Literature metrology and clinical data fusion lung cancer knowledge graph construction method and device

Abstract

The invention discloses a literature metrology and clinical data fusion lung cancer knowledge graph construction method and device, which can solve the problems of research and clinical disjunction, data fragmentation and AI'black box ', and realize systematic mining, verification and interpretable application of lung cancer diagnosis factors. The method comprises the following steps: (1) mining high-frequency and emerging diagnosis factors from scientific research literatures through a multi-dimensional algorithm, and generating academic prior weights; (2) enabling the literature candidate factors to correspond to real patient data, and performing structured matching to obtain an aligned clinical candidate factor set; (3) screening out a core factor having a reliable causal relationship with the lung cancer outcome from the clinical candidate factors, and providing a causal-driven skeleton for the knowledge graph; (4) establishing an initial knowledge graph based on D3 and G, and dynamically updating an edge weight based on newly added data; and (5) outputting a clinical interpretable report which comprises a puncture score Pbiopsy, driving factor contribution degree ranking, an evidence chain and a recommended path.

Description

Technical Field

[0001] This invention relates to the fields of medical information processing and artificial intelligence, and in particular to a method for constructing a lung cancer knowledge graph by fusing bibliometrics and clinical data, as well as a device for constructing a lung cancer knowledge graph by fusing bibliometrics and clinical data. Background Technology

[0002] Traditional medical knowledge graphs rely on a single data source (clinical guidelines, public databases, or manual annotation), which has significant limitations. Currently, in the diagnosis of pulmonary nodules and lung cancer, doctors often need to make comprehensive judgments based on multi-source data such as imaging, laboratory tests, pathology, and genetic testing, but this further highlights the following core issues:

[0003] (1) Disconnect between research and clinical practice: High-frequency research hotspots in academic literature (such as emerging biomarkers) lack correlation with real clinical data and are difficult to be directly used for diagnosis and treatment decisions;

[0004] (2) Data fragmentation: Clinical data (imaging features, test indicators, pathological conclusions, etc.) are stored in different systems, lacking a unified semantic knowledge structure, making it difficult to achieve cross-modal data correlation analysis;

[0005] (3) AI “black box” problem: Although existing auxiliary diagnostic models have certain predictive capabilities, the decision-making process lacks causal explanation and literature evidence, making it difficult for clinicians to trust or adopt their suggestions.

[0006] Existing technologies fail to address the aforementioned issues, either by constructing static knowledge graphs based solely on literature (lacking clinical adaptation mechanisms), by mining association rules from clinical data (without academic evidence backing), or by lacking causal verification algorithms (unable to distinguish between correlation and causation). Therefore, there is an urgent need for a closed-loop algorithm that integrates "hot literature knowledge" with "real clinical evidence" to automate the entire process from research hotspot discovery to clinically interpretable applications. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a method for constructing a lung cancer knowledge graph by integrating bibliometrics and clinical data. This method can solve the problems of disconnect between research and clinical practice, data fragmentation, and the "black box" nature of AI, and realize the systematic mining, verification, and interpretable application of lung cancer diagnostic factors.

[0008] The technical solution of this invention is: a method for constructing a lung cancer knowledge graph by fusing bibliometrics and clinical data, which includes the following steps:

[0009] (1) High-frequency and emerging diagnostic factors are mined from scientific research literature using a multi-dimensional algorithm to generate academic prior weights; the input is lung cancer-related literature from globally authoritative literature databases such as PubMed, PMC, CNKI, and patent databases; the output is a set of candidate factors from the literature, including W for each factor. lit (f) Values, topic tags, and time trends are denoted as dataset D1;

[0010] (2) Match the candidate factors in the literature with the real patient data and perform structured matching to obtain the aligned set of clinical candidate factors; the input is the original clinical multimodal data, denoted as dataset D0, which includes image DICOM files, test indicators, pathology reports, gene test results, and doctor text records, denoted as dataset D1; the output is the aligned set of clinical candidate factors, denoted as dataset D2, where each factor contains basic attributes, weight attributes, and clinical correlation attributes;

[0011] (3) From the clinical candidate factors, core factors with reliable causal relationships with lung cancer outcomes are screened to provide a causal-driven framework for the knowledge graph; the input is the dataset.

[0012] D2, clinical multimodal feature matrix M, pathological diagnosis outcome variable Y; the output is a set of core diagnostic factors and an optimized causal graph G. The set of core diagnostic factors includes the causal strength, confidence score and priority score of the factors, denoted as dataset D3. The optimized causal graph G only retains the causal edges between the factors in D3 and Y.

[0013] (4) Build an initial knowledge graph based on D3 and G, and dynamically update the edge weights based on the new data; where the input is dataset D3, causal graph G, evidence source data, and new data; the output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions;

[0014] (5) Output a clinically interpretable report, including: puncture score P biopsy Ranking of driving factor contributions, evidence chain, and recommendation path.

[0015] This invention addresses the issues of disconnect between research and clinical practice, data fragmentation, and the "black box" nature of AI by establishing a core algorithm loop of "literature hotspot discovery → clinical semantic alignment → causal structure verification → dynamic graph update → interpretable clinical output," thereby enabling the systematic mining, verification, and interpretable application of lung cancer diagnostic factors.

[0016] A knowledge graph construction device for lung cancer that integrates bibliometrics and clinical data is also provided, which includes:

[0017] The data mining module is configured to extract high-frequency and emerging diagnostic factors from scientific literature using a multi-dimensional algorithm, generating academic prior weights. The input consists of lung cancer-related literature from globally authoritative databases such as PubMed, PMC, CNKI, and patent databases. The output is a set of candidate factors from the literature, including the W value for each factor. lit (f) Values, topic tags, and time trends are denoted as dataset D1;

[0018] The alignment module is configured to match candidate factors from literature with real patient data in a structured manner to obtain an aligned set of clinical candidate factors. The input is the original clinical multimodal data, denoted as dataset D0, which includes DICOM images, laboratory indicators, pathology reports, gene test results, and physician text records, denoted as dataset D1. The output is the aligned set of clinical candidate factors, denoted as dataset D2, where each factor contains basic attributes, weight attributes, and clinical relevance attributes.

