Medical question and answer enhanced generation method and system based on dynamic weighted multi-knowledge base

Abstract

This invention relates to a method and system for enhancing medical question-answering based on a dynamically weighted multi-knowledge base, belonging to the field of medical information processing technology. It includes the following steps: constructing a multi-source medical knowledge base and preprocessing it to obtain standardized data tuples; constructing an adaptive timeliness weight calculation module based on half-life theory to calculate the timeliness weight of knowledge fragments; constructing a multi-dimensional authority weight calculation module based on AGREE II and bibliometrics to calculate authority weights; constructing a dynamic weighted fusion module based on a nonlinear product model to calculate the comprehensive retrieval score of knowledge fragments; and constructing a semantic-timeliness dual conflict detection and weighted voting resolution module to calculate the semantic similarity coefficient of the top k knowledge fragments, and calculate the credibility score through the semantic similarity coefficient and the comprehensive retrieval score, thereby obtaining a set of winning and outdated knowledge fragments. This invention can improve the interpretability of clinical decision support.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing technology, specifically relating to a method and system for enhancing medical question-answering generation based on dynamic weighted multi-knowledge bases. Background Technology

[0002] Medical knowledge is rapidly evolving, and traditional medical question-and-answer systems have failed to effectively address the issue of knowledge timeliness. According to recent research, the half-life of medical knowledge is approximately five years, meaning that about 50% of current medical knowledge will be updated or overturned within five years. However, most existing medical question-and-answer systems still employ static knowledge bases and fixed-weight knowledge fusion strategies. When doctors or patients ask questions requiring the latest medical evidence, these systems may provide outdated or even harmful advice.

[0003] Existing technologies suffer from the following problems: knowledge base updates are lagging, failing to reflect the latest medical guidelines and research progress in real time; knowledge integration lacks timeliness considerations, often assigning equal weight to medical knowledge from different sources and of different types; and conflict resolution mechanisms are insufficient, failing to effectively resolve and provide reliable advice when retrieved knowledge has timeliness conflicts. There is an urgent need for a dynamic weighting method that can balance the authority and timeliness of knowledge, addressing the problem of traditional medical question-and-answer systems being "correct but outdated," and providing clinicians and patients with more reliable and timely medical advice. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] This invention provides a method for enhancing the generation of medical question answers based on dynamic weighted multi-knowledge bases, comprising the following steps:

[0006] S1. Construct a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples;

[0007] S2. Construct an adaptive timeliness weight calculation module based on half-life theory to calculate the timeliness weight of knowledge fragments;

[0008] S3. Construct a multidimensional authority weight calculation module based on AGREE II and bibliometrics to calculate the authority weight;

[0009] S4. Construct a dynamic weighted fusion module based on a nonlinear product model to calculate the comprehensive retrieval score of knowledge fragments;

[0010] S5. Construct a semantic-time dual conflict detection and weighted voting resolution module, calculate the semantic similarity coefficient of the top k knowledge fragments, and calculate the credibility score by combining the semantic similarity coefficient and the comprehensive retrieval score, thereby obtaining the set of winning knowledge fragments and outdated knowledge fragments.

[0011] Furthermore, the multi-source medical knowledge base includes a clinical guideline database, a scientific literature database, a drug database, and a multimodal image database.

[0012] Furthermore, unstructured knowledge fragments from the multi-source medical knowledge base are preprocessed and vectorized for storage, generating a quintuple for each unstructured knowledge fragment. ,in, Represents text content; Vectorized representation; Indicates the publication timestamp; The source type label indicates four sources: clinical guideline database, scientific literature database, drug database, and multimodal image database. This represents the original score indicating static authority.

[0013] (Static Authority Raw Score) refers to a raw credibility score pre-calculated based on the knowledge source when the data is stored in the knowledge base (i.e., before the user asks a question). It is called "static" and "pre-calculated" because it is only related to the quality of the knowledge source itself and does not change with the time or content of the user's question.

[0014] Furthermore, considering the characteristic that the value of medical knowledge decreases over time, an exponential decay calculation model based on time difference is established, expressed by the following formula:

[0015] ,

[0016] in, This represents the timeliness weight of a knowledge fragment, with a value range of (0,1]. The closer the value is to 1, the stronger the timeliness. The time decay rate of the source type to which the knowledge fragment belongs determines the steepness of the weight decrease; Indicates the latest current year;

[0017] To address the issue of weight distortion, this invention first introduces the exponential decay model from bibliometrics: To determine the decay rate of different knowledge types ;in, Indicates the time period of knowledge fragments The surplus value, This represents the initial value; based on this model, the concept of "knowledge half-life" is introduced, which is the time required for the value of knowledge to decay to half of its initial value; the formula for calculating the time-depreciation rate of the source type of a knowledge fragment is:

[0018] ,

[0019] in, The average half-life of knowledge types is used; medical guidelines are updated approximately every 5-6 years (according to WHO, ESC, and AHA guidelines), with a corresponding decay rate λ=ln2 / 5≈0.1386; research hotspots in medical literature change approximately every 3 years (based on PubMed citation decay analysis), with a corresponding decay rate λ=ln2 / 3≈0.2310; drug instructions are updated approximately every year (FDA data), with a corresponding decay rate λ=ln2 / 1≈0.6931; for different disease domains, adaptive λ is allowed, calibrated through small-sample tasks.

[0020] To address the problem that traditional fixed-parameter models cannot cope with sudden bursts or stagnations of knowledge in the medical field, a dynamic self-learning parameter update mechanism is constructed. This invention does not stop at static parameters but introduces a feedback loop, enabling... It possesses self-evolution capabilities; update trigger condition: the system periodically monitors the proportion of outdated user-reported answers, and triggers an update when the monitored value shows an upward trend; the adaptive update formula for the timeliness decay rate of the source type of the knowledge fragment is:

[0021] ,

[0022] in, These represent the decay rates before and after the update, respectively. This indicates the difference between the current monitoring period's expiration rate and the previous period's rate; This represents the learning rate (empirically taken as 0.05-0.1), used to control the adjustment step size and prevent parameter oscillation; This represents the baseline aging rate constant; this mechanism enables the system to sense changes in the external environment.

