A tumor patient comprehensive risk stratification method and system based on data analysis

Abstract

This invention discloses a comprehensive risk stratification method and system for cancer patients based on data analysis. First, it extracts potential new nodes and their relationships from a pre-set database, traversing the paths between core nodes and potential new nodes to form a set of connection paths. Then, it calculates the shortest path length and the number of paths to obtain a connection density score, achieving a transformation from qualitative association to quantitative screening. Next, using a connection density score higher than a pre-set threshold as an entry condition, it extracts keyword vectors from the screened nodes, and calculates corrected semantic scores using semantic drift analysis to quantify semantic biases caused by differences in data sources and knowledge version updates. Finally, it integrates the screened nodes into a knowledge graph to form an enhanced knowledge graph, and outputs the comprehensive risk stratification results through path matching and weighted scoring.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph analysis, and in particular to a comprehensive risk stratification method and system for cancer patients based on data analysis. Background Technology

[0002] Comprehensive risk stratification of cancer patients is an important application area of ​​clinical decision support systems. Traditional risk stratification methods are usually based on fixed rules or single statistical models, directly correlating a patient's clinical indicators with the final prognostic event. In particular, for the risk prediction of complications in special patients, namely cancer patients, monitoring data lacks a systematic quantitative assessment of the correlation path between new clinical data and existing prognostic and complication knowledge. This leads to a decline in the model's predictive ability for complications and an inability to determine whether newly emerging complication nodes are worth including in the risk assessment model.

[0003] On the other hand, medical guidelines recording complications are updated regularly. Traditional methods rely solely on text similarity matching in the semantic screening stage, neglecting the reliability differences of the data sources themselves and the timeliness of medical knowledge. Different data sources have different ways of describing and recording the same risk factor, and many institutions' medical guidelines for cancer patients must be updated frequently. The definition and clinical significance of the same risk factor may change significantly between different versions, but traditional methods fail to quantify the degree of deviation of the data collection time relative to the guideline iteration cycle. This leads to nodes with large source deviations or outdated data being incorrectly identified as highly relevant nodes. In particular, when calculating the baseline drift value, traditional methods usually only process based on the preset offset value of the data source, failing to take into account factors affecting data quality, such as the qualification level of the data entry personnel, so that the baseline drift value cannot fully reflect the overall conceptual deviation risk of the node data in the source dimension. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive risk stratification method and system for cancer patients based on data analysis, which solves the above-mentioned technical problems pointed out in the prior art.

[0005] This invention provides a data analysis-based comprehensive risk stratification method for cancer patients, comprising the following steps:

[0006] Extract potential new nodes and their associated relationship data from the preset database. Based on the associated relationship data, traverse the paths between core nodes and potential new nodes in the existing knowledge graph corresponding to the preset database to obtain a set of connection paths between potential new nodes and core nodes.

[0007] The shortest path length from the potential new node to the core node is calculated based on the set of connection paths, and the connection density score of the potential new node is determined based on the shortest path length and the number of paths.

[0008] If the connection density score is higher than a preset threshold, keyword vectors are extracted from the attribute description text of the potential new node, and the semantic relevance score between the keyword vectors and the core topic vectors in the existing graph is calculated using the TF-IDF algorithm combined with semantic drift analysis to obtain a corrected semantic score; if the corrected semantic score is higher than the preset semantic relevance threshold, the potential new node passes the screening.

[0009] The potential new nodes selected through screening and the relationships between these nodes and key nodes in the process are integrated into the knowledge graph to form an enhanced knowledge graph centered on the correlation between complications and prognosis. Based on the enhanced knowledge graph, the comprehensive risk stratification results of the patients to be analyzed are output.

[0010] Preferably, the TF-IDF algorithm combined with semantic drift analysis is used to calculate the semantic relevance score between the keyword vector and the existing core topic vector of the graph, and the corrected semantic score is obtained. The steps include the following:

[0011] Obtain the attribute description text of potential new nodes and perform word segmentation to obtain an initial keyword set;

[0012] The TF-IDF algorithm is used to vectorize the initial keyword set to obtain keyword vectors.

[0013] Obtain the attribute description text of the core topic vector of the existing graph and the related nodes within one hop range of the core topic vector of the existing graph; then, perform word segmentation on the attribute text of each node and vectorize it using the TF-IDF algorithm to obtain the semantic vector of each node; finally, perform a weighted average of the semantic vectors of all core nodes and related nodes, and obtain the initial semantic relevance score by calculating the cosine similarity between the keyword vector and the core topic vector of the existing graph.