[0019] The screening module is configured to select core factors with a reliable causal relationship with lung cancer outcomes from clinical candidate factors, providing a causal-driven framework for the knowledge graph; the inputs are dataset D2, clinical multimodal feature matrix M, and pathological diagnosis outcome variable Y.

[0020] The output consists of a core set of diagnostic factors and an optimized causal graph G. The core set of diagnostic factors includes the causal strength, confidence score, and priority score of the factors, denoted as dataset D3. The optimized causal graph G retains only the causal edges between the factors and Y in D3.

[0021] The module is configured to build an initial knowledge graph based on D3 and G, and dynamically update edge weights based on newly added data. The inputs are dataset D3, causal graph G, evidence source data, and newly added data; the output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions.

[0022] The output module is configured to output clinically interpretable reports, including: puncture score P. biopsy Ranking of driving factor contributions, evidence chain, and recommendation path. Attached Figure Description

[0023] Figure 1 A flowchart of a method for constructing a lung cancer knowledge graph based on the fusion of bibliometrics and clinical data, according to the present invention, is shown.

[0024] Figure 2 A schematic diagram of the knowledge graph node relationships according to the present invention is shown. Detailed Implementation

[0025] like Figure 1 As shown, a method for constructing a lung cancer knowledge graph by fusing bibliometrics and clinical data includes the following steps:

[0026] (1) High-frequency and emerging diagnostic factors are mined from scientific research literature using a multi-dimensional algorithm to generate academic prior weights; the input is lung cancer-related literature from globally authoritative literature databases such as PubMed, PMC, CNKI, and patent databases; the output is a set of candidate factors from the literature, including W for each factor. lit (f) Values, topic tags, and time trends are denoted as dataset D1;

[0027] (2) Match the candidate factors in the literature with the real patient data and perform structured matching to obtain the aligned set of clinical candidate factors; the input is the original clinical multimodal data, denoted as dataset D0, which includes image DICOM files, test indicators, pathology reports, gene test results, and doctor text records, denoted as dataset D1; the output is the aligned set of clinical candidate factors, denoted as dataset D2, where each factor contains basic attributes, weight attributes, and clinical correlation attributes;

[0028] (3) From the clinical candidate factors, core factors with reliable causal relationship with lung cancer outcome are screened out to provide the causal-driven skeleton for knowledge graph; the input is dataset D2, clinical multimodal feature matrix M, and pathological diagnosis outcome variable Y; the output is the core diagnostic factor set and the optimized causal graph G. The core diagnostic factor set contains the causal strength, confidence score and priority score of the factors, denoted as dataset D3. The optimized causal graph G only retains the causal edges between the factors in D3 and Y.

[0029] (4) Build an initial knowledge graph based on D3 and G, and dynamically update the edge weights based on the new data; where the input is dataset D3, causal graph G, evidence source data, and new data; the output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions;

[0030] (5) Output a clinically interpretable report, including: puncture score P biopsy Ranking of driving factor contributions, evidence chain, and recommendation path.

[0031] This invention addresses the issues of disconnect between research and clinical practice, data fragmentation, and the "black box" nature of AI by establishing a core algorithm loop of "literature hotspot discovery → clinical semantic alignment → causal structure verification → dynamic graph update → interpretable clinical output," thereby enabling the systematic mining, verification, and interpretable application of lung cancer diagnostic factors.

[0032] Preferably, step (1) includes the following sub-steps:

[0033] (1.1) Document collection and preprocessing: Document metadata is collected in batches through API interface, and a structured corpus is formed after deduplication and standardization; all statistical data are analyzed, and the chi-square test or Fisher's exact test is used for count variables, and the independent samples t test is used for quantitative variables, with p value < 0.05;

[0034] (1.2) Topic and Feature Extraction: The BERTopic / LDA topic modeling algorithm is used to cluster research hotspots and identify potential literature candidate factors;

[0035] (1.3) Constructing a keyword co-occurrence network: Extracting keywords from each document and mapping them to...

[0036] UMLS / MeSH standard ontology; count the frequency of any two keywords appearing together in the same document to form a co-occurrence matrix; use keywords as nodes and co-occurrence frequency as edge weights, retain edges with a co-occurrence frequency ≥ 3 times, and construct an undirected co-occurrence network;

[0037] (1.4) Time trend analysis: The frequency of keyword occurrences is statistically analyzed by year to form a time series;

[0038] (1.5) Multi-dimensional weight calculation

[0039] W lit (f)=α·F(f)+β·C(f)+γ·PR(f)+δ·B(f)

[0040] Among them, W lit (f) represents the comprehensive document weight of diagnostic factor f; F(f) is the normalized frequency, calculated by dividing the total number of times keyword f appears in the documents by the maximum number of times the keyword appears in the corpus; C(f) is the co-occurrence network centrality (reflecting term relevance), calculated using node centrality. cooccur(f,i) is the co-occurrence frequency of f and keyword i, with the denominator being the maximum node degree in the network, ensuring C(f)∈[0,1]); PR(f) is the PageRank citation weight, constructing a document citation network where nodes are documents and directed edges represent citation relationships, and running the PageRank algorithm to obtain the PR value of each document; B(f) is the time burst weight; based on the Kleinberg burst detection algorithm, the frequency of f is statistically analyzed by year to form a time series T(f)=[t1,t2,…,t n ] Calculate the probability of each year being a sudden state, take the sum of the probabilities of the past 3 years and normalize it to get B(f)∈[0,1]; α,β,γ,δ are harmonic parameters that satisfy α+β+γ+δ=1. Through joint optimization by domain expert experience and data verification, ensure that the weight balances academic hotspots and citation authority.

