Search-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices

Abstract

This invention discloses a retrieval-enhanced generative evidence-based question-and-answer and decision support system for innovative drugs and medical devices, comprising an evidence collection module, a knowledge graph module, a retrieval enhancement module, a question-and-answer generation module, a security verification module, and an evidence update module. The evidence collection module collects drug and medical device data and automatically labels it with evidence-based level tags; the knowledge graph module constructs a dynamic evidence-based knowledge graph with timeliness attributes; the retrieval enhancement module performs a three-level hybrid retrieval, generating retrieval-enhanced context with evidence-based level scores; the question-and-answer generation module integrates context and a large language model to generate responses with evidence-based level labels; the security verification module performs regulatory status comparison, contraindication matching, and off-label use risk assessment, triggering risk warnings when anomalies occur; the evidence update module monitors timeliness attributes and automatically triggers response regeneration. This system solves the problems of scattered drug and medical device evidence, rigid knowledge representation, and lack of security verification in output, improving the reliability and timeliness of evidence-based question-and-answer.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to enhanced generative evidence-based question answering and decision support for innovative drugs and medical devices. Background Technology

[0002] With the accelerated research and approval process for innovative drugs and medical devices and the expansion of clinical applications, medical institutions and clinical decision-makers have an increasing demand for real-time access to and accurate interpretation of drug and medical device evidence. Currently, some technologies are attempting to provide intelligent support for medical question answering by combining knowledge graph construction with large language model generation.

[0003] However, these technologies have significant limitations: First, existing systems are designed for general medical scenarios and have not been specifically adapted to the characteristics of incomplete evidence systems and dynamically changing regulatory status for innovative drugs and medical devices, resulting in a disconnect between search results and actual clinical needs; Second, existing knowledge representation methods are static and lack dynamic quantification mechanisms for evidence timeliness and evidence-based levels, failing to reflect the objective laws of rapid iteration of evidence for innovative drugs and medical devices; Third, existing generative outputs have a single security verification level and fail to achieve three-dimensional cross-validation of real-time comparison of regulatory status, matching of individual contraindications, and off-label risk assessment, making it difficult to guarantee the reliability of outputs in high-risk decision-making scenarios.

[0004] Therefore, there is an urgent need for a drug and medical device-specific decision support solution that can deeply integrate evidence-based medicine level labeling, evidence timeliness-driven updates, and multi-dimensional safety verification. Summary of the Invention

[0005] The main objective of this invention is to provide a retrieval-enhanced generative evidence-based question answering and decision support system for innovative pharmaceuticals and medical devices, in order to solve the technical problems of scattered and disordered evidence, static and rigid knowledge representation, and lack of security verification in the existing technology.

[0006] To achieve the above objectives, this invention provides a retrieval-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices, characterized by comprising an evidence collection module, a knowledge graph module, a retrieval enhancement module, a question answering generation module, a security verification module, and an evidence updating module, wherein: The evidence collection module is used to connect to drug and medical device data sources, collect relevant data on innovative drugs and medical devices in real time, automatically label them with evidence-based level tags, and generate original evidence streams with time stamps. The knowledge graph module is used to receive the original evidence stream, extract the drug and medical device entities and regulatory approval status, establish evidence strength and timeliness attributes based on the evidence-based level tags, and construct a dynamically updated drug and medical device evidence-based knowledge graph. The retrieval enhancement module is used to receive user query requests, perform a three-level hybrid retrieval based on the drug and medical device evidence-based knowledge graph, and generate a retrieval enhancement context containing evidence-based level scores and time stamps. The question-and-answer generation module is used to integrate the search enhancement context with the large language model to generate evidence-based question-and-answer responses with literature citations and evidence-based level annotations; The security verification module is used to perform real-time comparison of the regulatory status, contraindication matching, and off-label risk assessment of the evidence-based question and answer response. When an abnormality occurs, a graded risk warning is triggered and fed back to the question and answer generation module for regeneration. The evidence update module is used to monitor the timeliness attribute of entity nodes in the drug and medical device evidence-based knowledge graph. When an evidence update or a change in regulatory status is detected, the retrieval enhancement module is automatically triggered to update the context and drive the question-and-answer generation module to regenerate the response, forming a timeliness-driven evidence-based decision support closed loop.

[0007] Optionally, the evidence collection module includes an evidence-based level labeling unit, used for: The original evidence collected is graded and judged according to the clinical trial stage, study design type and sample size; The judgment results are encoded as evidence-based medicine level labels and attached to the corresponding data to generate a stream of original evidence with level labels.