[0023] Furthermore, a standardized and normalized model for authority weights is constructed. Since the original static authority scores from different sources have inconsistent dimensions, the authority weights are calculated using the Min-Max normalization algorithm, as shown in the following formula:

[0024] ,

[0025] in, Indicates authority weight; This represents the minimum original authority score among the currently retrieved candidate knowledge sets; This represents the maximum value of the original authority score in the currently retrieved candidate knowledge set.

[0026] Furthermore, to address the technical limitations of traditional linear weighting (additive model) in effectively handling extreme cases (such as "authoritative but extremely outdated" knowledge still achieving high total scores), a nonlinear fusion model based on geometric weighting is constructed, expressed by the following formula:

[0027] ,

[0028] in, This represents the comprehensive retrieval score of the knowledge fragment, with a value range of [0,1], and is used for subsequent sorting and filtering; The contribution index, representing the weight of authority, controls the system's sensitivity to the reliability of evidence sources; The contribution index representing the timeliness weight controls the system's sensitivity to information freshness; This represents the dot product operation.

[0029] Furthermore, based on the knowledge fragments retrieved from the user's question, the comprehensive retrieval scores of the knowledge fragments are sorted from high to low, and the semantic similarity coefficient of the top k knowledge fragments is calculated, as shown in the following formula:

[0030] ,

[0031] in, , This represents the vectorized representation of two different knowledge fragments retrieved; the knowledge fragments are vectorized using a pre-trained language model (such as DeBERTa-v3 or BioBERT); Represents the semantic similarity coefficient; Indicates the length of the vector;

[0032] Set a threshold for contradiction Iterate through the weighted segments of the first k knowledge segments. ,when When two knowledge fragments are determined to be semantically identical, a standard text splicing strategy is used to integrate the two knowledge fragments into a context; when... If two knowledge fragments are deemed to have a substantial conflict of clinical recommendations, a weighted vote is performed on the knowledge fragments, and the decision score is calculated. The formula is as follows:

[0033] ,

[0034] in, Representing knowledge fragments Credibility score; A collection of knowledge fragments that support the same semantic viewpoint; This represents the comprehensive retrieval score of the j-th knowledge fragment in the set; This indicates the direct correlation between the knowledge fragment and the user's question Q;

[0035] The process involves iterating through the knowledge segments that participate in the voting and have substantial conflicts with clinical recommendations, identifying the knowledge segment with the highest adjudication score, and selecting it as the winning knowledge segment. The formula is as follows:

[0036] ,

[0037] in, This represents a winning knowledge fragment; Represents a set of conflicting knowledge fragments; This represents a knowledge fragment within a set of conflicting knowledge fragments; This represents the variable maximization operator;

[0038] The knowledge segment with the highest adjudication score This serves as the sole factual basis for passing to subsequent generation modules.

[0039] For the conflict set except Other knowledge fragments Labeling evidence as "outdated" may have some authority or relevance, but it loses out in the overall vote on timeliness.

[0040] ,

[0041] in, This represents a set of outdated knowledge fragments.

[0042] This invention also provides a medical question-answering enhancement generation system based on a dynamically weighted multi-knowledge base, comprising the following steps:

[0043] Data processing module: used to build a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples;

[0044] Timeliness Weight Calculation Module: Used to calculate the timeliness weight of knowledge fragments;

[0045] Authority Weight Calculation Module: Used to calculate authority weight;

[0046] Dynamic weighted fusion module: used to calculate the comprehensive retrieval score of knowledge fragments;

[0047] The dual conflict detection and weighted voting resolution module is used to calculate the semantic similarity coefficient of the top k knowledge fragments. The credibility score is calculated by combining the semantic similarity coefficient and the comprehensive retrieval score, and then the set of winning knowledge fragments and outdated knowledge fragments is obtained.

[0048] The Medical Question Answering Enhancement Generation and Interpretability Output Module is used to build structured prompt word templates and generate medical advice text, dynamic weighted heatmaps, and traceable evidence chains.

[0049] The advantages of this invention are:

[0050] This invention introduces an exponential decay model based on half-life theory and a dynamic λ update mechanism, combined with a product fusion model, to automatically identify and significantly reduce the weight of outdated knowledge. It overcomes the technical bottleneck of traditional RAGs in handling complex clinical conflicts by introducing a "semantic-timeliness" dual trigger mechanism. Unlike traditional simple text splicing, weighted voting is only initiated when the semantic contradiction exceeds a threshold τ, accurately resolving the conflict between recommendations from "old, highly authoritative guidelines" and "the latest groundbreaking literature." It also addresses the lack of transparency and trust in medical AI decision-making by embedding a dynamic weighted heatmap in the final result, transforming the complex algorithmic decision-making process into a trust credential visible to doctors. Doctors can intuitively see whether the system's answer is based on "the latest evidence" or "classic guidelines," and which older viewpoints have been excluded, greatly improving the interpretability of clinical decision support. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0052] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0053] Figure 2 This is the dynamic weighted heatmap of the present invention;

[0054] Figure 3 This is the dynamic weighted heatmap of Embodiment 3 of the present invention;

[0055] Figure 4 This is the dynamic weighted heatmap of Embodiment 4 of the present invention;

[0056] Figure 5 This is the dynamic weighted heatmap of Embodiment 5 of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1

[0059] In this embodiment, as Figure 1 As shown, this invention provides a method for enhancing the generation of medical questions and answers based on a dynamically weighted multi-knowledge base, the specific steps of which include:

[0060] S1. Construct a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples;

[0061] Specifically, the multi-source medical knowledge base includes a clinical guideline base, a scientific literature base, a drug database, and a multimodal image database;

[0062] The clinical guideline library covers the latest treatment guidelines published by authoritative institutions;

[0063] The scientific literature database is connected in real time to databases such as PubMed and WanFang to obtain high-quality medical papers from the past 5-10 years.

[0064] The drug database synchronizes with drug instructions and black box warning information published by authoritative institutions.

[0065] The multimodal image library integrates standardized diagnostic imaging cases such as ultrasound and MRI, along with expert consensus reports.