[0041] Preferably, step (2) includes the following sub-steps:

[0042] (2.1) Standardization of clinical data:

[0043] Structured data processing: Normalize the test indicators in D0 and map the diagnostic codes to the ICD-10 standard codes;

[0044] Image data: PyRadiomics tools were used to extract radiomics features of the nodule region and generate a structured feature matrix;

[0045] Unstructured text: Based on the BioBERT pre-trained model combined with the UMLS medical vocabulary, clinical entities are extracted from pathology / radiology reports, with an entity recognition F1 score ≥ 0.90;

[0046] Output: A standardized set of clinical entities, denoted as D. 01 This includes the entity name, data type, and original value.

[0047] (2.2) Cross-source semantic alignment:

[0048] Compare the literature candidate factor f in D1 with D 01 The alignment confidence S of the clinical entity e in the formula is calculated using the following formula. align (f,e)

[0049] S align (f,e)=λ1·cos sim (v f ,v e )+λ2·Jaccard(T f ,T e )+λ3·Ontology_dist(O f O e )

[0050] Among them, cos sim (v f ,v e f is the cosine similarity between the word vectors of the two words; the BioBERT model is used to encode f and e respectively, generating 768-dimensional word vectors v. f ,v e ,pass Calculate the cosine similarity of vectors, ranging from [-1, 1], and map the absolute value to [0, 1]; Jaccard(T f ,T e ) represents the Jaccard matching degree of the term set; extract the relevant term set T of f. f Extract the relevant term set T of e. e ;pass Calculate the Jaccard coefficient, in the range [0,1]; Ontology_dist(O f O e To unify the concept distance in the medical ontology, the Unified Medical Language System (UMLS) is chosen as the unified ontology, and f and e are mapped to concept nodes O in UMLS. f and O e Based on the hierarchical structure of UMLS, take O f and O e The distance value is obtained by taking the number of edges of the shortest path between them; dividing the distance value by the maximum possible distance in the ontology yields Ontology_dist(O f O e )∈[0,1]; λ1,λ2,λ3 are weight parameters, based on annotated literature terminology-clinical entity matching datasets, minimizing semantic alignment error through grid search; constraints λ1+λ2+λ3=1 and λ1≥0.4; Filter S align High-confidence matching pairs with (f,e)≥0.7 are used to ensure semantic consistency.

[0051] (2.3) Calculation of overall weight:

[0052] For successfully matched factors, a cross-source comprehensive score is generated by integrating literature weights and clinical characteristics:

[0053] W align (c)=μ1·W lit (f)+μ2·freq clinical (e)+μ3·OR(e)

[0054] Among them, W lit (f) represents the fusion, directly derived from the literature weight of factor f in D1, reflecting the comprehensive value of this factor in academic literature; freq clinical (e) represents the clinical frequency, which is the number of cases in D0 where clinical entity e appears / the total number of cases in D0, normalized to [0,1]; OR represents the association strength, which is the ratio of the probability of lung cancer malignancy when the factor is present to the probability of lung cancer malignancy when the factor is absent, based on the case-control analysis of D0; μ1, μ2, and μ3 are set by clinical experts in combination with clinical needs, satisfying μ1+μ2+μ3=1.

[0055] Preferably, step (3) includes the following sub-steps:

[0056] (3.1) Construction of causal graphs integrating prior knowledge from literature:

[0057] A joint NOTEARS and DAG-GNN algorithm is used to construct a causal graph with the dual objectives of data fitting and literature constraints.

[0058] Based on the feature matrix M and the outcome Y, the potential causal relationships between factors and between factors and Y are initially generated through data-driven learning, and the edge weight matrix W is preliminarily generated. ij This represents the causal strength of factor i on factor j;

[0059] Introducing W in D2 lit (f) As a prior constraint: For factors with high support from the literature, the causal edges they participate in are more difficult to sparsify in the optimization process.

[0060] Verify and ensure that the cause-effect graph is a directed acyclic graph (DAG), eliminating the logical contradiction that i→j and j→i;

[0061] (3.2) Multidimensional causal confidence score:

[0062] For each causal relationship f→Y, the confidence level is quantified from four dimensions, and a weighted fusion model is constructed.

[0063] The model generates causal confidence scores, and the formula is as follows:

[0064] Conf causal (f→Y)=ω1·ATE score +ω2·CI width +ω3·Lit support +ω4·Clinical consistency

[0065] Among them, the average treatment effect standardized value ATE score The average magnitude of the influence of factor f on outcome Y when it exists is a core indicator of the magnitude of causal effect. The calculation method involves performing 1000 bootstrap resampling operations on the edge f→Y in the causal graph, calculating the average treatment effect ATE for each sampling, and standardizing all ATE values ​​to the [0,1] interval. A larger ATE indicates a stronger causal driving force of X on Y. The confidence interval width is the reciprocal of CI. width Based on the Bootstrap sampling results, the 95% confidence interval of ATE is calculated. The width of the confidence interval is taken, and its reciprocal is calculated. This reciprocal is then standardized to the [0,1] interval. CI width A higher Lit value indicates smaller statistical fluctuations in the causal effect and a more reliable conclusion; support Based on dataset D1, the number of documents supporting a causal association between f and Y is counted, divided by the total number of documents related to either X or Y to obtain the percentage, ranging from [0,1]. A higher Lit_support indicates greater acceptance of the causal relationship in the academic field, enhancing the academic credibility of the conclusion; clinical consistency. consistencyTo assess the universality of this causal relationship across different clinical scenarios, the original clinical data D0 was grouped by hospital and year. The ATE (Absolute Value) for f→Y was calculated for each group, and the standard deviation of the ATE for each group was calculated. After standardizing the standard deviation, 1-standardized standard deviation was used to obtain the Clinical value. consistency Clinical consistency The higher the value, the more likely the causal relationship is not a random outcome of a specific sample and has clinical applicability; ω1, ω2, ω 3， ω4 is a weighting parameter that balances the contribution of the four dimensions to the total score, reflecting the importance and priority of different dimensions. The scoring is based on the core elements of causal inference, with a panel of experts consisting of two oncologists and two statisticians.