[0008] Optionally, the knowledge graph module includes a timeliness attribute construction unit, used for: Receive the original evidence stream with graded identifiers and extract drug molecular structure, device technical parameters, indications, contraindications, adverse reactions and regulatory approval status entities; Based on the evidence-based medicine level labels, the evidence strength values ​​between entities are calculated, and the publication time and expiration threshold are bound to each entity node to construct a drug and medical device evidence-based knowledge graph with time-sensitive attributes.

[0009] Optionally, the retrieval enhancement module includes a three-level hybrid retrieval unit, used for: Based on the dual-tower encoder, the query and evidence documents are mapped to a shared embedding space to calculate semantic similarity and perform dense paragraph retrieval. Based on the aforementioned evidence-based knowledge graph of pharmaceuticals and medical devices, the search for related entities is expanded through multi-hop relationships, and a graph structure search is performed. Based on the clinical trial stage index screening, evidence types that match the target drug or medical device development stage are used to perform stage-constrained searches. The three levels of search results are integrated to generate a search enhancement context that includes evidence-based ratings and time stamps.

[0010] Optionally, the question-and-answer generation module includes an evidence-based constraint generation unit, used for: The evidence-based rating in the enhanced retrieval context is used as a constraint condition embedded in the prompt text. The driving language model outputs evidence-based question-and-answer responses with reference citations and evidence-based level annotations under the given constraints.

[0011] Optionally, the security verification module includes a supervisory status synchronization unit, used for: The latest approval status of innovative drugs and medical devices in the database of the Center for Drug Evaluation can be obtained in real time through the interface. The latest approval status is compared with the regulatory status referenced in the evidence-based Q&A response. If they are inconsistent, an interception is triggered and the evidence update module is driven to start the regeneration process.

[0012] Optionally, the security verification module includes a contraindication matching unit, used for: The physiological parameters and medical history information in the patient's individual characteristics are analyzed and matched with the contraindication evidence in the drug and medical device evidence-based knowledge graph; When a match is successful, a taboo warning is generated and fed back to the question-and-answer generation module, which then drives the module to regenerate a response that avoids the taboo.

[0013] Optionally, the security verification module includes an off-label assessment unit, used for: The clinical applications recommended in the evidence-based Q&A responses are compared with the approved indications in the drug and medical device evidence-based knowledge graph; When out-of-scope use is detected, a risk level identifier is generated, and a preset security response template is invoked or the question and answer generation module is driven to regenerate based on the risk level.

[0014] Optionally, the evidence update module includes a timeliness monitoring unit, used for: Periodically scan the publication time and expiration threshold of each entity node in the drug and medical device evidence-based knowledge graph; When key evidence is detected to exceed the expiration threshold or an updated version is found, the search enhancement module is automatically triggered to update the search enhancement context.

[0015] Optionally, the evidence update module includes a state change response unit, used for: Monitor events related to changes in clinical trial status, retraction records, and changes in regulatory approval status; When the above events are detected, the retrieval enhancement module is driven to re-execute the three-level hybrid retrieval, and the question-and-answer generation module is triggered to regenerate evidence-based question-and-answer responses based on the updated retrieval enhancement context, forming a time-driven evidence-based decision support closed loop.

[0016] The beneficial effects of this invention are as follows: The automatic labeling of evidence-based levels in the evidence collection module standardizes and quantifies the source of evidence for innovative drugs and medical devices; the dynamic construction of timeliness attributes in the knowledge graph module overcomes the technical shortcomings of static and fixed knowledge representation; the three-level hybrid retrieval in the retrieval enhancement module improves retrieval accuracy in complex drug and medical device query scenarios; the generation of evidence-based constraints in the question-and-answer generation module ensures the traceability and labelability of the output content's evidence; the multi-dimensional cross-validation in the security verification module establishes a three-dimensional security barrier encompassing regulation, contraindications, and indications; and the timeliness monitoring and status change response in the evidence update module form a closed-loop decision support mechanism driven by the timeliness of evidence. The collaborative work of these modules enhances the reliability, timeliness, and security of evidence-based question-and-answer for innovative drugs and medical devices. Attached Figure Description

[0017] Figure 1 A module diagram of a search-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices provided in an embodiment of the present invention; Figure 2 A three-level hybrid retrieval flowchart of a retrieval-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the invention.

[0020] The main solution of this invention is: a search-enhanced generative evidence-based question-answering and decision support system for innovative drugs and medical devices. This system connects to multiple drug and medical device data sources through an evidence collection module and automatically labels them with evidence-based level tags. It constructs a drug and medical device evidence-based knowledge graph with timeliness attributes. Based on a three-level hybrid search, it generates a search-enhanced context containing evidence-based level scores. It integrates a large language model to output evidence-based question-answering responses with literature citations and evidence-based level labels. The system is then output after multi-dimensional security checks, including real-time comparison of regulatory status, contraindication matching, and off-label use risk assessment. Finally, it forms a closed-loop decision support system driven by evidence timeliness through timeliness monitoring and status change response.