[0066] Unstructured knowledge fragments from a multi-source medical knowledge base are preprocessed and vectorized for storage, generating a quintuple for each unstructured knowledge fragment. ,in, Represents text content; Vectorized representation; Indicates the publication timestamp; The source type label indicates four sources: clinical guideline database, scientific literature database, drug database, and multimodal image database. This represents the original score indicating static authority.

[0067] The static authority score refers to a pre-calculated credibility score based on the knowledge source, calculated when the data is stored in the knowledge base, i.e., before the user asks a question. It is only related to the quality of the knowledge source itself and does not change with the time or content of the user's question.

[0068] The calculation process for the static authority raw score is as follows:

[0069] Clinical guidelines: Quantified using the AGREE II (Appraisal of Guidelines for Research & Evaluation II) tool; evaluation dimensions include 23 items across 6 major areas: scope and purpose, participants, rigor of development, clarity of expression, applicability, and editorial independence;

[0070] ,

[0071] in, This represents the static authority score corresponding to knowledge fragments in clinical guidelines. Indicates the actual score; Indicates the minimum possible score; Indicates the maximum possible score;

[0072] The AGREE II (Appraisal of Guidelines for Research & Evaluation II) standard scoring algorithm is the international gold standard tool for evaluating the quality of clinical guidelines in the medical field. In the AGREE II evaluation system, multiple reviewers typically score 23 items in the guideline.

[0073] Maximum score: Assuming the guide is perfect and all reviewers give full marks to all entries in the field, 7 points is the highest total score that can be obtained in theory.

[0074] Calculation method:

[0075] Maximum possible score = 7 (full marks) × number of entries in the field × number of reviewers

[0076] Minimum possible score: The lowest total score that could be theoretically obtained if the guide is of extremely poor quality and all reviewers give the lowest score of 1 point to all items in this area.

[0077] Calculation method:

[0078] Minimum possible score = 1 (lowest score) × number of entries in the field × number of reviewers

[0079] Actual score: The sum of the actual scores given by all reviewers for all items in this area, which is the direct addition of the scores given by each expert.

[0080] Scientific literature:

[0081] ,

[0082] in, This represents the static authority score corresponding to knowledge fragments in scientific literature. Indicates the journal's impact factor; Indicates the number of citations;

[0083] Drug data:

[0084] ,

[0085] in, This represents the static authority score corresponding to a knowledge fragment related to drug data. This represents a preset constant, with a preset score set based on the publishing organization level. In one embodiment, this preset score is set to 10 points, which is the theoretical maximum score.

[0086] S2. Construct an adaptive timeliness weight calculation module based on half-life theory to calculate the timeliness weight of knowledge fragments;

[0087] Specifically, considering the characteristic that the value of medical knowledge decreases over time, an exponential decay calculation model based on time difference is established, and the formula is expressed as follows:

[0088] ,

[0089] in, This represents the timeliness weight of a knowledge fragment, with a value range of (0,1]. The closer the value is to 1, the stronger the timeliness. The time decay rate of the source type to which the knowledge fragment belongs determines the steepness of the weight decrease; Indicates the latest current year;

[0090] To address the issue of weight distortion, this invention first introduces the exponential decay model from bibliometrics: To determine the decay rate of different knowledge types ;in, Indicates the time period of knowledge fragments The surplus value, This represents the initial value; based on this model, the concept of "knowledge half-life" is introduced, which is the time required for the value of knowledge to decay to half of its initial value; the formula for calculating the time-depreciation rate of the source type of a knowledge fragment is:

[0091] ,

[0092] in, The average half-life of knowledge types is denoted as λ = ln2 / 5 ≈ 0.1386. Medical guidelines are updated approximately every 5-6 years, with a corresponding decay rate of λ = ln2 / 5 ≈ 0.1386. Research hotspots in medical literature change approximately every 3 years, with a corresponding decay rate of λ = ln2 / 3 ≈ 0.2310. Drug instructions are updated approximately every year, with a corresponding decay rate of λ = ln2 / 1 ≈ 0.6931. For different disease domains, adaptive λ is allowed, calibrated through small-sample tasks.

[0093] To address the problem that traditional fixed-parameter models cannot cope with sudden bursts or stagnations of knowledge in the medical field, a dynamic self-learning parameter update mechanism is constructed. This invention does not stop at static parameters but introduces a feedback loop, enabling... It possesses self-evolution capabilities; update trigger condition: the system periodically monitors the proportion of outdated user-reported answers, and triggers an update when the monitored value shows an upward trend; the adaptive update formula for the timeliness decay rate of the source type of the knowledge fragment is:

[0094] ,

[0095] in, These represent the decay rates before and after the update, respectively. This indicates the difference between the current monitoring period's expiration rate and the previous period's rate; This represents the learning rate, which is empirically set to 0.05-0.1. It is used to control the adjustment step size and prevent parameter oscillation. This represents the baseline aging rate constant; this mechanism enables the system to sense changes in the external environment.

[0096] S3. Construct a multidimensional authority weight calculation module based on AGREE II and bibliometrics to calculate the authority weight;

[0097] Specifically, since the original static authority scores from different sources have inconsistent dimensions, the authority weights are calculated using a Min-Max normalization algorithm, as shown in the following formula:

[0098] ,

[0099] in, Indicates authority weight; This represents the minimum original authority score among the currently retrieved candidate knowledge sets; This represents the maximum original authority score among the currently retrieved candidate knowledge sets;

[0100] : Classified as Tier 1, the gold standard, such as high-quality guidelines or FDA black box warnings;

[0101] The criteria for classifying a candidate as Tier 2 are strong, supported by highly cited literature.

[0102] It is classified as Tier 3, based on the relevant evidence.

[0103] S4. Construct a dynamic weighted fusion module based on a nonlinear product model to calculate the comprehensive retrieval score of knowledge fragments;

[0104] Specifically, to address the technical limitation of traditional linear weighting in effectively handling extreme cases, such as authoritative but extremely outdated knowledge still achieving high total scores, a nonlinear fusion model based on geometric weighting is constructed, expressed by the following formula:

[0105] ,

[0106] in, This represents the comprehensive retrieval score of the knowledge fragment, with a value range of [0, 1], and is used for subsequent sorting and filtering; The contribution index, representing the weight of authority, controls the system's sensitivity to the reliability of evidence sources; The contribution index representing the timeliness weight controls the system's sensitivity to information freshness; This represents the dot product operation.