[0066] (3.3) Prioritization and screening of core factors:

[0067] For factor f, a multi-dimensional weighted fusion model is constructed to rank the factors using a priority formula, resulting in a priority score:

[0068] Priority(f) = θ1·Conf causal (f→Y)+θ2·W lit (f)+θ3·AUC improve (f)+θ4·Clinical freq (f)

[0069] in:

[0070] AU Cimprove To measure the improvement in AUC of a lung cancer diagnostic model after incorporating factor f, the XGBoost algorithm was used. The tree depth was 5, the learning rate was 0.1, and the number of iterations was 100. The training and validation sets were split from D0 in a 7:3 ratio. The AUC of the validation set was calculated separately for the model with and without factor f, and the difference was taken as the AUC. improve (f), standardized to the [0,1] interval;

[0071] Clinical freq θ1 represents the frequency of occurrence of factor f in D0; θ2, θ3, and θ4 are parameters set by experts to balance the contribution of the four dimensions in the priority score and reflect the clinical decision priority of each dimension; a priority threshold of Priority(f) ≥ 0.7 is set, and only factors with a priority ≥ 0.7 are retained as core diagnostic factors.

[0072] Preferably, in step (3.2), the core elements of causal inference include: effect size, statistical reliability, academic support, and clinical applicability.

[0073] Preferably, in step (4), the document ID comes from the document database of step (1) and corresponds to the original document ID that supports the factor relationship in D3; the clinical case identifier comes from D0 of step (2) and corresponds to the real case that verifies the factor relationship.

[0074] Preferably, step (4) includes the following sub-steps:

[0075] (4.1) Map initialization:

[0076] An initial knowledge graph was built based on D3 and G, transforming abstract factors and causal relationships into structured graph nodes and edges, which were stored using the Neo4j graph database. The nodes and edges of the knowledge graph followed the following principles:

[0077] Node definition: Define node types based on the clinical types of factors in D3, and associate each node with basic attributes;

[0078] Edge Definition: Based on the causal relationship types in G, edge types are defined. The core attributes of an edge include: causal confidence derived from D3. causal Evidence source data, initial timestamp;

[0079] (4.2) Dynamic updates:

[0080] The edge weights are dynamically updated based on newly added data, using the following formula:

[0081] W edge (t)=η·W edge (t-1)+(1-η)·[μ·Causal strength +v·New evidence ]

[0082] Among them, W edge (t) and W edge (t-1) represents the edge weights at the current time t and the previous version t-1, respectively; η is the historical weight decay factor, which controls the retention ratio of historical weights and balances the stability of old evidence with the timeliness of new evidence, η∈[0,1]; μ and ν are the weight parameters of new evidence, satisfying μ+v=1; Causal strength The causal strength of the newly added data is calculated by quantifying the causal strength of the newly added data in step (3); evidence To increase the support of new evidence, the calculation method is the percentage of new literature that supports the relationship plus the percentage of new cases that verify the relationship, normalized to [0,1].

[0083] Preferably, in step (4), the nodes are core diagnostic factors and disease outcomes; the edges are causal or related relationships, labeled with real-time updated Wedge(t); the attributes include: the Confcausal associated with the edge, the source of evidence, and the timestamp; version tracing supports querying the graph status at any point in time.

[0084] Preferably, step (5) includes the following sub-steps:

[0085] (5.1) Patient data mapping:

[0086] Based on the semantic alignment in step (2), the features of patient P are mapped to nodes of KG, and the correlation factor values ​​of the corresponding nodes are obtained from KG and standardized to f. i And causal confidence Conf i ;

[0087] (5.2) Puncture score calculation:

[0088] The necessity score for puncture was calculated by integrating patient characteristics, causal confidence level, and clinical weights.

[0089] P biopsy =σ(∑[β i ·f i Conf i ]+β0)

[0090] Among them, σ uses the sigmoid function to compress the linear output to [0,1], which is convenient for clinical interpretation;

[0091] β i β0 represents the factor contribution coefficient, quantifying the actual impact weight of each factor on the puncture decision; β0 is the set baseline offset.

[0092] (5.3) Contribution and Evidence Chain Generation:

[0093] Contribution calculation:

[0094]

[0095] Evidence chain generation:

[0096] Extracting the causal path from factors to the outcome from KG, and relating the sources of evidence for step (4),

[0097] To form structured evidence.

[0098] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a device for constructing a lung cancer knowledge graph by fusing bibliometrics and clinical data. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0099] The data mining module is configured to extract high-frequency and emerging diagnostic factors from scientific literature using multi-dimensional algorithms, generating academic prior weights; the input is a globally authoritative literature database.

[0100] Lung cancer-related literature from PubMed, PMC, CNKI, and patent databases; the output is a set of candidate factors for the literature, including the W value for each factor. lit (f) Values, topic tags, and time trends are denoted as dataset D1;

[0101] The alignment module is configured to match candidate factors from literature with real patient data in a structured manner to obtain an aligned set of clinical candidate factors. The input is the original clinical multimodal data, denoted as dataset D0, which includes DICOM images, laboratory indicators, pathology reports, gene test results, and physician text records, denoted as dataset D1. The output is the aligned set of clinical candidate factors, denoted as dataset D2, where each factor contains basic attributes, weight attributes, and clinical relevance attributes.

[0102] The screening module is configured to select core factors with a reliable causal relationship with lung cancer outcomes from clinical candidate factors, providing a causal-driven framework for the knowledge graph; the inputs are dataset D2, clinical multimodal feature matrix M, and pathological diagnosis outcome variable Y.

[0103] The output consists of a core set of diagnostic factors and an optimized causal graph G. The core set of diagnostic factors includes the causal strength, confidence score, and priority score of the factors, denoted as dataset D3. The optimized causal graph G retains only the causal edges between the factors and Y in D3.

