highest score of 1; at 12 months, the score is 0.5; at 24 months, the score is 0.25, decreasing exponentially over time; cases older than 60 months are directly set to 0, completely excluding historical cases under outdated systems and ensuring that the recommended results comply with current management standards.

[0044] Step 103: Calculate the decoupled text vectors of each candidate event in the query event and adverse event case set by combining the structured label matrix, and update the text similarity to the decoupled text similarity based on the decoupled text vectors; wherein, the structured label matrix is ​​generated based on the encoding results of event category, event department and event level.

[0045] In one embodiment, the decoupled text vectors of each candidate event in the query event and adverse event case set are calculated by combining the structured label matrix, and the text similarity is updated to decoupled text similarity based on the decoupled text vectors, specifically including: The event text sequences of each candidate event in the adverse event case set are concatenated to construct an event text vector matrix; The event text vector matrix and the structured tag matrix are standardized to obtain a standard event text vector matrix and a standard structured tag matrix; the expression for the standardization process is as follows: ; ; In the formula, In the standard event text vector matrix, the value of the i-th adverse event sample in the j-th text dimension is given. Let X be the original feature value of the i-th adverse event sample in the j-th dimension. Let be the sample mean of the j-th dimension in the event text vector matrix X; Let be the sample standard deviation of the j-th dimension in the event text vector matrix X; In a standard structured label matrix, the value of the i-th adverse event sample in the k-th label dimension is given. In the structured label matrix, the original value of the i-th adverse event sample in the k-th label dimension is given. This represents the sample mean of the k-th label dimension in the structured label matrix; is the sample standard deviation of the k-th label dimension in the structured label matrix; i is the sample index, ranging from 1 to n, where n is the total number of candidate event samples; j is the text dimension index, ranging from 1 to d, where d is the dimension size of the text vector; k is the structured label dimension index, ranging from 1 to m, where m is the sum of the dimensions of the three one-hot encodings for event category, department, and level. The semantic component matrix of the text, which can be interpreted by the structured label matrix, is obtained based on latent principal component analysis; the expression of the semantic component matrix is ​​as follows: ; In the formula, It is a semantic component matrix of text that can be interpreted by a structured tag matrix, with dimensions n×d; This is the text score matrix extracted using partial least squares, with dimensions n× ; The number of potential principal components is used to characterize the core information in the text related to the structured tag matrix; The text payload matrix is ​​extracted based on partial least squares and has a dimension of d×k. It is used to represent the weight mapping relationship of each text dimension on the potential principal components. This is the transpose of the text payload matrix; Subtracting the text semantic component matrix from the standard event text vector matrix yields the pure semantic residual matrix; the expression for the pure semantic residual matrix is: ; In the formula, It is a purely semantic residual matrix; The standard event text vector matrix has dimensions n×d; The semantic residual matrix is ​​denormalized to restore it to its original numerical scale, resulting in the decoupled text vector matrix; the expression for the decoupled text vector is: ; In the formula, To decouple the value of the i-th adverse event sample in the text vector matrix in the j-th text dimension; For decoupled text vector matrices; Let be the value of the i-th adverse event sample in the j-th text dimension of the standardized residual matrix; Text similarity is updated based on decoupled text vectors, generating decoupled text similarity between the query event and each candidate event; the expression for decoupled text similarity is: ; In the formula, To decouple text similarity; Custom weighting coefficients; Cosine similarity is used to decouple pure semantic vectors; This is the first global semantic vector after decoupling and correction; For the first The second global semantic vector of each candidate event after decoupling and correction corresponds to the decoupling text vector matrix. The i-th row vector; This refers to the degree of overlap between entities.