[0107] S5. Construct a semantic-time dual conflict detection and weighted voting resolution module, calculate the semantic similarity coefficient of the top k knowledge fragments, and calculate the credibility score by combining the semantic similarity coefficient and the comprehensive retrieval score, thereby obtaining the set of winning knowledge fragments and outdated knowledge fragments.

[0108] Specifically, based on the knowledge fragments retrieved from the user's question, the comprehensive retrieval scores of the knowledge fragments are sorted from high to low, and the semantic similarity coefficient of the top k knowledge fragments is calculated. The formula is as follows:

[0109] ,

[0110] in, , This represents the vectorized representation of two different knowledge fragments retrieved; the knowledge fragments are vectorized after being processed by a pre-trained language model. Represents the semantic similarity coefficient; The length of the vector is indicated; the pre-trained language model is either a DeBERTa-v3 model or a BioBERT model.

[0111] Set a threshold for contradiction Iterate through the weighted segments of the first k knowledge segments. ,when When two knowledge fragments are determined to be semantically identical, a standard text splicing strategy is used to integrate the two knowledge fragments into a context; when... If two knowledge fragments are deemed to have a substantial conflict of clinical recommendations, a weighted vote is performed on the knowledge fragments, and the decision score is calculated. The formula is as follows:

[0112] ,

[0113] in, Representing knowledge fragments Credibility score; A collection of knowledge fragments that support the same semantic viewpoint; This represents the comprehensive retrieval score of the j-th knowledge fragment in the set; This indicates the direct correlation between the knowledge fragment and the user's question Q;

[0114] The process involves iterating through the knowledge segments that participate in the voting and have substantial conflicts with clinical recommendations, identifying the knowledge segment with the highest adjudication score, and selecting it as the winning knowledge segment. The formula is as follows:

[0115] ,

[0116] in, This represents a winning knowledge fragment; Represents a set of conflicting knowledge fragments; This represents a knowledge fragment within a set of conflicting knowledge fragments; This represents the variable maximization operator;

[0117] The knowledge segment with the highest adjudication score This serves as the sole factual basis for passing to subsequent generation modules.

[0118] For the conflict set except Other knowledge fragments Labeling evidence as "outdated" may give it some authority or relevance, but it loses out in the overall vote on timeliness.

[0119] ,

[0120] in, This represents a set of outdated knowledge fragments.

[0121] Example 2

[0122] This embodiment provides a medical question-answering enhancement generation system based on a dynamically weighted multi-knowledge base, which executes the aforementioned medical question-answering enhancement generation method based on a dynamically weighted multi-knowledge base, including:

[0123] Data processing module: used to build a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples;

[0124] Timeliness Weight Calculation Module: Used to calculate the timeliness weight of knowledge fragments;

[0125] Authority Weight Calculation Module: Used to calculate authority weight;

[0126] Dynamic weighted fusion module: used to calculate the comprehensive retrieval score of knowledge fragments;

[0127] The dual conflict detection and weighted voting resolution module is used to calculate the semantic similarity coefficient of the top k knowledge fragments. The credibility score is calculated by combining the semantic similarity coefficient and the comprehensive retrieval score, and then the set of winning knowledge fragments and outdated knowledge fragments is obtained.

[0128] The Medical Question Answering Enhancement Generation and Interpretability Output Module is used to build structured prompt word templates and generate medical advice text, dynamic weighted heatmaps, and traceable evidence chains.

[0129] Specifically,

[0130] 1. Construction of enhanced prompts based on the ruling result: The system receives the winning knowledge fragments from the output. And its conflict markers. To suppress the illusion of large models and force them to follow the system's decision, a structured cue word template is constructed:

[0131] Instruction layer: Embedded "time-sensitive constraint instructions".

[0132] Prompt example: "You are an evidence-based medicine assistant. Please base your arguments strictly on the following winning evidence (posted in...") Year, comprehensive weight Answering user questions. Note: The system has detected outdated viewpoints that conflict with this (posted on...). (Year), please clearly state in your answer that this old viewpoint has been revised by the latest evidence.

[0133] Context layer: Fill in only Text content that is deemed outdated must not be used as factual evidence input into the model, thus cutting off interference from outdated information at the source.

[0134] 2. Large-Scale Model Inference and Generation: The constructed prompts are input into the large-scale language model (LLM). Based on the contextual information of the input, the large-scale model performs logical reasoning and text reconstruction to generate natural language text that conforms to medical terminology standards. During this process, the model is restricted to using only retrieved high-weight knowledge as factual basis, ensuring that although the generated answer is generated by AI, its core arguments are entirely controlled by high-weight medical evidence.

[0135] 3. Multidimensional Interpretable Result Encapsulation System: Aligns the generated text with intermediate computational data to produce a final product containing three dimensions:

[0136] (a) Medical advice text: intuitive natural language response;

[0137] (b) Dynamic weighted heatmap (core visualization feature): calculated by system call. and Draw a two-dimensional coordinate heatmap:

[0138] Horizontal axis (X-Axis): Represents the weight of timeliness. The more recent the date, the further to the right it is;

[0139] Vertical axis (Y-Axis): Represents authority weight. The more authoritative the source, the higher it is considered;

[0140] Node rendering: Map all candidate fragments retrieved in this search to nodes on the graph.

[0141] Winner node: Highlighted, usually located in the upper right or lower right area;

[0142] Loser nodes: Displayed in gray or warning color, visually indicating that although they may have high authority and are located at the top, they are located on the left due to their extremely low timeliness and are eliminated through a voting mechanism.

[0143] In one embodiment, the dynamic weighted heatmap is as follows: Figure 2 As shown.

[0144] (c) Chain of evidence: List the supporting evidence The original document title, DOI link, and publication year.

[0145] Example 3

[0146] This embodiment demonstrates the application of the TimeRAG method of the present invention in emergency triage. Taking acute heart failure triage as an example, the system needs to retrieve the latest guidelines to provide accurate triage recommendations.