[0104] The module is configured to build an initial knowledge graph based on D3 and G, and dynamically update edge weights based on newly added data. The inputs are dataset D3, causal graph G, evidence source data, and newly added data; the output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions.

[0105] The output module is configured to output clinically interpretable reports, including: puncture score P. biopsyRanking of driving factor contributions, evidence chain, and recommendation path.

[0106] The following section focuses on the diagnostic scenario of benign and malignant lung nodules in lung cancer, briefly describing the implementation process of each step in this invention:

[0107] 1. Implementation of Literature Data Collection and Processing

[0108] Data source: Retrieved lung cancer / pulmonary nodule related literature from 2014 to 2024 from PubMed, CNKI and the National Patent Database, filtered out duplicate literature (12,305 articles) and low-quality literature with an impact factor <1 (7,120 articles), and retained 60,217 structured articles;

[0109] Core operations:

[0110] Use the BERTopic tool to cluster research hotspots (such as "liquid biopsy" and "radiomics") to identify potential diagnostic factors;

[0111] Construct a keyword co-occurrence network (mapped to the UMLS standard ontology) and combine it with Kleinberg burst detection to analyze the time trend; the edge retention threshold of the co-occurrence network can be set to "co-occurrence frequency ≥ 3 times";

[0112] The comprehensive weight W of the literature is calculated using a multi-dimensional formula. lit (f) Balancing frequency, relevance, authority, and recent popularity. The harmonic parameters for the multi-dimensional weighting formula can be found at α = 0.25, β = 0.2, γ = 0.3, and δ = 0.25.

[0113] Output: Candidate diagnostic factor set D1 (125 factors in total, including W) lit (f) Hashtags, time trends, etc.

[0114] 2. Implementation of Clinical Multimodal Data Alignment

[0115] Data source: Clinical data of patients undergoing lung biopsy at the hospital (including CT images, laboratory indicators, pathology reports, and gene testing results), with a total of 2,988 cases having complete data.

[0116] Core operations:

[0117] Standardization: PyRadiomics was used to extract radiomics features, and the BioBERT model extracted clinical entities (such as “spiculated hairs” and “elevated CEA”) from the text.

[0118] Semantic alignment: Calculate the alignment confidence between document factors and clinical entities according to the formula, and screen high-confidence matching pairs; cross-source semantic alignment weights can be referenced at λ1=0.5, λ2=0.3, λ3=0.2;

[0119] By integrating literature weights with clinical characteristics (frequency, OR value), a cross-source comprehensive weight W is generated. align (c) The fusion parameters can be referenced as μ1 = 0.4, μ2 = 0.3, and μ3 = 0.3;

[0120] Output: Aligned clinical candidate factor set D2 (83 factors in total, including matching relationships, W) align (c) Clinical relevance attributes.

[0121] 3. Implementation of Causal Structure Learning and Core Factor Screening

[0122] Inputs: D2, clinical multimodal feature matrix M (2988 cases × 83 columns (factors)), pathological outcome variables (benign / malignant).

[0123] Core operations:

[0124] A causal graph was constructed using the NOTEARS+DAG-GNN algorithm, and W was introduced. lit (f) serves as a priori constraint to ensure acyclic structures;

[0125] Causal confidence is quantified from four dimensions: causal effect, statistical stability, literature support, and clinical consistency; the weights for causal confidence can be referenced as ω1 = 0.4, ω2 = 0.3, ω3 = 0.15, and ω4 = 0.15.

[0126] High-priority core factors were selected by combining the performance improvement value of the diagnostic model and the clinical frequency ranking.

[0127] Output: Core diagnostic factor set D3 (6 factors in total, including causal confidence and priority), causal graph structure G.

[0128] 4. Implementation of Knowledge Graph Construction and Dynamic Updates

[0129] Core operations:

[0130] Initialization: The graph was constructed using Neo4j, with nodes defined as core factors / disease types and edges defined as causal / correlation relationships. The confidence level of D3 was associated with the source of evidence (document ID, case identifier).

[0131] Graph size: The initial graph contains 1,186 nodes (12 diseases, 78 imaging features, 39 test indicators, etc.) and 12,893 edges (178 causal edges, 2,345 related edges, etc.).

[0132] Dynamic updates: New literature / clinical data are regularly included (processed in steps 1-3), edge weights are adjusted according to the formula, version traceability is retained, and dynamic update parameters can be set to η=0.7, μ=0.6, v=0.4, with the update cycle set to "once per quarter".

[0133] Output: Dynamic knowledge graph KG (including nodes, edges, weights, evidence, and version records, such as "Initial version 2024-01-15" and "Updated version 2024-04-20").

[0134] 5. Clinically explainable output implementation

[0135] Inputs: KG, multimodal data of a single patient (such as CT images, test results).

[0136] Core operations:

[0137] Patient characteristics are mapped to KG nodes, and factor values ​​and causal confidence scores are extracted.

[0138] The necessity score for puncture is calculated using a formula, and the contribution of each factor is quantified.

[0139] Extract causal paths and evidence chains from KG to generate structured reports.

[0140] Output: Clinically interpretable report (including puncture score, factor contribution ranking, chain of evidence, and recommended pathway).

[0141] Output report example: Puncture score P biopsy =0.806 (high necessity), contribution ranking "ctDNA increase (35%) > spiculation sign (24%)", evidence chain includes "supporting literature PMID 3812XXX (328 articles), verification case P2019XXX (668 cases)".

[0142] The beneficial technical effects of the present invention are as follows:

[0143] 1. Cross-source fusion algorithm: Breaking through multimodal data barriers and solving the "research-clinical disconnect"

[0144] Based on the literature comprehensive weight formula (W) in step 1 lit (f) and the cross-source semantic alignment algorithm in step 2 (S align (f,e)) to achieve standardized and quantitative integration of academic research (candidate factors of literature) and clinical multimodal data (radiomics, test indicators, pathological entities, etc.); through UMLS ontology mapping and BioBERT semantic matching, to unify the semantic scope of literature terms and clinical entities, and completely break the disconnect between "academic hotspots without clinical validation and clinical data without academic support".