[0046] The original text description naturally contains event category, department, and level information. Directly fusing the original text similarity and structured score is equivalent to calculating the score twice for the same set of business information, which amplifies the weight of the category / department and reduces the contribution of the process text. In this embodiment of the invention, redundant information that can be interpreted by one-hot tags is removed from the text through partial least squares (PLS), retaining only the pure event process semantics. The decoupled text and the corrected similarity are then used for the final fusion. The text vectors of all candidate events in the case set are stacked row by row in sample order to generate the original text vector matrix. Each row in the matrix corresponds to one sample, and each column corresponds to one dimension of the semantic vector. Z-score standardization is applied to both the text vector matrix and the structured standard matrix to eliminate dimensional interference. After standardization, all dimensions of both types of matrices have a mean of 0 and a variance of 1, eliminating numerical scale differences and ensuring that the PLS extracted components are not affected by the magnitude of the values. Then, the NIPALS iterative PLS algorithm is used with the standard text vector matrix... As the independent variable and the structured label matrix as the dependent variable, extract... One potential principal component was identified based on cross-validation. The dimension is 1 / 3 of the structured label dimension. Then, multiplying the text score matrix extracted using partial least squares with the transpose of the text load matrix yields the text semantic component matrix. This matrix retains only the semantic content in the text that can be interpreted by category, department, and level labels, i.e., redundant and repetitive information. Next, element-wise subtraction is performed to remove redundant information bound to structured labels within the text, retaining only pure process semantics that cannot be interpreted by category, department, or level. Semantic residual matrix. Each row of vectors represents the unique textual features remaining after removing interference from categories, departments, and levels, such as event operations, disease conditions, and treatment procedures. For example, in the original text "cardiology department, level II, extravasation of medication (category)", the semantics corresponding to "cardiology department, level II, extravasation of medication (category)" will be represented by the text semantic component matrix. After stripping away the data, the residuals retain only the pure process descriptions of "infusion leakage, skin redness and swelling, and insufficient patrols," thus eliminating the duplication of scores for structured information at the source. Finally, based on the calculated decoupled text vectors, the original similarity fusion logic is reused, but replaced with the cosine value of the decoupled pure semantic vectors to obtain the decoupled text similarity.

[0047] Step 104: Use pre-set adaptive factor weights to perform weighted fusion of category matching score, department matching score, grade proximity score, time decay score and decoupled text similarity to generate a comprehensive score between the query event and each candidate event; among which, the types of factors include category factor, department factor, grade factor, time factor and text factor.

[0048] In one embodiment, a pre-defined factor adaptive weighting is used to weight and fuse category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event, specifically including: Based on the major category of the queried event, the weights of each factor under the corresponding category are retrieved; the weights are configured based on Dirichlet prior weights and include categories such as diagnosis and treatment operations, nursing processes, medicines and equipment, and public facilities. Based on the preset multi-factor fusion scoring function, the system calls category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event. The expression for the multi-factor fusion scoring function is as follows: ; In the formula, For comprehensive scoring; The weights are respectively assigned to the text factor, category factor, department factor, grade factor, and time factor, and satisfy the following conditions: ; To decouple text similarity; Scoring based on category matching; Departmental matching score; For rating matching score; For time decay; When the event field of a query event indicates a missing value, the available factors are dynamically normalized to generate a corrected comprehensive score. The expression for dynamic normalization is: ; In the formula, This is the revised overall score; To query the set of factors currently available for the event; The weight of the m-th available factor; The matching score for the m-th available factor; This is the dimension index for the fusion factor, with values ​​ranging from 1 to 5, corresponding to the text factor, category factor, department factor, grade factor, and time factor, respectively.

[0049] In this embodiment of the invention, a Dirichlet prior weighting is adopted, with initial normalized weights determined by medical quality control experts based on the management priority of different adverse events. The weights naturally satisfy... Adapted to Dirichlet distribution simplex constraints. Each event category is equipped with a dedicated five-factor weight configuration: (1) Diagnostic and therapeutic procedures: Clinical practice relies heavily on cases from the same department and at the same risk level, therefore the weight of structured factors related to department and risk level is relatively high; (2) Nursing process category: The core of nursing reference is the details of the operation text, with text factors having the highest weight, while also taking into account the timeliness of the updated standards; (3) Pharmaceuticals and medical devices: Drug risks are strongly correlated with their category and severity, with category and grade having higher weighting. (4) Public facilities: Facility accidents have the highest reference value among similar categories, and the category weighting is the largest.