[0147] Knowledge base construction: The system integrates the 2024 version of the NCCN acute heart failure triage guidelines, relevant PubMed literature published from 2022 to 2025, and the FDA drug labeling updated in 2023. Each knowledge fragment includes a timestamp. and source type .

[0148] Timeliness weight calculation This embodiment is based on the exponential decay model in bibliometrics. The attenuation rate is determined by the following formula: .

[0149] NCCN Guidelines =Guide:

[0150] Basis for setting: Based on updated statistics from WHO guidelines, the half-life is set. Year;

[0151] Attenuation rate calculation: ;

[0152] Weight calculation: .

[0153] PubMed literature, =References:

[0154] Setting basis: The half-life is set based on PubMed reference decay analysis. Year;

[0155] Attenuation rate calculation: ;

[0156] Taking 2023 literature as an example, the weight calculation is as follows: .

[0157] FDA drug labeling =Medication:

[0158] Basis for setting: The half-life is set based on the FDA's drug label revision cycle. Year;

[0159] Attenuation rate calculation: ;

[0160] Weight calculation: .

[0161] Authority weight calculation: This system adopts a heterogeneous scoring normalization mechanism to normalize the raw scores from different sources. Mapped to the [0,1] interval. Assuming the current search pool contains... =10, =0.

[0162] NCCN Guidelines:

[0163] The AGREE II standard was used for scoring.

[0164] Raw score calculation: (Scope 7 + Personnel 6 + Rigorousness 7 + Clarity 7 + Application 6 + Independence 7) / (6 × 7) × 10 = 9.52 points;

[0165] Normalized weights: .

[0166] PubMed literature:

[0167] The evaluation is based on a combination of impact factor (IF) and citation count.

[0168] Raw score calculation: Assume the article was published in The Lancet (IF=168) and has been cited 20 times. =5.1 + 3.0 = 8.1 points;

[0169] Normalized weights: .

[0170] FDA drug label:

[0171] Pre-set authority level scoring is used;

[0172] Raw score calculation: FDA / NMPA official documents belong to the highest regulatory level and are directly assigned full marks. point;

[0173] Normalized weights: .

[0174] Dynamic weight fusion Parameter optimization based on Bayesian method Calculate the overall score for each segment:

[0175] Guide excerpt: ;

[0176] Excerpt from the document: ;

[0177] Drug fragments: ;

[0178] Conclusion: Although the FDA label has the highest authority (1.0), its total weight is significantly lower than the latest guidelines in 2024 because it is based on 2023 data and the drugs have extremely short half-lives. This is consistent with the clinical logic of not recommending obsolete drugs.

[0179] Conflict Detection and Resolution: When the system detects differences in triage criteria between the 2020 old version of the guidelines and the 2024 new version, specifically, the system calculates the semantic similarity between the two. If the semantic similarity is lower than a preset conflict threshold, the system will resolve the conflict. When a discrepancy is detected, the conflict resolution process is triggered.

[0180] Semantic conflict triggering: The DeBERTa-v3 model is used to calculate the semantic vectors of the old and new guide segments. The semantic similarity between the two is then calculated. ;because If a substantial conflict of clinical recommendations is found, the system will forcibly activate a weighted voting mechanism.

[0181] Weighted voting decision: The system utilizes calculated comprehensive weights, new guidelines The old guidelines assumed Conduct the vote. Calculate the final score. :

[0182] The new guidelines' viewpoint score: Due to its higher weighting for timeliness, it dominates the weighted numerator and receives a higher trust score.

[0183] The old guideline viewpoints received a low score due to significant weight decay in timeliness, and were therefore marked as outdated evidence.

[0184] Final results output, visualization, and generation: The system uses the winning 2024 revised guidelines as the core input to the large model, generating the following enhanced response:

[0185] Medical advice: It is recommended to start intravenous diuretic therapy immediately, based on node [B] in the figure, which represents the latest authoritative evidence in 2024.

[0186] Conflict Warning: The old 2021 guideline for node [A] in the diagram suggested "observe first," but this suggestion has been corrected by the latest evidence for node [B]. Please refer to the winning evidence.

[0187] Weighted heatmap: In the output two-dimensional heatmap, such as Figure 3 As shown, node [B], representing a new clinical breakthrough, is located in the upper right corner, representing high timeliness and authority, and is displayed as a deep red solid dot (Winner). Node [A], representing an older guideline, is located at the top center, representing high authority but insufficient timeliness, and is displayed as a gray semi-transparent dot (Loser), visually indicating to doctors that the evidence is outdated. This figure is a dynamic weighted two-dimensional heatmap generated for the emergency heart failure triage scenario. This figure illustrates how this invention solves the misleading problem of highly authoritative but outdated medical knowledge through a timeliness-authority dual-dimensional coordinate system. It demonstrates how this invention addresses the technical pain point of blindly worshipping authority and ignoring timeliness in traditional medical question-and-answer systems. Through visualization, this figure clearly explains to doctors the decision-making basis for why the system adopts node [B] and rejects node [C], that is, the system prioritizes the latest and most authoritative evidence, effectively improving the interpretability and trustworthiness of clinical auxiliary decision-making.

[0188] The positions of the nodes in the diagram strictly correspond to the weight calculation results in this embodiment: Node [C] Drug Label: Located in the upper left corner of the chart, This node represents the FDA drug label. Although it has a perfect score in static authority and is at the top, its timeliness weight is extremely low due to its early publication date and the drug's extremely short half-life. Based on the nonlinear product model of this invention, its extremely low timeliness weight lowers the overall score, causing it to be judged as outdated evidence (Loser) and displayed in gray. Node [A] Old Guidelines: Located at the top left of the chart. This node represents the 2021 version of the guidelines. While highly authoritative, it lacks timeliness compared to the latest breakthroughs, thus losing in the weighted vote and appearing gray. Node [B] New Clinical Breakthrough: Located in the upper right corner of the chart. This node represents the latest clinical evidence from 2024, possessing both high authority and timeliness. It received the highest comprehensive search score and was judged by the system as the winning evidence, displayed as a dark red highlighted solid dot.