[0145] 2. Causal learning mechanism: Enhancing the interpretability of AI

[0146] Breaking the "black box" dilemma relies on the "literature-prior-constrained causal graph construction algorithm" in step 3.1 and the multi-dimensional causal confidence model in step 3.2, upgrading from "data association" to "causal cognition"—only retaining factors with a direct causal relationship to lung cancer outcomes (benign / malignant) and eliminating spurious variables; at the same time, through a four-dimensional scoring of "causal effect + statistical stability + literature support + clinical consistency", it provides traceable causal evidence for each decision, significantly improving clinicians' trust in and adoption rate of AI suggestions.

[0147] 3. Dynamic update system: Ensures the timeliness of knowledge and avoids the lag of static knowledge bases.

[0148] Based on the edge weight iteration formula (W) in step 4.2 edge (t)) supports the continuous evolution of knowledge graphs with new evidence (newly published literature and new clinical cases, all of which have been fully verified in steps 1-3); it balances the "stability of old evidence" and the "timeliness of new evidence" through the historical weight decay factor (η), while retaining the ability to trace all versions (such as edge weights and evidence sources at any point in time), thus completely solving the pain points of traditional static knowledge bases that are "slow to update, difficult to iterate, and unable to adapt to medical progress".

[0149] 4. High clinical adaptability: Provides actionable decision-making suggestions to improve diagnostic and treatment efficiency.

[0150] The puncture scoring algorithm and factor contribution decomposition model in step 5 directly output clinically understandable structured results: providing a quantitative score for the necessity of puncture (such as P). biopsy =0.85 corresponds to "high puncture demand"), and clarifies the relative contribution of each factor to the decision (such as "spiculation sign contributes 40%)", while linking the literature ID in the knowledge graph with the clinical case evidence chain; assisting doctors in accurately judging the treatment path, reducing unnecessary lung punctures, imaging re-examinations and other operations, reducing the waste of medical resources, and improving the efficiency of the lung cancer diagnosis process.