[0021] Because existing technologies are designed for general medical scenarios and are not specifically adapted to the characteristics of incomplete evidence systems and dynamically changing regulatory status of innovative drugs and medical devices, their knowledge representation is static and lacks a dynamic quantification mechanism for evidence-based levels. The generated output does not achieve three-dimensional cross-validation of real-time comparison of regulatory status and off-label risk assessment. The retrieval-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices provided by this invention can achieve dynamic management of drug and medical device evidence through automatic annotation of evidence-based levels and evidence timeliness-driven updates. It provides evidence-based decision-making basis with multi-dimensional security verification and transforms the assessment results into actionable clinical medication guidance, thereby meeting the needs of evidence reliability, timeliness and safety in clinical decision support for innovative drugs and medical devices.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can understand it.

[0023] Reference Figure 1 This invention provides a retrieval-enhanced generative evidence-based question answering and decision support system for innovative drugs and medical devices, comprising: an evidence collection module, a knowledge graph module, a retrieval enhancement module, a question answering generation module, a security verification module, and an evidence updating module.

[0024] Among them, innovative drugs and medical devices refer to drugs and medical devices that are in the clinical trial stage or early market launch stage and have not yet formed a complete clinical evidence system. Their evidence sources are characterized by being scattered, time-sensitive, and subject to dynamic changes in regulatory status. Evidence-based question and answer refers to generating response content with literature traceability based on verifiable clinical research evidence. Decision support refers to providing clinical decision-makers with recommendations on the use of drugs and medical devices with safety verification.

[0025] As a preferred embodiment of the present invention, in order to provide accurate medical and regulatory compliance support for the automatic labeling of evidence-based levels, the closed-loop driven evidence timeliness, and the off-label risk assessment in this system, innovative drugs and medical devices are specifically limited to drugs and medical device products that are in Phase I to Phase III of clinical trials or have submitted a marketing invention but have not yet completed the approval of all indications, and preferably meet the requirement of having evidence accumulation of at least 500 follow-up data in real-world evidence collection terminals.

[0026] Specifically, the types of evidence sources, categorized by strength of evidence, include: randomized controlled trial evidence, real-world evidence, single-arm trial evidence, case series evidence, and expert opinion evidence. The weighting coefficients for each type of evidence are preset according to the GRADE standard. In actual clinical applications, the system is deployed in a cluster of computing devices containing a central processing unit (CPU) and a graphics processing unit (GPU) or a distributed cloud computing server cluster. The computing device cluster is equipped with a non-volatile storage array for long-term storage of data returned from the drug and medical device review database interface, clinical trial platform registration information, real-world evidence terminal collection records, enterprise R&D pipeline status data, and generated drug and medical device evidence-based knowledge graph data. The system achieves data interaction and command invocation between the evidence collection module, knowledge graph module, retrieval enhancement module, question-and-answer generation module, security verification module, and evidence update module through a high-speed data bus and memory sharing mechanism within the computing devices.

[0027] The evidence collection module connects to pharmaceutical and medical device data sources via standardized interfaces. These data sources specifically include the clinical trial database of the Center for Drug Evaluation, the medical device innovation approval channel, the clinical trial registration platform, real-world evidence collection terminals, and pharmaceutical and medical device company R&D pipeline databases. The evidence collection module is equipped with an evidence-based medicine level labeling unit. This unit classifies and determines the collected raw evidence based on clinical trial phase, study design type, and sample size, encodes the determination results as evidence-based medicine level labels, and attaches them to the corresponding data, generating a raw evidence stream with time stamps.

[0028] When performing the grading assessment, the evidence-based grading unit uses a weighted scoring function to calculate the evidence-based grading score for each piece of evidence. The formula is as follows: in, This represents the evidence-based score of the i-th piece of original evidence; The value represents the clinical trial stage coefficient of the i-th original evidence, with a value of 0.2 for Phase I, 0.4 for Phase II, 0.6 for Phase III, 0.8 for inventions submitted for market approval, and 1.0 for those that have completed approval for all indications. The coefficient represents the study design type coefficient for the i-th piece of original evidence. The coefficient is 1.0 for randomized controlled trials, 0.8 for real-world studies, 0.6 for single-arm trials, 0.4 for case series, and 0.2 for expert opinions. The normalized value of the sample size of the i-th original evidence is represented by the ratio of the actual sample size to the preset baseline sample size of 5000 cases, with an upper limit cutoff of 1.0; α, β and γ are the weight parameters of the clinical trial stage coefficient, the study design type coefficient and the normalized value of the sample size, respectively, satisfying α+β+γ=1. In this embodiment, the preferred values ​​are α=0.3, β=0.5 and γ=0.2.