[0050] Extract the top-level category from the query event field and match the corresponding group. The weights serve as the baseline coefficients for this fusion calculation. When all fields corresponding to the five factors of the query event are complete, the multi-factor fusion scoring function is used directly. The five terms in the formula are multiplied by their respective matching indicators and exclusive weights, and then summed. This indicates the semantic contribution of decoupled plain text, representing only the similarity of event operations and process details, with redundant interference from categories / departments removed; Event classification dimensions and their contribution to business matching; : The department where the incident occurred contributed to the scene matching; : Risk severity matches contribution; The case's timeliness is a contributing factor. All sub-items have a score range of [0,1] and a total weight of 1; therefore, the final comprehensive score is calculated as follows: The score naturally falls between 0 and 1. A higher score indicates a stronger comprehensive reference value for the candidate event across all dimensions, including text flow, business classification, departmental scenario, risk level, and timeliness, and can be directly used for descending sorting. A minimum filtering threshold of 0.4 is set; candidates with a comprehensive score below 0.4 are directly eliminated to reduce interference from low-reference cases. If a queried event has blank fields (e.g., no event level or no reporting time), the corresponding factor has no valid score and cannot be directly substituted into the standard formula. Otherwise, the missing dimension will be calculated as 0 by default, lowering the total score and causing sorting distortion. Therefore, a dynamic normalization correction mechanism is designed. Only the available factors are weighted and summed, then divided by the sum of the weights of the valid factors to re-normalize, ensuring that the corrected score still remains in the 0-1 range, eliminating scoring bias caused by missing fields. For example, querying an event with a missing time field, the time factor... Failure , only The weighted average of the four factors is then divided by the sum of the four factors to achieve normalization. This prevents the time score from being forcibly set to 0, which would lower the overall matching accuracy and ensure that the sorting results still conform to the business logic even in scenarios with missing fields.

[0051] Step 105: Sort the candidate events in the adverse event case set according to the comprehensive score, select the top-ranked candidate events to generate a similar event retrieval recommendation list for the query event, and generate retrieval recommendation information based on the comprehensive score of each candidate event.

[0052] In one embodiment, the candidate events in the adverse event case set are sorted according to the comprehensive score, a preset number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list for the query event, and retrieval recommendation information is generated based on the comprehensive score of each candidate event, specifically including: Based on the comprehensive score, all candidate events in the adverse event case set are sorted in descending order, and a predetermined number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list; the expression for the similar event retrieval recommendation list is: ; In the formula, This outputs a list of similar event recommendations containing the top K similar candidate events for a given query event q. To retrieve the first K sorted results; The events are sorted in descending order of their overall score, meaning that events with higher matching scores are ranked higher. For comprehensive scoring; Based on a predefined recommendation explanation vector, a quantitative explanation for each factor dimension in the comprehensive score is generated; whereby the expression for the recommendation explanation vector is: ; In the formula, Recommendation explanation vector; Weighted contribution score to the semantic dimension of the text; Weighted contribution score for event category dimension; Weighted contribution score for the departmental dimension of the event; Weighted contribution score for the event level dimension; Weighted contribution over time; The core similarity segments between each candidate event and the query event in the similar event retrieval recommendation list are determined based on the weighted contribution score of the text semantic dimension. Natural language recommendation reasons for candidate events are generated based on the scores of each sub-item in the comprehensive score and the core similar segments. Retrieval recommendation information for candidate events is generated based on itemized quantitative explanations, core similar segments, and natural language recommendation reasons.

[0053] In this embodiment of the invention, the overall score is used as the sorting keyword to perform descending sorting; the higher the overall score, the higher the event's position in the sorting list. Samples with a matching degree lower than 0.4 are pre-selected and do not participate in the sorting. After sorting, a TopK truncation operation is performed, with a default preset number. It also supports administrators in customizing and adjusting the K value, ultimately generating a set of similar recommendations specific to the query event q. The overall score is then broken down into the weighted contribution scores of five factors, visually displaying the contribution of each dimension to the total score, thus solving the problem of "high overall score, but not knowing which matching factor played a key role." The contribution of decoupled text semantics represents the total score gain due to similarity in event operations and process details; This indicates the contribution of the event category, representing the total score gain brought about by the matching of the classification hierarchy; The contribution of the department scene represents the total score gain brought about by matching the department where the incident occurred; This indicates the contribution of risk level, representing the total score gain due to similar severity levels; The contribution of timeliness represents the overall score gain due to the freshness of the case. The contribution percentage of each dimension can be directly compared using the recommendation explanatory vector. For example, the highest text contribution score indicates a high degree of overlap in the core operational processes of the two events; a very low time contribution score indicates that the case is old and has limited reference value, providing quantitative data support for manual judgment. Then, based on the attention weights of the decoupled pure semantic vectors, the text semantic weighted contribution is combined. High- and low-level execution segment extraction: The higher the text dimension contribution, the higher the overlap in the text descriptions of the two events. The system automatically selects 3-5 sentences with the highest matching degree from the original text as core similar segments and highlights them. Only key business statements such as operational procedures, symptom descriptions, and handling errors are retained in the segments, while irrelevant background text is removed. The aim is to allow managers to quickly identify highly similar plots between two events without reading the entire text, and to quickly determine whether the case has review and reference value. Furthermore, based on the scores of each sub-item in the comprehensive score and the core similar segments, a structured recommendation reason is automatically generated, as shown in the example format below: This case has a XX% overall relevance to the current event. Recommendation basis: Event category matching: Both are [XX type] adverse events, and the categories are completely identical; Scenario matching: All occurred in [XX department], and belong to the [XX specialty system]; Similar severity: Both events are classified as [Level XX], indicating similar levels of risk; Similar core plot points: They closely match in terms of [key plot descriptions], etc. Case timeliness: The incident occurred XX months ago, making it highly relevant to the relevant regulations.