[0151] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing a lung cancer knowledge graph by integrating bibliometrics and clinical data, characterized in that: It comprises the following steps: (1) Mining high-frequency and emerging diagnostic factors from scientific literature through a multi-dimensional algorithm to generate academic prior weights; Wherein the input is lung cancer related literature in the global authoritative literature database PubMed, PMC, CNKI, patent library; the output is a set of literature candidate factors, including the W lit (f) value, topic label, time trend, denoted as data set D1; (2) Corresponding the literature candidate factors with real patient data, structured matching is performed to obtain the aligned clinical candidate factor set; Wherein the input is the clinical multi-modal original data, denoted as data set D0, including image DICOM files, test indicators, pathology reports, gene test results, and doctor's text records, data set D1; The output is the aligned clinical candidate factor set, denoted as data set D2, each factor contains basic attributes, weight attributes, and clinical association attributes; (3) From the clinical candidate factors, the core factors with reliable causal relationship with lung cancer outcome are screened out to provide a causally driven skeleton for the knowledge graph; wherein the input is data set D2, clinical multi-modal feature matrix M, and pathological diagnosis outcome variable Y; the output is the core diagnostic factor set and the optimized causal graph G, the core diagnostic factor set contains the causal strength, confidence score, and priority score of the factors, denoted as data set D3, and the optimized causal graph G only retains the causal edges of the factors and Y in D3; (4) Building an initial knowledge graph based on D3 and G, and dynamically updating the edge weights based on new data; wherein the input is data set D3, causal graph G, evidence source data, and new data; The output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions; (5) Output a clinically interpretable report, including: puncture score P biopsy , driving factor contribution ranking, evidence chain and recommended path. 2.The method of claim 1, wherein the method comprises: The step (1) comprises the following sub-steps: (1.1) Literature collection and preprocessing: obtaining literature metadata in batches through API interface, and forming a structured corpus after de-duplication and standardization processing; All statistical data are analyzed, the count variables are subjected to chi-square test or Fisher's exact test, and the quantitative variables are subjected to independent sample t-test, with p value < 0.05; (1.2) Topic and feature extraction: using BERTopic / LDA topic modeling algorithm to cluster research hotspots and identify potential literature candidate factors; (1.3) Building keyword co-occurrence network: extracting the keywords of each literature, mapping to UMLS / MeSH standard ontology; counting the frequency of any two keywords appearing together in the same literature to form a co-occurrence matrix; taking keywords as nodes and co-occurrence frequency as edge weight, retaining edges with co-occurrence frequency ≥ 3 times, and constructing an undirected co-occurrence network; (1.4) Time trend analysis: counting the keyword appearance frequency by year to form a time series; (1.5) Multi-dimensional weight calculation W lit (f) = a · F(f) + β · C(f) + γ · PR(f) + δ · B(f) Wherein, W lit (f) is the literature comprehensive weight of diagnostic factor f; F(f) is the normalized frequency, and the calculation method is that the total number of occurrences of the keyword f in the literature is divided by the maximum number of occurrences of the keyword in the corpus; C(f) is the co-occurrence network centrality (reflecting the relevance of the terms), and the node degree centrality is used cooccur(f,i) is the co-occurrence frequency of f and the keyword i, and the denominator is the maximum node degree in the network, ensuring that C(f) ∈ [0, 1]); PR(f) is the PageRank reference weight, a literature reference network is constructed, the nodes are literatures, the directed edges represent the reference relationship, and the PageRank algorithm is run to obtain the PR value of each literature; B(f) is the time burst weight; based on the Kleinberg burst detection algorithm, the occurrence frequency of f is counted by year to form a time series T(f) = [t1, t2, …, t n ], the probability of each year belonging to the burst state is calculated, the sum of the probabilities of the last 3 years is taken and normalized to obtain B(f) ∈ [0, 1]; α, β, γ, δ are harmonic parameters, satisfying α+β+γ+δ=1, and are optimized through the joint optimization of field expert experience and data verification to ensure the balance between academic hotspots and reference authority. 3.The method of claim 2, wherein the method comprises: The step (2) comprises the following sub-steps: (2.1) Standardization of clinical data: Structured data processing: normalizing the test indicators in D0, and mapping the diagnostic codes to ICD-10 standard codes; Image data: using PyRadiomics tool to extract imageomics features of nodule regions to generate structured feature matrix; Unstructured text: based on BioBERT pre-training model combined with UMLS medical word table, extracting clinical entities from pathology / radiology reports, with entity recognition F1 value ≥ 0.90; Output: A set of standardized clinical entities, denoted as D 01 , containing entity name, data type, and original value. (2.2) Cross-source semantic alignment: The alignment confidence S is calculated for a literature candidate factor f in D1 and a clinical entity e in D 01 by the following formula align (f,e) S align (f,e) = λ1 · cos sim (v f ,v e )+λ2·Jaccard(T f ,T e )+λ3 • Ontology_dist(O f ,O e ) where cos sim (v f ,v e ) is the cosine similarity between two word vectors; BioBERT model is used to encode f and e respectively to generate 768-dimensional word vectors v f ,v e ; the vector cosine similarity is calculated by , ranging from [-1, 1], and after taking the absolute value, it is mapped to [0, 1]; Jaccard(T f ,T e ) is the Jaccard matching degree of the term set; the relevant term set T f of f is extracted; the relevant term set T e of e is extracted; the Jaccard coefficient is calculated by , ranging from [0, 1]; Ontology_dist(O f ,O e ) is the concept distance in the unified medical ontology; the Unified Medical Language System (UMLS) is selected as the unified ontology, and f and e are mapped to the concept nodes O f and O e in UMLS; based on the hierarchical structure of UMLS, the number of edges of the shortest path between O f and O e is taken as the distance value; the distance value is divided by the maximum possible distance in the ontology to obtain Ontology_dist(O f ,O e ) ∈ [0, 1]; λ1, λ2, λ3 are weight parameters, based on the annotated literature term-clinical entity matching dataset, the semantic alignment error is minimized by grid search; the constraint conditions are λ1+λ2+λ3=1 and λ1≥0.4; high-confidence matching pairs with S align (f,e) ≥ 0.7 are screened to ensure semantic consistency; (2.3) Comprehensive weight calculation: For the matched factors, the literature weight and clinical characteristics are fused to generate a cross-source comprehensive score: W align (c) = μ1 · W lit (f) + μ2 · freq clinical (e) + μ3 · OR(e) where W lit (f) is the fusion, directly taken from the literature weight of factor f in D1, reflecting the comprehensive value of the factor in academic literature; freq clinical (e) is the clinical frequency, the number of cases in which clinical entity e appears in D0 / the total number of cases in D0, normalized to [0, 1]; OR is the association strength, calculated based on case-control analysis of D0, as the ratio of the probability of lung cancer malignancy when the factor is present and the probability of lung cancer malignancy when the factor is not present; μ1, μ2, μ3 are set by clinical experts in combination with clinical needs, satisfying μ1+μ2+μ3=1. 4.The method of claim 3, wherein the method further comprises: The step (3) comprises the following sub-steps: (3.1) Fusion of literature prior causal graph construction: A NOTEARS and DAG-GNN joint algorithm is used to construct the causal graph with the dual objectives of data fitting and literature constraint: Based on the feature matrix M and the outcome Y, the potential causal relationships between factors and Y are learned through data-driven learning of the edge weight matrix W, W ij represents the causal strength of factor i on factor j; W introduced in D2 lit (f) As a prior constraint: factors with high literature support are harder to be sparsified in the optimization. Verify and ensure that the causal graph is a directed acyclic graph (DAG), and exclude the logical contradiction of i→j and j→i; (3.2) Multi-dimensional causal confidence score: For each f→Y causal relationship, the confidence is quantified from four dimensions, and a weighted fusion model is constructed to generate the causal confidence score, and the model formula is as follows: Conf causal (f→Y) = η1·ATE score + ω2·CI width + η3·Lit support + η4·Clinical consistency wherein, the average treatment effect standardized value ATE score , the average influence range of the outcome Y when the factor f exists, is the core index of the size of the causal effect, the calculation method is to perform 1000 times of Bootstrap resampling on the edge f→Y in the causal graph, calculate the average treatment effect ATE each time, and normalize all ATE values to the interval [0, 1]. The larger the ATE is, the stronger the causal driving force of X on Y is; the reciprocal of the confidence interval width CI width Based on the Bootstrap sampling results, the 95% confidence interval of ATE is calculated, the width of the confidence interval is calculated, the reciprocal of the width is calculated, and the reciprocal is normalized to the interval [0, 1]. The higher the CI width , the smaller the statistical fluctuation of the causal effect is, and the more reliable the conclusion is; the literature support degree Lit support Based on the data set D1, the number of literatures supporting the causal relationship between f and Y is counted, and then divided by the number of all literatures related to X or Y, to obtain the proportion, which ranges from 0 to 1. The higher the Lit_support is, the higher the recognition degree of the causal relationship in the academic field is, and the academic credibility of the conclusion is enhanced; the clinical consistency Clinical consistency , measures the universality of the causal relationship in different clinical scenarios. The clinical original data D0 is grouped according to the hospital and year dimensions, the ATE of f→Y is calculated for each group of data, the standard deviation of each group of ATE is calculated, the standard deviation is standardized, and then 1-standardized standard deviation is obtained Clinical consistency ; the higher the Clinical consistency , the more the causal relationship is not a random result of a specific sample, and the more the causal relationship has clinical universality; ω1, ω2, ω 3， ω4 is a weight parameter, which balances the contribution degrees of the four dimensions in the total score and reflects the importance priority of different dimensions. An expert group composed of two oncologists and two statisticians scores based on the core elements of causal inference. (3.3) Core factor priority ranking and screening: For factor f, a multi-dimensional weighted fusion model is constructed to rank the priority formula, and the priority score is obtained: Priority(f) = θ1- Conf causal (f → Y) + θ2- W lit (f) + θ3- AUC improve (f) + θ4 • Clinical freq (f) Wherein: AU Cimprove The AUC of the model after adding the f factor to the lung cancer diagnosis model; the diagnosis model adopts the XGBoost algorithm, the tree depth is 5, the learning rate is 0.1, the iteration number is 100, the training set and the validation set are split from D0 in the proportion of 7:3, the AUC of the validation set containing f model and the AUC of the validation set not containing f model are calculated respectively, and the difference is AUC improve (f), normalized to the interval [0, 1]; Clinical freq is the frequency of occurrence of the factor f in D0; θ1, θ2, θ3, θ4 are expert-set parameters balancing the contribution of the four dimensions in the priority score, reflecting the priority of clinical decision-making for each dimension; a priority threshold of Priority(f) ≥ 0.7 is set, and only factors with a priority ≥ 0.7 are retained as core diagnostic factors. 5.The method of claim 4, wherein the method further comprises: In the step (3.2), the core elements of causal inference include: effect size, statistical reliability, academic support, and clinical applicability. 6.The method of claim 5, wherein the method further comprises: In the step (4), the literature ID comes from the literature database in step (1) to support the original literature ID of the factor relationship in D3; the clinical case identifier comes from D0 in step (2) to verify the factor relationship.