[0029] The knowledge graph module is equipped with an entity extraction unit and a time-sensitive attribute construction unit. The entity extraction unit uses a named entity recognition model pre-trained on biomedical text to extract drug and medical device-related entities from the original evidence stream. Since the literature descriptions of innovative drugs and medical devices involve complex molecular structural terminology and device technical parameters, the entity extraction unit specifically selects a pre-trained model with biomedical adaptive capabilities and performs secondary optimization on the system's drug and medical device corpus to improve the accuracy of professional terminology recognition.

[0030] After the evidence collection module completes its calculations, it transmits the original evidence stream with time stamps to the knowledge graph module and synchronously writes the evidence-based level labels into the non-volatile storage array for subsequent retrieval.

[0031] The knowledge graph module receives raw evidence streams with expiration stamps from the evidence collection module. The knowledge graph module is configured with an entity extraction unit and an expiration attribute construction unit. The entity extraction unit uses a named entity recognition model pre-trained on biomedical text to extract entities related to drug molecular structure, device technical parameters, indications, contraindications, adverse reactions, clinical trial phases, and regulatory approval status from the raw evidence stream. The expiration attribute construction unit receives evidence-based medicine level labels output by the evidence-based medicine level labeling unit, calculates the evidence strength value between entities based on the evidence-based medicine level labels, and binds the publication time and expiration threshold to each entity node, constructing a drug and device evidence-based knowledge graph with expiration attributes.

[0032] The formula for calculating the inter-entity evidence strength value of the timeliness attribute building unit is as follows: in, This represents the strength of evidence value between entity u and entity v, where u and v are identifiers of different entity nodes in the knowledge graph; This represents the set of all original evidence that simultaneously relates entity u and entity v; The evidence-based score of the kth original evidence is calculated by the evidence-based score labeling unit; λ represents the time decay coefficient, which is preferably 0.001 per hour in this embodiment. The time interval between the release time of the kth piece of original evidence and the current time is expressed in hours; e is a natural constant.

[0033] The evidence-based knowledge graph of pharmaceuticals and medical devices is stored in a graph database, and the timeliness attribute index of entity nodes is synchronously transmitted to the retrieval enhancement module.

[0034] The retrieval enhancement module receives natural language query requests input by the user. (See reference...) Figure 2The search enhancement module is configured with an intent parsing submodule and a three-level hybrid search unit. The intent parsing submodule parses the innovative drug / device names, clinical application scenarios, and evidence type requirements in the query, generating a structured query representation. The three-level hybrid search unit performs a three-level hybrid search based on the drug / device evidence-based knowledge graph, generating a search enhancement context that includes evidence-based score and time stamp.

[0035] The first level of the three-level hybrid retrieval unit is dense paragraph retrieval, which uses a dual-tower encoder to map queries and evidence documents to a shared embedding space to calculate semantic similarity. The query-side encoder of the dual-tower encoder converts the user query text into a query vector, and the document-side encoder converts the evidence documents into document vectors. The matching score is calculated using cosine similarity. The formula is as follows: in, represents the semantic similarity score for dense paragraph retrieval; q represents the query vector, output by the query-side encoder; d represents the document vector, output by the document-side encoder. Indicates the L2 norm of the query vector; · represents the L2 norm of a document vector; · represents the vector dot product operation.

[0036] The second level of the three-level hybrid retrieval unit is graph structure retrieval, which expands the retrieval of related entities based on the evidence-based knowledge graph of pharmaceuticals and medical devices through multi-hop relationships. Graph structure retrieval starts from the entity nodes extracted by the intent parsing submodule, and performs a breadth-first traversal along the edges of treatment relationships, contraindication relationships, and interaction relationships in the knowledge graph. The traversal depth is set to 2 hops, and it returns all related entities and their relationship attributes on the traversal path.

[0037] The third level of the three-tiered hybrid search unit is the stage-constrained search, which filters evidence types that match the target drug / device development stage based on the clinical trial stage index. The stage-constrained search, based on the evidence type requirements parsed by the intent parsing submodule, filters the clinical trial stage coefficient T from the drug / device evidence-based knowledge graph. i Entity nodes that meet the threshold condition, where the threshold is dynamically set according to user needs, preferably T. i ≥0.4.