[0189] Example 4

[0190] This embodiment demonstrates the application of the TimeRAG method of the present invention in recommending tumor drug treatment regimens. Taking the treatment regimen for HER2-low breast cancer as an example, the system needs to resolve the time-sensitivity conflict between chemotherapy recommended in older guidelines and ADC drugs recommended by the latest clinical breakthroughs, providing physicians with recommendations that conform to the latest evidence-based medicine.

[0191] Knowledge base construction:

[0192] The system, targeting the field of breast cancer treatment, integrates the following core data resources through an API interface and annotates each knowledge chunk with structured metadata, including publication timestamps. and source type :

[0193] Candidate Evidence A: Derived from the NCCN Breast Cancer Clinical Practice Guidelines (2021 Edition), which recommends that standard single-agent chemotherapy be given priority for HER2-negative / low-expressing patients; =2021, =Guide;

[0194] Candidate evidence B: The results of the DESTINY-Breast04 clinical trial, published in the New England Journal of Medicine (NEJM) in 2024, confirm that T-DXd is significantly superior to chemotherapy in patients with low HER2 expression. =2024, =Literature;

[0195] Candidate Evidence C: Derived from the FDA-approved drug label, with expanded indications approved for T-DXd in 2022; =2022, =Medicine;

[0196] Current system time T: set to 2025.

[0197] Timeliness weight calculation :

[0198] This embodiment is based on the exponential decay model in bibliometrics. Calculate the timeliness weight of each piece of evidence. Where T is the current year. The year the evidence was published. The attenuation rate is calculated using the following formula: .

[0199] Regarding candidate evidence A:

[0200] Basis for setting: Clinical guidelines are updated relatively frequently, so a half-life was set. =5 years;

[0201] Attenuation rate calculation: ;

[0202] Weighting calculation: Time difference Year;

[0203] .

[0204] Regarding candidate evidence B:

[0205] Basis for setting: Cutting-edge literature in the field of oncology is updated rapidly, hence the setting of half-life. =5 years;

[0206] Attenuation rate calculation: ;

[0207] Weighting calculation: Time difference Year;

[0208] .

[0209] Regarding candidate evidence C:

[0210] Basis for setting: The drug's instructions are extremely sensitive to time, hence the setting of the half-life. =1 year;

[0211] Attenuation rate calculation: ;

[0212] Weighting calculation: Time difference Year;

[0213] .

[0214] The low result of 0.125 reflects the potential risks of relying solely on drug instructions from three years ago, and the system tends to depend on more recent literature evidence.

[0215] Authority weight calculation :

[0216] The system employs a heterogeneous scoring normalization mechanism to normalize raw scores from different sources. Mapped to the [0,1] interval. Assuming the current search pool contains... .

[0217] Regarding candidate evidence A:

[0218] Scoring method: The AGREE II tool was used for six-dimensional scoring;

[0219] Original score: Based on calculations, this guide performs excellently in dimensions such as rigor and applicability, with a converted score of 9.2 points;

[0220] Normalized weights: .

[0221] Regarding candidate evidence B:

[0222] Scoring method: Based on journal impact factor (IF) and citation frequency;

[0223] Original score: NEJM has an extremely high impact factor (IF>100), and the article has a large number of citations, resulting in a high overall score. =8.8 points;

[0224] Normalized weights: .

[0225] Dynamic weight fusion :

[0226] Introducing the authority index of hyperparameters determined by Bayesian optimization and timeliness index Calculate the comprehensive retrieval score for each segment. .

[0227] Candidate Evidence A:

[0228] ;

[0229] Candidate Evidence B:

[0230] ;

[0231] Conclusion: Although older guidelines are slightly more authoritative, their earlier publication date significantly diminishes the weight of timeliness, resulting in a significantly lower overall score compared to the latest clinical literature. This reflects the system's logic of considering not only authoritative sources but also the most up-to-date evidence.

[0232] Conflict Detection and Resolution:

[0233] When a user asks, "What is the preferred treatment for HER2-low expression breast cancer?", the system retrieves the above high-scoring fragment and triggers the conflict resolution process:

[0234] 1. Semantic conflict triggering:

[0235] The system calls a pre-trained language model (DeBERTa-v3) to construct semantic vectors;

[0236] Conflict analysis: Evidence A advocates "chemotherapy", while evidence B advocates "the use of T-DXd";

[0237] Calculation results: Both discuss the same condition but reach contradictory conclusions; the calculated semantic similarity is... ;

[0238] Judgment: Set threshold It is 0.35, because If a substantial conflict of clinical recommendations is found, the system will forcibly activate a weighted voting mechanism.

[0239] 2. Weighted voting decision:

[0240] The system utilizes computation As a "vote weighting":

[0241] The total weight of those supporting "chemotherapy" is 0.645, from evidence A;

[0242] The total weight supporting "T-DXd" is 0.787, from evidence B;

[0243] The verdict: Evidence B scored higher and was determined to be the winning evidence; Evidence A was marked as outdated evidence (Loser).

[0244] 3. Final Result Output, Visualization and Generation: Based on winning evidence B, the system constructs a Prompt, generating the following response, along with interpretable charts:

[0245] Medical recommendation text: "Based on the latest results of the DESTINY-Breast04 clinical trial (2024), T-DXd (trastuzumab) is recommended as an alternative to standard chemotherapy for patients with HER2-low expression breast cancer."

[0246] Conflict warning notice: "Note: The 2021 version of the old guidelines recommended prioritizing chemotherapy, but this view has been revised by the latest clinical evidence. Please refer to the conclusions of the new research."

[0247] Weighted heatmap display: In the output two-dimensional coordinate graph, such as Figure 4 As shown, Evidence B, the new literature, is located in the middle right corner, indicating extremely high timeliness and authority, and is displayed as a dark red highlighted node; Evidence A, the old guideline, is located at the top left corner, indicating low timeliness and high authority, and is displayed as a gray semi-transparent node, visually indicating to doctors that this authoritative guideline has been downgraded by the system due to its outdated timeliness. The positions of the nodes in the figure strictly correspond to the weight calculation results in this embodiment: Evidence C, the drug label, is located in the top left corner of the chart. , This shows that although it possesses the highest regulatory authority, it was deemed outdated by the system due to severe time lag; Evidence A, the old version of the guidelines, is located at the top left of the chart. , The first line represents the old standard for recommended chemotherapy in 2021. Due to the diminishing weight of time-sensitive evidence, its overall score is lower than that of the new evidence. Evidence B, the new clinical breakthrough, is located in the middle of the right side of the chart. , This represents the T-DXd treatment protocol published in the NEJM in 2024. Although its static authority is slightly lower than that of guidelines and FDA labels, it achieved the highest overall score in the nonlinear product model due to its extremely high weighting for timeliness. It was marked as a dark red highlighted solid node (Winner) and became the sole basis for the system's generated recommendations.