7. The method of claim 6, wherein the method is characterized by: The step (4) comprises the following sub-steps: (4.1) Graph initialization: Based on D3 and G, the initial knowledge graph is built, and the abstract factors and causal relationships are converted into structured graph nodes and edges, which are stored in a Neo4j graph database; the nodes and edges of the knowledge graph follow the following principles: Node definition: define node types according to the clinical types of factors in D3, and each node is associated with basic attributes; Edge Definition: Define edge type based on the type of causal relationship in G. The core attributes of an edge include: Causal Confidence Conf taken from D3, Evidence Source Data, Initial Timestamp causal ​ (4.2) Dynamic update: Based on the newly added data, the edge weight is dynamically updated, and the formula is as follows: W edge (t) = η · W edge (t-1) + (1 - η) · [μ · Causal strength + v • New evidence ] where W edge (t) and W edge (t-1) are the edge weights at the current time t and the last version t-1, respectively; η is a historical weight decay factor that controls the retention proportion of historical weights, balances the stability of old evidence and the timeliness of new evidence, η ∈ [0, 1]; μ and v are new evidence weight parameters, satisfying μ + v = 1; Causal strength is the causal strength of the new data, which is obtained by quantifying the causal strength of the new data through step (3); New evidence is the support degree of the new evidence, and the calculation method is the proportion of the literature supporting the relationship in the new literature + the proportion of the cases verifying the relationship in the new cases, normalized to [0, 1]. 8.The method of claim 7, wherein the method further comprises: In the step (4), the nodes are core diagnostic factors and disease outcomes; the edges are causal or related relationships, marked with real-time updated Wedge(t); the attributes include Confcausal, evidence source, and timestamp associated with the edges; version traceability supports querying the graph state at any time point. 9.The method of claim 8, wherein the method further comprises: The step (5) comprises the following sub-steps: (5.1) Patient data mapping: Based on the semantic alignment of step (2), the features of the patient P are mapped to the nodes of the KG, and the corresponding node association factor values are standardized to f i and the causal confidence Conf i ; (5.2) Puncture score calculation: By fusing patient characteristics, causal confidence, and clinical weight, the puncture necessity score is calculated: P biopsy = σ(∑[β i ·f i ·Conf i ]+β0) where σ uses a sigmoid function to compress the linear output to [0, 1], facilitating clinical interpretation; β i is the contribution coefficient of the factor, quantifying the actual influence weight of each factor on the puncture decision; β0 is the baseline offset set (5.3) Contribution degree and evidence chain generation: Contribution degree calculation: Evidence chain generation: From the KG, the causal path of the factor to the outcome is extracted, and the evidence source in step (4) is associated to form a structured evidence.

10. The bibliometrics and clinical data fusion lung cancer knowledge graph construction device is characterized in that: It includes: A mining module configured to mine high-frequency and emerging diagnostic factors from scientific research literature through multi-dimensional algorithms to generate academic prior weights; Wherein the input is lung cancer related literature in the global authoritative literature database PubMed, PMC, CNKI, patent library; the output is a set of literature candidate factors, including the W lit (f) value, topic label, time trend, denoted as dataset D1; An alignment module configured to structure match the literature candidate factors with real patient data to obtain an aligned clinical candidate factor set; Wherein the input is the clinical multi-modal raw data, denoted as data set D0, including image DICOM files, test indicators, pathology reports, gene test results, and doctor text records, data set D1; The output is the aligned clinical candidate factor set, denoted as data set D2, each factor containing basic attributes, weight attributes, and clinical association attributes; The screening module is configured to screen out core factors having reliable causal relationship with lung cancer outcomes from clinical candidate factors, to provide a causally driven skeleton for the knowledge graph; wherein the input is a data set D2, a clinical multi-modal feature matrix M, and a pathological diagnosis outcome variable Y; The output is a core diagnosis factor set and an optimized causal graph G, the core diagnosis factor set containing causal strength, confidence score, and priority score of the factors, denoted as a data set D3, and the optimized causal graph G retaining only the causal edges of the factors and Y in D3; The building module is configured to build an initial knowledge graph based on D3 and G, and to dynamically update edge weights based on new data; wherein the input is the data set D3, the causal graph G, evidence source data, and new data; The output is a dynamic knowledge graph KG, which contains nodes, edges, attributes, and versions. an output module configured to output a clinically interpretable report including: a puncture score P biopsy , a driving factor contribution ranking, a chain of evidence, and a recommended pathway.

Images