[0038] The formula for fusing the results of three levels of retrieval in a three-level hybrid retrieval unit is as follows: in, This represents the final retrieval score after fusion; Indicates the score for dense paragraph retrieval; The graph structure retrieval score is calculated as the normalized result of the sum of evidence strength values ​​on the path of related entities. The value represents the phased constraint retrieval score, which is 1.0 when the phased threshold is met, and 0 otherwise. w1, w2 and w3 are the weight parameters of the three-level retrieval, which satisfy w1+w2+w3=1. In this embodiment, the preferred values ​​are w1=0.5, w2=0.3 and w3=0.2.

[0039] The search enhancement module sorts the search results in descending order based on the final search score, and selects the top K evidence fragments to construct a search enhancement context. In this embodiment, the preferred value for K is 5. The search enhancement context includes the evidence source, evidence-based score, confidence score, and timestamp for each evidence fragment, where the confidence score is the final search score. .

[0040] After the retrieval enhancement module completes its calculations, it transmits the retrieval enhancement context to the question-answer generation module.

[0041] The question-and-answer generation module is connected to the retrieval enhancement module and the large language model fine-tuned for biomedical tasks via a data bus. The question-and-answer generation module receives the retrieval enhancement context output by the retrieval enhancement module. It is equipped with an evidence-based constraint generation unit, which embeds the evidence-based score from the retrieval enhancement context as a constraint into the prompt text, driving the large language model to output evidence-based question-and-answer responses with literature citations and evidence-based score annotations under these constraints.

[0042] The evidence-based constraint generation unit constructs the template structure for enhanced prompt text as follows: The task instruction segment clearly defines the model's role as a clinical drug and medical device decision support assistant, emphasizing the priority of medication safety; the validated evidence segment retrieves evidence fragments from the enhanced context in descending order of evidence level, with each evidence fragment labeled with its source identifier and evidence level; the user query segment references the original query text. The evidence-based constraint generation unit encodes the evidence level score as constraint weights, embedding them at the beginning of the validated evidence segment in the prompt text. The constraint weights are assigned to the evidence level score according to the GRADE standard. The mapping is divided into three levels: high, medium, and low, with corresponding constraint weights of 1.0, 0.6, and 0.3, respectively.

[0043] During the generation process, the large language model adjusts the citation priority of each evidence fragment according to the constraint weight, giving priority to citing high-level evidence with a constraint weight of 1.0, and automatically attaches an evidence level label at the end of the generated response, which includes the highest and lowest evidence level range of the evidence on which the response is based.

[0044] After the question-and-answer generation module completes its calculations, it transmits the evidence-based question-and-answer responses to the security verification module and simultaneously writes the list of cited evidence generated during the generation process into the non-volatile storage array.

[0045] The safety verification module is connected to the question-and-answer generation module, the knowledge graph module, and the database interface of the Center for Drug Evaluation via a data bus. The safety verification module receives evidence-based question-and-answer responses from the question-and-answer generation module. The safety verification module is equipped with a regulatory status synchronization unit, a contraindication matching unit, and an off-label indication assessment unit.

[0046] The regulatory status synchronization unit obtains the latest approval status of innovative drugs and medical devices from the Center for Drug Evaluation's database in real time via an interface. This unit compares the latest approval status with the regulatory status referenced in the evidence-based Q&A responses. The comparison is based on the consistency between the regulatory approval status attributes of entity nodes in the drug and medical device evidence-based knowledge graph and the status returned by the interface. When an inconsistency is detected, the regulatory status synchronization unit generates a status conflict signal and feeds it back to the Q&A generation module, driving it to regenerate the response.

[0047] The contraindication matching unit parses the physiological parameters and medical history information from the patient's individual characteristics and matches them with contraindication entities in the drug and medical device evidence-based knowledge graph. The matching process employs entity alignment technology, inputting the patient's individual characteristic text into the same named entity recognition model used during knowledge graph construction, extracting contraindication-related entities, and calculating semantic similarity with the contraindication entity nodes of the target drug or medical device. The formula is as follows: in, J represents the contraindication matching score; J represents the set of all contraindication entity nodes associated with the target drug and medical device. This represents the contraindication evidence vector extracted from the patient's individual characteristics; Let represent the j-th contraindication verification vector of the target drug / device; cos represents the cosine similarity function. When When the threshold of 0.85 is exceeded, the contraindication matching unit determines that the match is successful, generates a contraindication warning and sends it to the question and answer generation module, which then drives the module to regenerate the response content that avoids the contraindication.