[0014] Obtain the data source identifier of potential new nodes, and query the preset semantic drift table based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier;

[0015] The data collection timestamps for potential new nodes, the medical guideline update timestamps for the risk factor types to which potential new nodes belong, and the record time difference values ​​for the risk factor types to which potential new nodes belong in different data sources are obtained. The time-lapse factor is calculated by combining the collection timestamps, medical guideline update timestamps, and record time difference values.

[0016] Multiply the baseline drift value by the time-sensitivity drift factor to obtain the semantic drift amount;

[0017] The initial semantic relevance score is attenuated and corrected using the semantic drift, resulting in the corrected semantic score.

[0018] Preferably, the attribute description text includes description text of risk factors related to complications and description text of risk factors related to baseline status.

[0019] Preferably, the timeliness drift factor is calculated as follows: For the medical guideline dimension, obtain the medical guideline version index 'a' and update cycle index 'i', calculate the absolute difference between the data collection timestamp corresponding to the current potential new node and the weighted average of all key time nodes in the 'i' update cycle of the 'a'th medical guideline version, divide by the time span of the update cycle to obtain the normalized time deviation for that version and that cycle, multiply by the preset weight coefficient corresponding to that version and that cycle, and sum the above products for all versions and all cycles to obtain the guideline dimension drift value; For the data source dimension, obtain the data source index 'b' and record batch index 'i', calculate the absolute difference between the data collection timestamp and the weighted average of all record times in the 'i'th record batch of the 'b'th data source, divide by the time span of the record batch to obtain the normalized time deviation for that source and that batch, multiply by the preset weight coefficient corresponding to that source and that batch, and sum the above products for all sources and all batches to obtain the data source dimension drift value; add the guideline dimension drift value and the data source dimension drift value to obtain the timeliness drift factor.

[0020] Preferably, the data source identifier of the potential new node is obtained, and a preset semantic drift table is queried based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier, including the following steps:

[0021] Extract the data source field from the metadata of potential new nodes as a data source identifier. The data source identifier includes the electronic medical record system identifier, the laboratory information system identifier, or the image archiving system identifier.

[0022] Query the preset semantic drift scale to obtain the initial drift value corresponding to the data source identifier. The semantic drift scale pre-stores the degree of conceptual shift caused by inconsistencies in the expression of the same medical term among different medical data sources.

[0023] Obtain the qualification level of data entry personnel for potential new nodes, and query the preset personnel confidence table based on the qualification level to obtain the personnel confidence coefficient;

[0024] The baseline drift value is obtained by weighting and summing the initial drift value with the personnel confidence coefficient.

[0025] Preferably, the data source identifier of the potential new node is obtained, and a preset semantic drift table is queried based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier, including the following steps:

[0026] Obtain all data source identifiers corresponding to potential new nodes, form a source identifier set, and query the preset semantic drift table for each source identifier to obtain the initial drift value corresponding to each source;

[0027] If the source identifier set contains only one source identifier, then the initial drift value of that source is directly used as the baseline drift value; if it contains multiple source identifiers, then each source identifier is sorted from high to low according to its corresponding initial drift value to form a first sorted list, and the first source with the highest initial drift value and the second source with the lowest initial drift value are extracted.

[0028] The record fragments from the first source and the second source in the attribute description text of the potential new node are obtained respectively, and the record features of the record fragments are extracted by natural language processing tools.

[0029] The record features from the first and second sources are used to determine conflicts, and then a correction and adjustment process is performed to obtain the baseline drift value.

[0030] Preferably, conflict determination is performed on the recording features from the first and second sources, and a correction and adjustment process is combined to obtain the baseline drift value, including the following steps:

[0031] When there is a conflict between the record features of the first source and the second source, the first historical record set of the first source is obtained and the consistency score of the first historical record set within the first source is calculated; and the second historical record set of the second source is obtained and the consistency score of the second source is calculated.

[0032] The consistency scores within the first and second sources are analyzed and processed using preset high consistency thresholds and preset low consistency thresholds to obtain the first candidate drift value and the second candidate drift value, respectively.

[0033] The first candidate drift value and the second candidate drift value are cross-compared, the deviation between the two is calculated, and the temporary reference drift value is obtained by analyzing the deviation between the two and the preset acceptable deviation range.

[0034] The temporary baseline drift value is compared with each candidate drift value in the source identifier set to calculate the deviation; based on the deviation and a preset deviation tolerance threshold, the baseline drift value is obtained through iterative trend analysis.

[0035] Preferably, conflict determination refers to determining whether the recording time granularity is consistent, or whether the numerical units are uniform, or whether the recording personnel roles are consistent, or whether the number of overlapping keyword sets is lower than a preset overlap threshold.