[0054] Finally, the generated comprehensive score, sub-item quantitative explanations, core similar segments, and natural language recommendation reasons are integrated and packaged. The integrated information can be directly output to the system front end, supporting the export of event review pages and training reports.

[0055] As an optimized solution of this invention, after generating retrieval recommendation information, it also includes building an active learning closed loop and knowledge accumulation process to realize the reuse of recommendation results, iteration of manual feedback, and continuous accumulation of case knowledge. The following processing steps are included: (1) Case analysis assistance: Output the complete event process, handling measures, and rectification suggestions of similar events obtained from the search, as a reference for the analysis and handling of newly emerging adverse events, and assist managers in quickly formulating risk response plans; (2) Automatic generation of review materials: Select representative similar cases by event category, department where the incident occurred, and risk level, automatically summarize the common causes and standardized improvement measures of similar events, and output complete materials that can be directly used for departmental training and quality control review; (3) Accumulation of key case library: Store events marked as high reference value by hand into the key case library separately, and automatically give matching priority bonus when searching later; at the same time, extract the key points of handling, root causes, and rectification plans from the adverse event case library in a structured manner to build a standardized adverse event knowledge base; (4) Dirichlet-multinomial Bayes weight adaptive optimization: Based on business feedback, automatically iterate the normalized weights of the five major fusion factors to solve the problem of matching logic deviation caused by fixed weights and unconstrained parameter tuning from the root. Its implementation logic includes: ① Prior distribution initialization: For the four major top-level event categories, use the initial weight vector marked by medical quality control experts. Constructing the Dirichlet Prior Distribution ; where w is a 5-dimensional weight vector, corresponding to the weights [α,β,γ,δ,ε] of the five fusion factors: text, category, department, grade, and time. It satisfies the normalization constraint that the sum of all weight components is 1, and is the random variable to be estimated. This is a standard symbol for probability and statistics, representing "following a distribution of...". , short for Dirichlet distribution, is a probability distribution defined on a multidimensional simplex, specifically used to model a set of proportional / weighted variables that sum to 1 and satisfy normalization constraints; prior hyperparameter vector. Each component satisfies a priori concentration ① Fixed value of 100, equivalent to 100 historical feedback samples, balancing weight stability and iteration flexibility, avoiding drastic fluctuations or long-term stagnation in small sample feedback; ② Feedback counting statistics: Collecting three levels of manual feedback: "fully adopted, partially adopted, and not adopted". Fully adopted samples are counted cumulatively according to the score ratio of each factor. Partially adopted samples are used to incentivize low-contribution factors and increase their retrieval weight. Unrecognized feedback without a clear reason is only logged and does not participate in the count update to prevent abnormal feedback from polluting the model; ③ Posterior normalized weight solution: Utilizing the conjugate property of Dirichlet distribution and multinomial distribution, the expected value of the posterior distribution is taken as the update weight. The calculation formula is as follows: , ensure that the total weight of the five factors of text, category, department, level and time after the update is always equal to 1; ④ Weight fluctuation constraint: the adjustment range of single factor weight in a single iteration shall not exceed the initial value ±30%, and after exceeding the threshold, it shall be reset to the original value. A monthly batch update cycle is set to balance the iteration speed and business stability; (5) Long-tail case incremental update: new rare subclass samples are regularly included in the long-tail event pool, the category semantic transfer model is iteratively optimized, and the similar case recall accuracy of small sample adverse events is continuously improved; (6) Full business closed loop feedback: the Dirichlet adaptive weight after iterative update is fed back to the multi-factor fusion scoring link to form a sustainable business closed loop of new event retrieval output similar cases - manual judgment feedback - automatic weight optimization + case knowledge accumulation.

[0056] In this embodiment of the invention, a multi-factor fusion-based adverse event retrieval device is also provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the above-described multi-factor fusion-based adverse event retrieval method.