[0048] The off-label assessment unit compares the clinical applications recommended in the evidence-based Q&A responses with the approved indications in the drug-device evidence-based knowledge graph. The off-label assessment unit calculates the comparison results using text similarity, with the following formula: in, Indicates the degree of deviation from the indication; This represents the text vector representing the clinical applications recommended in the evidence-based question and answer response; This represents the approved indication text vector. When... When the deviation exceeds a preset threshold of 0.5, the off-label use assessment unit determines that off-label use has occurred and generates a risk level label. The risk level is divided into three levels based on the degree of deviation: 0.5 ≤ <0.7 indicates low risk, 0.7≤ <0.9 indicates medium risk. A score of ≥0.9 indicates high risk. Based on the risk level, the off-label assessment unit will either call a preset safety response template or drive the question-and-answer generation module to regenerate the response.

[0049] After the security verification module completes its calculations, it outputs the evidence-based Q&A responses that pass the verification to the terminal display device, transmits the feedback signals that fail the verification to the evidence update module, and writes the verification log to the non-volatile storage array.

[0050] The evidence update module is connected to the knowledge graph module, retrieval enhancement module, and question-answer generation module via a data bus. The evidence update module is equipped with a timeliness monitoring unit and a status change response unit.

[0051] The timeliness monitoring unit periodically scans the publication time and expiration threshold of each entity node in the drug and medical device evidence-based knowledge graph. In this embodiment, the scanning period is preferably set to 3600 seconds. The timeliness monitoring unit calculates the remaining effective duration of each entity node using the following formula: in, Indicates the remaining valid duration; Indicates the failure threshold of an entity node; This indicates the time interval between the current time of the entity node's publication. When... When the timeliness monitoring unit determines that the entity node is invalid, it automatically triggers the retrieval enhancement module to update the retrieval enhancement context and removes the evidence fragment corresponding to the entity node from the retrieval results.

[0052] The status change response unit monitors events related to changes in clinical trial status, retraction records, and regulatory approval status. Monitoring sources include notifications of registration information changes on the clinical trial platform, updates to the academic journal retraction database, and approval status announcements from the Center for Drug Evaluation. When such events are detected, the status change response unit updates the attribute values ​​of the corresponding entity nodes in the drug-device evidence-based knowledge graph according to the event type, drives the retrieval enhancement module to re-execute the three-level hybrid retrieval, and simultaneously triggers the question-and-answer generation module to regenerate evidence-based question-and-answer responses based on the updated retrieval enhancement context.

[0053] After the evidence update module completes its calculations, it synchronously writes the updated drug and medical device evidence-based knowledge graph into a non-volatile storage array and feeds back the update event records to the evidence collection module, driving it to start a new round of data collection and evidence-based level labeling, thus forming a time-driven evidence-based decision support closed loop.

[0054] To further illustrate the execution logic and evidence-based decision-making process of the technical solution of this invention in actual clinical scenarios, we will now take an evidence-based question-and-answer application scenario of the innovative anti-tumor drug pembrolizumab combined with chemotherapy for first-line treatment of advanced non-small cell lung cancer as an example to specifically deduce the quantitative calculation and time-driven process of the above-mentioned search-enhanced generative evidence-based question-and-answer and decision support system.

[0055] Clinicians input a query request through the terminal interface: "The latest evidence-based data and safety precautions regarding pembrolizumab in combination with pemetrexed for first-line treatment of EGFR mutation-negative advanced non-small cell lung cancer." The search enhancement module receives this query request, and the intent parsing submodule extracts the innovative drug / device name as pembrolizumab, the clinical application scenario as first-line treatment of advanced non-small cell lung cancer, and the required evidence type as randomized controlled trial evidence.

[0056] The search enhancement module performs a three-level hybrid search based on a drug and medical device evidence-based knowledge graph. Dense paragraph retrieval uses a dual-tower encoder to calculate the semantic similarity between the query and the evidence documents. Graph structure retrieval expands along treatment relationship edges to retrieve indications and contraindications associated with pembrolizumab. Stage constraint retrieval filters clinical trial stage coefficients that satisfy T... i Entity nodes with a score ≥ 0.4. Assuming a dense paragraph retrieval score. The graph structure retrieval score is 0.82. The staged constraint retrieval score is 0.75. The value is 1.0. Substituting these values ​​into the three-level hybrid retrieval fusion formula, with weight parameters w1=0.5, w2=0.3, and w3=0.2, the final retrieval score is calculated. =0.5×0.82+0.3×0.75+0.2×1.0=0.845.

[0057] The search enhancement module sorts evidence fragments in descending order based on the final search score and selects the top 5 to construct the search enhancement context. Assuming the top-ranked evidence fragment is the result of the randomized controlled trial KEYNOTE-189, with an evidence-based grade score of G... i The evidence-based rating unit (EBRT) is calculated to be 0.92, and the time interval between the release time and the current time is [not specified]. The time decay is 876 hours, and the time decay coefficient λ is 0.001 per hour. Substituting these values ​​into the evidence strength calculation formula, the evidence strength value corresponding to this evidence fragment is... =0.92× ≈0.92×0.416≈0.383. The confidence score of this evidence fragment is the final retrieval score of 0.845, and the timestamp is the current system time.