[0036] In another aspect, the present invention also provides a comprehensive risk stratification system for cancer patients based on data analysis, including a path extraction module, a connectivity density analysis module, a screening module, and an enhanced stratification module;

[0037] The path extraction module is used to extract potential new nodes and their associated relationship data from a preset database, and to traverse the paths between core nodes and potential new nodes in the existing knowledge graph corresponding to the preset database based on the associated relationship data, thereby obtaining a set of connection paths between potential new nodes and core nodes.

[0038] The connection density analysis module is used to calculate the shortest path length from the potential new node to the core node based on the set of connection paths, and to determine the connection density score of the potential new node based on the shortest path length and the number of paths.

[0039] The filtering module is used to extract keyword vectors from the attribute description text of potential new nodes if the connection density score is higher than a preset threshold, and to calculate the semantic relevance score between the keyword vectors and the core topic vectors in the existing graph using the TF-IDF algorithm combined with semantic drift analysis to obtain a corrected semantic score; if the corrected semantic score is higher than the preset semantic relevance threshold, the potential new node passes the filtering.

[0040] The enhanced stratification module is used to integrate the potential new nodes that have passed the screening and the relationships between the potential new nodes and the key nodes in the process into the knowledge graph, forming an enhanced knowledge graph with the correlation between complications and prognosis as the core. Based on the enhanced knowledge graph, the comprehensive risk stratification results of the patients to be analyzed are output.

[0041] In another aspect, the present invention also provides a storage medium comprising a stored program, wherein, when the program is executed, the device on which the storage medium is located executes the aforementioned data analysis-based comprehensive risk stratification method for cancer patients.

[0042] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0043] Analysis of the data analysis-based comprehensive risk stratification method and system for cancer patients provided by this invention reveals that, in practical application, firstly, candidate nodes to be evaluated are acquired. Potential new nodes and their associations with existing knowledge graphs are extracted from a pre-set database. Paths between core nodes and potential new nodes are traversed to form a set of connection paths, providing structured network data for subsequent quantitative analysis. Then, these path relationships are transformed into comparable quantitative indicators. A connection density score is obtained by calculating the shortest path length and the number of paths, reflecting the degree of association between potential new nodes and core prognostic events. A higher score indicates that the node should be included in the knowledge graph, achieving a transformation from qualitative association to quantitative screening. Finally, whether the connection density score is higher than a pre-set threshold is used as an entry condition. The process involves ensuring that the nodes to be analyzed possess basic prognostic correlation strength. Then, keyword vectors are extracted from the selected nodes, and semantic drift analysis is used to calculate corrected semantic scores. The semantic bias caused by differences in data sources and knowledge version updates is quantified by multiplying the baseline drift value by the timeliness drift factor, ensuring that only nodes with strong semantic relevance and good data timeliness can enter the final integration stage. Finally, the selected nodes and their relationships are integrated into a knowledge graph, forming an enhanced knowledge graph with clinical treatment records as the starting point, complication events as the hub, and prognostic risk as the endpoint. Based on this, path matching is performed for the patients to be analyzed, and a weighted scoring method is used to calculate the comprehensive risk value and output stratified results according to thresholds, so that the output can reflect the complete causal relationship from clinical treatment to complications to prognosis. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the main process of a comprehensive risk stratification method for cancer patients based on data analysis.

[0045] Figure 2 This is a schematic diagram simulating potential new nodes in a comprehensive risk stratification method for cancer patients based on data analysis.

[0046] Figure 3 This is a schematic diagram simulating the relationship in a comprehensive risk stratification method for cancer patients based on data analysis.

[0047] Figure 4 This is a schematic diagram simulating the association path in a comprehensive risk stratification method for cancer patients based on data analysis.

[0048] Figure 5 This is a schematic diagram of an enhanced knowledge graph simulation in a data-driven comprehensive risk stratification method for cancer patients.

[0049] Figure 6 This is a schematic diagram of the overall architecture of a comprehensive risk stratification system for cancer patients based on data analysis.