[0057] In this embodiment of the invention, a computer-readable storage medium is also provided, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the above-described adverse event retrieval method based on multi-factor fusion when it is running.

[0058] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to carry out the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a multi-factor fusion-based adverse event retrieval device.

[0059] Multi-factor fusion-based adverse event retrieval devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These devices may include, but are not limited to, processors, memory, and displays. Those skilled in the art will understand that the above-mentioned components are merely examples of multi-factor fusion-based adverse event retrieval devices and do not constitute a limitation on them. The device may include more or fewer components, combinations of certain components, or different components. For example, multi-factor fusion-based adverse event retrieval devices may also include input / output devices, network access devices, buses, etc.

[0060] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the multi-factor fusion-based adverse event retrieval device, connecting all parts of the device via various interfaces and lines.

[0061] The memory can be used to store computer programs and / or modules. The processor implements various functions of the multi-factor fusion-based adverse event retrieval device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area can store data created based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0062] The module for retrieving adverse events based on multi-factor fusion, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. Those skilled in the art can understand and implement this without any inventive effort.

[0063] This invention provides a multi-factor fusion-based adverse event retrieval method. It abandons the single-text matching mode and fully mines the multi-dimensional structured core fields of adverse event cases, such as category, department, safety level, and occurrence time. Combined with unstructured event text features, it achieves multi-factor collaborative retrieval, comprehensively covering the attribute and content features of adverse events. By constructing a structured tag matrix to generate decoupled text vectors, it optimizes and updates text similarity, effectively avoiding the matching bias problems caused by semantic redundancy and feature coupling in traditional text retrieval, making the text matching results more closely match the professional scenario characteristics of medical adverse events. Simultaneously, it specifically adds four structured scoring dimensions: category, department, level, and time. Combined with time decay scores, it aligns with the time-sensitive application characteristics of adverse events, taking into account both the inherent attribute matching of events and the dynamic adaptation of the time dimension. Furthermore, it employs factor adaptive weights to complete the weighted fusion of multi-dimensional scores, which can autonomously adapt to the retrieval needs of different query events, avoiding the retrieval rigidity caused by fixed weights, and ensuring the scientific and reasonable nature of the comprehensive score. By generating a search recommendation list based on accurate comprehensive scoring and ranking, it not only achieves efficient and accurate screening of similar adverse events, but also ensures the hierarchy and reference value of search results. It can provide accurate and effective case support for medical safety review, risk prevention and control, and event tracing, significantly improving the intelligence level and clinical application value of medical adverse event retrieval.

[0064] Example 2 See Figure 2 , Figure 2This is a schematic diagram of a multi-factor fusion-based adverse event retrieval system provided in one embodiment of the present invention. The embodiment of the present invention provides a multi-factor fusion-based adverse event retrieval system, including: a similarity calculation module 2011, a dimension matching calculation module 202, a similarity update module 203, a comprehensive score calculation module 204, and a retrieval result generation module 205; The similarity calculation module 201 is used to calculate the text similarity between the query event and each candidate event in the adverse event case set; wherein, the adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level and event department; The dimension matching calculation module 202 is used to calculate the category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate events based on the event field representation; wherein, the event field representation is constructed based on field information and event time; The similarity update module 203 is used to calculate the decoupled text vectors of each candidate event in the query event and adverse event case set by combining the structured label matrix, and update the text similarity to the decoupled text similarity based on the decoupled text vectors; wherein, the structured label matrix is ​​generated based on the coding results of event category, event department and event level; The comprehensive score calculation module 204 is used to perform weighted fusion of category matching score, department matching score, grade proximity score, time decay score and decoupled text similarity using pre-set factor adaptive weights to generate a comprehensive score between the query event and each candidate event; wherein, the types of factors include category factors, department factors, grade factors, time factors and text factors; The search result generation module 205 is used to sort each candidate event in the adverse event case set according to the comprehensive score, select a preset number of top-ranked candidate events to generate a similar event search recommendation list for the query event, and generate search recommendation information based on the comprehensive score of each candidate event.