[0058] The question-and-answer generation module integrates the retrieved enhanced context with the large language model. The evidence-based constraint generation unit encodes an evidence-based score of 0.92 as a constraint weight of 1.0, embedding the start position of the validated evidence segment in the prompt text. Under the constraints, the large language model generates an evidence-based question-and-answer response, including data on the objective response rate, median progression-free survival, and median overall survival of pembrolizumab combined with pemetrexed, and labels the evidence-based score as "high," citing the original paper of the KEYNOTE-189 trial published in the *New England Journal of Medicine*.

[0059] The security verification module performs multi-dimensional cross-validation on the evidence-based Q&A responses. The regulatory status synchronization unit obtains the latest approval status of pembrolizumab from the Center for Drug Evaluation's database in real time via an interface, comparing and showing that the indication has been approved, consistent with the regulatory status cited in the response. The contraindication matching unit analyzes individual patient characteristics, assuming a history of active autoimmune disease, and matches it with contraindication evidence entities in the drug-device evidence-based knowledge graph. Substituting these into the contraindication matching formula, the patient's contraindication evidence vector is obtained. Contraindications to pembrolizumab confirm body vector cosine similarity The calculated value is 0.91, which exceeds the preset threshold of 0.85. The contraindication matching unit determines that the match is successful, generates a contraindication warning, and sends it back to the question and answer generation module.

[0060] After receiving the contraindication warning, the question and answer generation module regenerates the response, adding a contraindication warning to the original evidence-based question and answer content: "This patient has a history of active autoimmune disease, which falls under the category of contraindications for the use of pembrolizumab. It is recommended to evaluate alternative treatment options."

[0061] The off-label assessment unit compares the recommended clinical application in the regenerated response with the approved indications. The response recommends first-line treatment for EGFR mutation-negative advanced non-small cell lung cancer, and this application is included in the approved indications. Substituting these into the off-label deviation calculation formula... =1-1.0=0, which does not exceed the preset threshold of 0.5, indicating that there is no off-label use.

[0062] The evidence update module continuously monitors the timeliness attributes of each entity node in the drug and device evidence-based knowledge graph. Assuming the timeliness monitoring unit scans to the entity node corresponding to the KEYNOTE-189 trial, and this node was published on April 1, 2024, with an expiration threshold set at 8760 hours for randomized controlled trial evidence, and the current time is May 15, 2024, the time interval since publication is... It is 1056 hours. Substituting this into the formula for calculating the remaining valid time, =8760-1056=7704 hours If the value is >0, the entity node is still in a valid state.

[0063] Suppose that the state change response unit detects the release event of the updated version of the KEYNOTE-189 trial, KEYNOTE-189 long-term follow-up results, updates the attribute values ​​of the corresponding entity nodes in the drug-device evidence-based knowledge graph according to the event type, drives the retrieval enhancement module to re-execute the three-level hybrid retrieval, and triggers the question-answer generation module to regenerate evidence-based question-answer responses based on the updated retrieval enhancement context, incorporating the 5-year overall survival data from the long-term follow-up.

[0064] The evidence update module synchronously writes the updated drug and medical device evidence-based knowledge graph into the non-volatile storage array and feeds back the update event records to the evidence collection module, driving it to start a new round of data collection and evidence-based level labeling, forming a time-driven evidence-based decision support closed loop.

[0065] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores drug and medical device evidence-based knowledge graph data, raw evidence stream data, and retrieval enhancement context data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external drug and medical device data sources via a network connection. When executed by the processor, the computer program implements a retrieval enhancement generative evidence-based question answering and decision support method for innovative drugs and medical devices.