[0248] Example 5

[0249] This embodiment demonstrates the application of the TimeRAG method of the present invention in complex medical imaging diagnosis. Taking prenatal diagnosis of fetal heart malformation as an example, the system needs to resolve the conflict between traditional two-dimensional ultrasound diagnostic standards and the latest multimodal imaging consensus (integrating four-dimensional STIC technology and MRI), providing ultrasound physicians with high-precision auxiliary diagnostic evidence.

[0250] Knowledge base construction:

[0251] The system, designed for the field of prenatal ultrasound diagnosis, integrates the following core data resources and provides structured metadata annotations for each knowledge fragment, including publication timestamps. and source type :

[0252] Candidate evidence A: derived from the 2020 edition of the "Technical Specifications for Prenatal Two-Dimensional Ultrasound Diagnosis", which mainly relies on two-dimensional sections for malformation screening, and has a limited detection rate for complex malformations; =2020, =Documents / Standards;

[0253] Candidate evidence B: derived from the "2024 Expert Consensus on Multimodal Imaging of Fetal Heart Malformations" published in the Chinese Journal of Ultrasound in Medicine, which recommends combining MRI with ultrasound multimodal diagnosis; =2024, =Guide;

[0254] Candidate Evidence C: Derived from the latest clinical application report on "4D Ultrasound Spatiotemporal Correlation Imaging (STIC)" technology published in 2025; =2025, =Literature;

[0255] Current system time T: set to 2025.

[0256] Timeliness weight calculation :

[0257] This embodiment is based on the exponential decay model in bibliometrics. Calculate the timeliness weight of each piece of evidence. Among them, For the current year, The year the evidence was published. This represents the attenuation rate.

[0258] Regarding candidate evidence B:

[0259] Basis for setting: New technology literature belongs to the category of scientific literature, and a half-life is set accordingly. Annual decay rate ;

[0260] Weighting calculation: Time difference Year;

[0261] .

[0262] Regarding candidate evidence C:

[0263] Basis for setting: Expert consensus belongs to the guideline category, and the half-life is set accordingly. Annual decay rate ;

[0264] Weighting calculation: Time difference Year;

[0265] .

[0266] Regarding candidate evidence A:

[0267] Basis for setting: As an older technical standard, it is classified as a document, and the attenuation rate is... ;

[0268] Weighting calculation: Time difference Year;

[0269] .

[0270] Authority weight calculation :

[0271] The system employs a heterogeneous scoring normalization mechanism;

[0272] Taking candidate evidence B as an example, the AGREE II standard is used for quantitative scoring:

[0273] AGREE II Dimension Scoring: Scope and Purpose: 7 points; Participants: 7 points; Rigor in Formulation: 8 points (higher weight); Clarity of Expression: 7 points; Applicability: 7 points; Editorial Independence: 7 points. Standardized Calculation: Assuming each item has a maximum score of 7 points, the total number of items is 6 (simplifying the calculation logic to match the original text).

[0274] .

[0275] Dynamic weight fusion :

[0276] Based on nonlinear geometric fusion model Introducing Bayesian optimization parameters .

[0277] Candidate Evidence B:

[0278]

[0279] Candidate Evidence A:

[0280] Due to timeliness weight The lower the score, the lower the final composite score after the product model.

[0281] Conflict Detection and Resolution:

[0282] When a user asks, "What are the diagnostic criteria for complex fetal heart malformations?", the system triggers the conflict resolution process:

[0283] 1. Semantic conflict triggering:

[0284] The system detected a difference between evidence A and evidence B in the diagnostic pathway;

[0285] Calculate semantic similarity, if contradiction degree If so, it is determined to be a substantial conflict.

[0286] 2. Weighted voting decision:

[0287] Evidence B: With a high overall weight of 0.913, it dominated the voting and was judged as the Winner;

[0288] Evidence A: Due to insufficient timeliness, the contribution weight is extremely low. The original example mentions that the contribution weight is only 56.3% relative value, and it is judged as a loser.

[0289] 3. Final Result Output, Visualization and Generation: The system generates the following enhanced answer:

[0290] The medical recommendation states: "Based on the 2024 expert consensus, it is recommended to use multimodal imaging techniques combining ultrasound and MRI for diagnosis. For complex cases, four-dimensional STIC technology can be used as an adjunct to improve diagnostic accuracy."

[0291] Conflict warning notice: "Note: The old version of the standard in 2020 mainly relied on two-dimensional ultrasound, which may have the risk of missing diagnosis for complex malformations. It has been supplemented and revised by the latest multimodal consensus."

[0292] Weighted heatmap display: In the output chart, such as Figure 5 As shown, evidence B is displayed as a highlighted red dot in the upper right corner, while evidence A is displayed as a dark dot in the lower left corner, clearly indicating to the doctor the basis for accepting evidence B. This diagram visually demonstrates the system's evidence adjudication process when facing complex malformation diagnoses:

[0293] Evidence A: Located in the lower left area of ​​the chart. This node represents the two-dimensional ultrasound standard released in 2020. Due to its limited technical means, its technical authority in diagnosing complex malformations is not as high as the latest multimodal methods. Moreover, it was released earlier, so it is at a disadvantage in terms of timeliness and authority, and is shown as a gray dot.

[0294] Evidence B: Located in the upper right corner of the chart. This node represents the multimodal imaging expert consensus released in 2024, which combines the advantages of MRI and ultrasound, and has extremely high timeliness and authority. It was identified by the system as the dark red highlighted node Winner.