[0065] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0066] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for retrieving adverse events based on multi-factor fusion, characterized in that, include: Calculate the text similarity between the query event and each candidate event in the adverse event case set; wherein, the adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level, and event department; The category matching score, department matching score, grade proximity score, and time decay score of the query event and the candidate events are calculated based on the event field representation; wherein, the event field representation is constructed based on the field information and the event time; The structured label matrix is ​​used to calculate the decoupled text vectors of the query event and each candidate event in the adverse event case set, and the text similarity is updated to the decoupled text similarity based on the decoupled text vectors; wherein, the structured label matrix is ​​generated based on the encoding results of the event category, the event department, and the event level; The category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity are weighted and fused using pre-set factor adaptive weights to generate a comprehensive score between the query event and each candidate event; wherein, the types of factors include category factors, department factors, grade factors, time factors, and text factors; The expression for the decoupled text vector is: ； In the formula, To decouple the value of the i-th adverse event sample in the text vector matrix in the j-th text dimension; Let be the value of the i-th adverse event sample in the j-th text dimension of the standardized residual matrix; Let X be the sample standard deviation of the j-th dimension in the event text vector matrix X; Let be the sample mean of the j-th dimension in the event text vector matrix X; The expression for the decoupled text similarity is: ； In the formula, To decouple text similarity; Custom weighting coefficients; Cosine similarity is used to decouple pure semantic vectors; This is the first global semantic vector after decoupling and correction; For the first The second global semantic vector of candidate events after decoupling and correction; To query events; For the first One candidate event; This refers to the degree of overlap between entities. The candidate events in the adverse event case set are sorted according to the comprehensive score, and a preset number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list for the query event. Retrieval recommendation information is generated based on the comprehensive score of each candidate event.

2. The adverse event retrieval method based on multi-factor fusion as described in claim 1, characterized in that, The adverse event case set contains several structured adverse event samples. Each adverse event sample retains at least the following fields: event category, event process, event analysis, adverse event level for medical safety, and the department involved. Specifically, these include: Adverse event samples are collected through preset dimensions; wherein, the dimensions include event time, event category, and event department; For each adverse event sample, retain at least the following fields: event category, event details, event analysis, adverse event level for medical safety, and the department involved. One-hot encoding is performed on the field information of the event category, the event department, and the event level to generate one-hot tag vectors for each field, thus obtaining the structured tag matrix; Based on a pre-trained entity recognition model in the medical field, several core entities are extracted from the field information of the event process and event analysis and mapped to a unified standard code to obtain entity normalization results and generate an entity code set. The event process and the field information of the event analysis are concatenated to generate the text representation basis of the adverse event, and the event text sequence is generated by combining the entity normalization result; The event time is extracted from the field information, and the event field representation is constructed based on the event time, the three-level classification result of the event category, the event process, the event analysis, the event level, the event department, and the entity code list; wherein, the entity code list is determined according to the extracted core entities and the mapped standard codes; Event labeling results for adverse event samples are generated based on the comparison between the total number of adverse event samples under each event subclass and the preset subclass sample threshold; wherein, the event labeling results include long-tail events and rare events; A structured sample of malfunctioning events is generated based on the structured tag matrix, the entity encoding set, the event text sequence, the event field representation, and the event tagging result.

3. The adverse event retrieval method based on multi-factor fusion as described in claim 2, characterized in that, The calculation of the text similarity between the query event and each candidate event in the adverse event case set specifically includes: The field information of the event process and the event analysis is preprocessed, and the preprocessed text is input into a pre-trained medical domain model for encoding to generate a first global semantic vector of the query event and a second global semantic vector of each candidate event; The entity overlap degree between the first global semantic vector and each of the second global semantic vectors is calculated based on the entity encoding set of adverse events; wherein, the expression for the entity overlap degree is: ； In the formula, To query events; For the first One candidate event; For the degree of overlap between entities; The set of entity codes for the query event; For the first A set of entity codes for candidate events; The text similarity between the query event and each candidate event is calculated based on the entity overlap; wherein the expression for the text similarity is: ； In the formula, For text similarity, Custom weighting coefficients; Cosine similarity of global semantic vectors; This is the first global semantic vector for the query event; For the first The second global semantic vector of the candidate events.