[0066] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0067] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0068] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0069] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0072] The databases involved in the various embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the various embodiments provided by this invention may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A search-enhanced generative evidence-based question-answering and decision support system for innovative drugs and medical devices, characterized in that: It includes an evidence collection module, a knowledge graph module, a retrieval enhancement module, a question-and-answer generation module, a security verification module, and an evidence update module, among which: The evidence collection module is used to connect to drug and medical device data sources, collect relevant data on innovative drugs and medical devices in real time, automatically label them with evidence-based level tags, and generate original evidence streams with time stamps. The knowledge graph module is used to receive the original evidence stream, extract the drug and medical device entities and regulatory approval status, establish evidence strength and timeliness attributes based on the evidence-based level tags, and construct a dynamically updated drug and medical device evidence-based knowledge graph. The retrieval enhancement module is used to receive user query requests, perform a three-level hybrid retrieval based on the drug and medical device evidence-based knowledge graph, and generate a retrieval enhancement context containing evidence-based level scores and time stamps. The question-and-answer generation module is used to integrate the search enhancement context with the large language model to generate evidence-based question-and-answer responses with literature citations and evidence-based level annotations; The security verification module is used to perform real-time comparison of the regulatory status, contraindication matching, and off-label risk assessment of the evidence-based question and answer response. When an abnormality occurs, a graded risk warning is triggered and fed back to the question and answer generation module for regeneration. The evidence update module is used to monitor the timeliness attribute of entity nodes in the drug and medical device evidence-based knowledge graph. When an evidence update or a change in regulatory status is detected, the retrieval enhancement module is automatically triggered to update the context and drive the question-and-answer generation module to regenerate the response, forming a timeliness-driven evidence-based decision support closed loop.

2. The system according to claim 1, characterized in that, The evidence collection module includes an evidence-based level labeling unit, used for: The original evidence collected is graded and judged according to the clinical trial stage, study design type and sample size; The judgment results are encoded as evidence-based medicine level labels and attached to the corresponding data to generate a stream of original evidence with level labels.

3. The system according to claim 2, characterized in that, The knowledge graph module includes a timeliness attribute construction unit, used for: Receive the original evidence stream with graded identifiers and extract drug molecular structure, device technical parameters, indications, contraindications, adverse reactions and regulatory approval status entities; Based on the evidence-based medicine level labels, the evidence strength values ​​between entities are calculated, and the publication time and expiration threshold are bound to each entity node to construct a drug and medical device evidence-based knowledge graph with time-sensitive attributes.

4. The system according to claim 3, characterized in that, The retrieval enhancement module includes a three-level hybrid retrieval unit, used for: Based on the dual-tower encoder, the query and evidence documents are mapped to a shared embedding space to calculate semantic similarity and perform dense paragraph retrieval. Based on the aforementioned evidence-based knowledge graph of pharmaceuticals and medical devices, the search for related entities is expanded through multi-hop relationships, and a graph structure search is performed. Based on the clinical trial stage index screening, evidence types that match the target drug or medical device development stage are used to perform stage-constrained searches. The three levels of search results are integrated to generate a search enhancement context that includes evidence-based ratings and time stamps.

5. The system according to claim 4, characterized in that, The question-and-answer generation module includes an evidence-based constraint generation unit, used for: The evidence-based rating in the enhanced retrieval context is used as a constraint condition embedded in the prompt text. The driving language model outputs evidence-based question-and-answer responses with reference citations and evidence-based level annotations under the given constraints.

6. The system according to claim 5, characterized in that, The security verification module includes a monitoring status synchronization unit, used for: The latest approval status of innovative drugs and medical devices in the database of the Center for Drug Evaluation can be obtained in real time through the interface. The latest approval status is compared with the regulatory status referenced in the evidence-based Q&A response. If they are inconsistent, an interception is triggered and the evidence update module is driven to start the regeneration process.

7. The system according to claim 6, characterized in that, The security verification module includes a contraindication matching unit, used for: The physiological parameters and medical history information in the patient's individual characteristics are analyzed and matched with the contraindication evidence in the drug and medical device evidence-based knowledge graph; When a match is successful, a taboo warning is generated and fed back to the question-and-answer generation module, which then drives the module to regenerate a response that avoids the taboo.

8. The system according to claim 7, characterized in that, The security verification module includes an off-label treatment assessment unit, used for: The clinical applications recommended in the evidence-based Q&A responses are compared with the approved indications in the drug and medical device evidence-based knowledge graph; When out-of-scope use is detected, a risk level identifier is generated, and a preset security response template is invoked or the question and answer generation module is driven to regenerate based on the risk level.

9. The system according to claim 8, characterized in that, The evidence update module includes a timeliness monitoring unit, used for: Periodically scan the publication time and expiration threshold of each entity node in the drug and medical device evidence-based knowledge graph; When key evidence is detected to exceed the expiration threshold or an updated version is found, the search enhancement module is automatically triggered to update the search enhancement context.

10. The system according to claim 9, characterized in that, The evidence update module includes a state change response unit, used for: Monitor events related to changes in clinical trial status, retraction records, and changes in regulatory approval status; When the above events are detected, the retrieval enhancement module is driven to re-execute the three-level hybrid retrieval, and the question-and-answer generation module is triggered to regenerate evidence-based question-and-answer responses based on the updated retrieval enhancement context, forming a time-driven evidence-based decision support closed loop.

Images