[0295] Evidence C: Located on the right side of the chart. The figure, representing the latest STIC technical report from 2025, appears as timely supporting evidence, reinforcing the diagnostic recommendation of winning evidence B.

[0296] Example 6

[0297] The performance of the proposed method and existing methods is compared on the TimeMedQA dataset, as shown in Table 1:

[0298] Table 1. Performance Comparison between the Method of the Present Invention and Existing Methods

[0299]

[0300] Experimental results show that TimeRAG improves accuracy by 12.3% and 15.8% compared to LINS and MedRAG models, respectively, and reduces the outdated recommendation rate from 25.6% to 14.9%. The timeliness score increases from 69.8% for LINS to 88.4%, indicating that TimeRAG can more effectively utilize the latest medical knowledge.

[0301] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing the generation of medical question-answering based on dynamic weighted multi-knowledge bases, characterized in that, Includes the following steps: S1. Construct a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples; Unstructured knowledge fragments from a multi-source medical knowledge base are preprocessed and vectorized for storage, generating a quintuple for each unstructured knowledge fragment. ,in, Represents text content; Vectorized representation; Indicates the publication timestamp; The source type label indicates four sources: clinical guideline database, scientific literature database, drug database, and multimodal image database. This represents the original score indicating static authority. S2. Construct an adaptive timeliness weight calculation module based on half-life theory to calculate the timeliness weight of knowledge fragments; the specific process is as follows: An exponential decay calculation model based on time difference is established, and the formula is expressed as follows: ， in, Indicates the timeliness weight of knowledge fragments; The time-depreciation rate of the source type to which the knowledge fragment belongs; This indicates the latest year; the formula for calculating the time-lapse rate of the source type of a knowledge fragment is: ， in, Indicates the average half-life of knowledge types; S3. Construct a multidimensional authority weight calculation module based on AGREE II and bibliometrics to calculate the authority weight; the specific process is as follows: A standardized and normalized model of authority weights is constructed, and the authority weights are calculated using the Min-Max normalization algorithm. The formula is as follows: ， in, Indicates authority weight; This represents the minimum original authority score among the currently retrieved candidate knowledge sets; This represents the maximum original authority score among the currently retrieved candidate knowledge sets; S4. Construct a dynamic weighted fusion module based on a nonlinear product model to calculate the comprehensive retrieval score of knowledge fragments; S5. Construct a semantic-timeliness dual conflict detection and weighted voting resolution module, calculate the semantic similarity coefficient of the top k knowledge fragments, and obtain the credibility score by combining the semantic similarity coefficient and the comprehensive retrieval score, thereby obtaining the set of winning knowledge fragments and outdated knowledge fragments; the specific process is as follows: Based on the knowledge fragments retrieved from the user's question, the knowledge fragments are sorted from high to low according to their comprehensive retrieval scores. The semantic similarity coefficient of the top k knowledge fragments is calculated, and the formula is expressed as follows: ， in, , This represents the vectorized representation of two different knowledge fragments retrieved. Represents the semantic similarity coefficient; Indicates the length of the vector; Set a threshold for contradiction Iterate through the weighted segments of the first k knowledge segments. ,when When two knowledge fragments are determined to be semantically identical, a standard text splicing strategy is used to integrate the two knowledge fragments into a context; when... If two knowledge fragments are deemed to have a substantial conflict of clinical recommendations, a weighted vote is performed on the knowledge fragments, and the decision score is calculated. The formula is as follows: ， in, Representing knowledge fragments Credibility score; A collection of knowledge fragments that support the same semantic viewpoint; This represents the comprehensive retrieval score of the j-th knowledge fragment in the set; This indicates the direct correlation between the knowledge fragment and the user's question Q; The process involves iterating through the knowledge segments that participate in the voting and have substantial conflicts with clinical recommendations, identifying the knowledge segment with the highest adjudication score, and selecting it as the winning knowledge segment. The formula is as follows: ， in, This represents a winning knowledge fragment; Represents a set of conflicting knowledge fragments; This represents a knowledge fragment within a set of conflicting knowledge fragments; This represents the variable maximization operator; For the conflict set except Other knowledge fragments Determined to be outdated knowledge fragment: ， in, This represents a set of outdated knowledge fragments.

2. The medical question-answering enhancement generation method based on dynamic weighted multi-knowledge base as described in claim 1, characterized in that, The multi-source medical knowledge base includes a clinical guidelines database, a scientific literature database, a drug database, and a multimodal image database.

3. The medical question-answering enhancement generation method based on dynamic weighted multi-knowledge base according to claim 2, characterized in that, The adaptive update formula for the timeliness decay rate of the source type of a knowledge fragment is: ， in, These represent the decay rates before and after the update, respectively. This indicates the difference between the current monitoring period's expiration rate and the previous period's rate; Indicates the learning rate; This represents the baseline overdue rate constant.

4. The medical question-answering enhancement generation method based on dynamic weighted multi-knowledge base according to claim 3, characterized in that, A geometrically weighted nonlinear fusion model is constructed, expressed by the following formula: ， in, This represents the overall retrieval score for a knowledge fragment; The contribution index representing the weight of authority; The contribution index representing the weight of timeliness; This represents the dot product operation.

5. A medical question-answering enhancement generation system based on a dynamically weighted multi-knowledge base, executing the medical question-answering enhancement generation method based on a dynamically weighted multi-knowledge base as described in claim 1, characterized in that, include: Data processing module: used to build a multi-source medical knowledge base and perform preprocessing to obtain standardized data tuples; Timeliness Weight Calculation Module: Used to calculate the timeliness weight of knowledge fragments; Authority Weight Calculation Module: Used to calculate authority weight; Dynamic weighted fusion module: used to calculate the comprehensive retrieval score of knowledge fragments; The dual conflict detection and weighted voting resolution module is used to calculate the semantic similarity coefficient of the top k knowledge fragments. The credibility score is calculated by combining the semantic similarity coefficient and the comprehensive retrieval score, and then the set of winning knowledge fragments and outdated knowledge fragments is obtained. The Medical Question Answering Enhancement Generation and Interpretability Output Module is used to build structured prompt word templates and generate medical advice text, dynamic weighted heatmaps, and traceable evidence chains.

Images