4. The adverse event retrieval method based on multi-factor fusion as described in claim 2, characterized in that, The calculation of category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate events based on the event field representation specifically includes: Obtain the three-level classification result of the event category, and calculate the category matching score based on the three-level classification result; wherein, the matching rule expression of the category matching score is: ； In the formula, For category matching functions; To query events; For the first One candidate event; , These are event subclasses for query events and candidate events, respectively. This is a category main class mapping function used to map event subclasses to their respective intermediate level main classes; This is a category mapping function used to map the main class of an event to its corresponding top-level category; The department matching score is calculated based on the event department represented in the event field and a pre-defined clinical department matching rule; wherein, the expression of the clinical department matching rule is: ； In the formula, For department matching functions; , These are the specific departments to which the query event and candidate events belong, respectively. For departmental first-level classification mapping function; The level proximity score is calculated based on the event level in the event field and a pre-defined level matching rule; wherein the expression for the level matching rule is: ； In the formula, It is a proximity function of the order of precedence; , These are the event levels for the query event and the candidate event, respectively. This is a level mapping function that maps event levels I-IV to 4, 3, 2, and 1 respectively, from highest to lowest severity. This is the attenuation coefficient for grade differences, with a default value of 0.35; The time decay score is calculated based on the event time represented in the event field and a pre-defined time decay function; wherein the expression for the time decay function is: ； In the formula, It is a time decay function; The time interval between the query event and the candidate events; This is the time half-life parameter, with a default value of 12 months. This is the maximum effective time window, with a default value of 60 months.

5. The adverse event retrieval method based on multi-factor fusion as described in claim 1, characterized in that, The process of calculating decoupled text vectors for the query event and each candidate event in the adverse event case set by combining the structured label matrix, and updating the text similarity to decoupled text similarity based on the decoupled text vectors, specifically includes: The event text sequences of each candidate event in the adverse event case set are concatenated to construct an event text vector matrix; The event text vector matrix and the structured tag matrix are standardized to obtain a standard event text vector matrix and a standard structured tag matrix; wherein the expression for the standardization process is: ； ； In the formula, In the standard event text vector matrix, the value of the i-th adverse event sample in the j-th text dimension is given. Let X be the original feature value of the i-th adverse event sample in the j-th dimension. Let be the sample mean of the j-th dimension in the event text vector matrix X; Let be the sample standard deviation of the j-th dimension in the event text vector matrix X; In a standard structured label matrix, the value of the i-th adverse event sample in the k-th label dimension is given. In the structured label matrix, the original value of the i-th adverse event sample in the k-th label dimension is given. This represents the sample mean of the k-th label dimension in the structured label matrix; is the sample standard deviation of the k-th label dimension in the structured label matrix; i is the sample index, ranging from 1 to n, where n is the total number of candidate event samples; j is the text dimension index, ranging from 1 to d, where d is the dimension size of the text vector; k is the structured label dimension index, ranging from 1 to m, where m is the sum of the dimensions of the three one-hot encodings for event category, department, and level. The text semantic component matrix, which can be interpreted by the structured label matrix, is obtained based on latent principal component analysis; wherein, the expression of the text semantic component matrix is: ； In the formula, It is a semantic component matrix of text that can be interpreted by a structured tag matrix, with dimensions n×d; This is the text score matrix extracted using partial least squares, with dimensions n× ; The number of potential principal components is used to characterize the core information in the text related to the structured tag matrix; The text payload matrix is ​​extracted based on partial least squares and has a dimension of d×k. It is used to represent the weight mapping relationship of each text dimension on the potential principal components. This is the transpose of the text payload matrix; Subtracting the text semantic component matrix from the standard event text vector matrix yields the pure semantic residual matrix; wherein the expression for the pure semantic residual matrix is: ； In the formula, It is a purely semantic residual matrix; The standard event text vector matrix has dimensions n×d; The pure semantic residual matrix is ​​denormalized to restore it to the original numerical scale, resulting in a decoupled text vector matrix. The text similarity is updated based on the decoupled text vector to generate the decoupled text similarity between the query event and each candidate event.

6. The adverse event retrieval method based on multi-factor fusion as described in claim 1, characterized in that, The process employs pre-defined adaptive weighting to weight and fuse the category matching score, department matching score, grade proximity score, time decay score, and decoupled text similarity to generate a comprehensive score between the query event and each candidate event. Specifically, this includes: Based on the major category of the query event, the weights of each factor under the corresponding category are called; wherein, the weights are configured based on Dirichlet prior weights, including categories such as diagnosis and treatment operations, nursing processes, medicines and equipment, and public facilities; Based on a preset multi-factor fusion scoring function, the category matching score, the department matching score, the grade proximity score, the time decay score, and the decoupled text similarity are called to generate a comprehensive score between the query event and each candidate event. The expression of the multi-factor fusion scoring function is as follows: ； In the formula, For comprehensive scoring; The weights are respectively assigned to the text factor, category factor, department factor, grade factor, and time factor, and satisfy the following conditions: ; To decouple text similarity; Scoring based on category matching; Departmental matching score; For rating matching score; For time decay; Specifically, when the event field of the query event indicates a missing value, the available factors are dynamically normalized to generate a corrected comprehensive score. The expression for the dynamic normalization is: ； In the formula, This is the revised overall score; To query the set of factors currently available for the event; The weight of the m-th available factor; The matching score for the m-th available factor; This is the dimension index for the fusion factor, with values ​​ranging from 1 to 5, corresponding to the text factor, category factor, department factor, grade factor, and time factor, respectively.

7. The adverse event retrieval method based on multi-factor fusion as described in claim 1, characterized in that, The process of sorting candidate events in the adverse event case set according to the comprehensive score, selecting a predetermined number of top-ranked candidate events to generate a similar event retrieval recommendation list for the query event, and generating retrieval recommendation information based on the comprehensive score of each candidate event specifically includes: Based on the comprehensive score, all candidate events in the adverse event case set are sorted in descending order, and a predetermined number of the top-ranked candidate events are selected to generate a similar event retrieval recommendation list; wherein, the expression for the similar event retrieval recommendation list is: ； In the formula, This outputs a list of similar event recommendations containing the top K similar candidate events for a given query event q. To retrieve the first K sorted results; A collection of adverse event cases; The events are sorted in descending order of their overall score, meaning that events with higher matching scores are ranked higher. For comprehensive scoring; Based on a predefined recommendation explanation vector, a quantitative explanation for each factor dimension in the comprehensive score is generated; wherein, the expression for the recommendation explanation vector is: ； In the formula, Recommendation explanation vector; Weighted contribution score to the semantic dimension of the text; Weighted contribution score for event category dimension; Weighted contribution score for the departmental dimension of the event; Weighted contribution score for the event level dimension; Weighted contribution score for the time dimension; Based on the weighted contribution score of the text semantic dimension, the core similarity segments between each candidate event in the similar event retrieval recommendation list and the query event are determined; Natural language recommendation reasons for candidate events are generated based on the scores of each sub-item in the comprehensive score and the core similar segments. Based on the itemized quantitative explanation, the core similar segments, and the natural language recommendation reasons, retrieval recommendation information for candidate events is generated.

8. A multi-factor fusion-based adverse event retrieval system, characterized in that, include: The module includes a similarity calculation module, a dimension matching calculation module, a similarity update module, a comprehensive score calculation module, and a search result generation module. The similarity calculation module is used to calculate the text similarity between the query event and each candidate event in the adverse event case set; wherein, the adverse event case set contains several structured adverse event samples, and each adverse event sample retains at least the field information of event category, event process, event analysis, medical safety adverse event level and event department; The dimension matching calculation module is used to calculate the category matching score, department matching score, grade proximity score, and time decay score between the query event and the candidate event based on the event field representation; wherein, the event field representation is constructed based on the field information and event time; The similarity update module is used to calculate the decoupled text vectors of the query event and each candidate event in the adverse event case set by combining the structured label matrix, and update the text similarity to the decoupled text similarity based on the decoupled text vectors; wherein, the structured label matrix is ​​generated based on the encoding results of the event category, the event department and the event level; The comprehensive scoring calculation module is used to perform weighted fusion of the category matching score, the department matching score, the grade proximity score, the time decay score, and the decoupled text similarity using pre-set factor adaptive weights to generate a comprehensive score between the query event and each candidate event; wherein, the types of factors include category factors, department factors, grade factors, time factors, and text factors; The expression for the decoupled text vector is: ； In the formula, To decouple the value of the i-th adverse event sample in the text vector matrix in the j-th text dimension; Let be the value of the i-th adverse event sample in the j-th text dimension of the standardized residual matrix; Let X be the sample standard deviation of the j-th dimension in the event text vector matrix X; Let be the sample mean of the j-th dimension in the event text vector matrix X; The expression for the decoupled text similarity is: ； In the formula, To decouple text similarity; Custom weighting coefficients; Cosine similarity is used to decouple pure semantic vectors; This is the first global semantic vector after decoupling and correction; For the first The second global semantic vector of candidate events after decoupling and correction; To query events; For the first One candidate event; This refers to the degree of overlap between entities. The search result generation module is used to sort each candidate event in the adverse event case set according to the comprehensive score, select a preset number of candidate events with the highest ranking to generate a similar event search recommendation list for the query event, and generate search recommendation information based on the comprehensive score of each candidate event.

9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the adverse event retrieval method based on multi-factor fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the adverse event retrieval method based on multi-factor fusion as described in any one of claims 1 to 7.

Images