[0050] Figure reference numerals: Path extraction module 10, Connectivity density analysis module 20, Filtering module 30, Enhanced stratification module 40. Detailed Implementation

[0051] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0052] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0053] Example 1

[0054] like Figure 1 As shown, Embodiment 1 of the present invention provides a comprehensive risk stratification method for cancer patients based on data analysis, including the following steps:

[0055] Step S10: Extract potential new nodes from the preset database (e.g., Figure 2 As shown, Figure 2 In the diagram, nodes connected to the existing knowledge graph by dashed lines are potential new nodes. Potential new nodes refer to newly acquired patient clinical characteristic data points not yet included in the existing knowledge graph, such as newly recorded specific complications or adverse events like "Grade 3 myelosuppression," "immune-related pneumonia," "postoperative bleeding," "liver injury," and "gastrointestinal reactions." These nodes also include their associated relationship data (relationship data refers to the semantic or clinical association between potential new nodes and existing nodes in the existing knowledge graph, such as the causal relationship between "myelosuppression" and "chemotherapy cycle," or the precipitating relationship between "septic shock" and "immunotherapy," etc.). Figure 3 The association data of the core nodes (as shown) and potential new nodes are obtained by traversing the existing knowledge graph corresponding to the preset database based on the association data. The core nodes (core nodes refer to key nodes in the knowledge graph that are directly related to the patient's prognosis (prognosis is a medical term that refers to the prediction and assessment of the future development of the disease), including but not limited to nodes that reflect poor prognosis such as "death", "recurrence", "metastasis" and "serious complications") and potential new nodes, and the set of connection paths between potential new nodes and core nodes is obtained.

[0056] The pre-defined database refers to a database of clinical characteristics and prognosis of cancer patients, with clinical data related to the diagnosis and treatment process as its core. This database stores patients' diagnostic information, pathological stage, gene testing results, records of various clinical treatments (such as chemotherapy, radiotherapy, immunotherapy, and targeted therapy), records of complications that occurred during the diagnosis and treatment process (such as infection, bleeding, thrombosis, bone marrow suppression, immune-related pneumonia, etc.), prognostic follow-up data (such as survival status and recurrence), and the correlation between the above data.

[0057] Step S20: Calculate the shortest path length from the potential new node to the core node based on the set of connection paths (wherein, the shortest path length is calculated using the breadth-first search algorithm, starting from the potential new node, traversing adjacent nodes layer by layer until the core node is visited for the first time, which is the shortest path length).

[0058] The number of paths is counted using a depth-first search algorithm. Within a limited search depth (no more than 10 hops), the number of different paths that can lead from potential new nodes to core nodes is counted. The number of paths (referring to the number of paths in the knowledge graph from potential new nodes to core nodes (core nodes, such as death, recurrence, metastasis, and serious complications, are key nodes related to prognosis) is also counted.

[0059] In path statistics, these core nodes may have associated pathways with complication nodes (such as severe myelosuppression) in the complication database, such as... Figure 4 As shown, however, the core node itself is limited to prognostic nodes; the shortest path length refers to the minimum number of edges traversed from a potential new node to a core node, reflecting the density between nodes; the number of paths refers to the total number of all possible connected paths between a core node and a potential new node, reflecting the multiple associations between nodes), and the connection density score of the potential new node is determined based on the shortest path length and the number of paths (calculated as follows). This score quantifies the average reachability of potential new nodes to core prognostic nodes via multiple paths. A higher score indicates stronger connectivity between the node and the core node in the network, suggesting a stronger association between the node and the core prognostic event, and thus a greater likelihood of inclusion in the graph. The number of paths refers to the total number of all connected paths from the potential new node to the core node (the search depth is limited to no more than 10 hops to avoid weakening the association due to excessively long paths). The shortest path length refers to the minimum number of edges traversed between them. Adding 1 to the denominator prevents division by zero when the shortest path length is 0, and also ensures the monotonicity of the score decreasing with increasing path length. This score reflects the tightness of the association between potential new nodes and core prognostic nodes. The connection density score is introduced because the number of paths reflects the multiple associations between a node and the core prognostic event; more paths indicate that the node influences the prognosis through multiple pathways, thus deserving a higher score. The shortest path length reflects the directness of the association between the node and the core prognostic event; shorter paths indicate a more direct association, thus deserving a higher score.

[0060] For example, in a knowledge graph, if the shortest path length from the potential new node "severe myelosuppression" to the core node "serious complications" is 2 ("severe myelosuppression—infection—serious complications"), and the number of paths is 3 (including "severe myelosuppression—infection—serious complications", "severe myelosuppression—delayed chemotherapy—serious complications", "severe myelosuppression—prolonged hospitalization—serious complications"), then the connection density score = 3 / (2+1) = 1.0.

[0061] Step S30: If the connection density score is higher than (otherwise, remove the potential new node) a preset threshold (this threshold is used to screen nodes closely associated with the core prognostic event, and can be configured according to the clinical scenario, generally set to 0.3 (the basis for setting this threshold is: select nodes known to be strongly associated with the core prognostic event in historical data (such as 'septic shock') as a positive sample set, calculate their connection density scores, and take the 25th percentile of the positive sample set scores as the lower limit of the threshold to ensure that the screened nodes have a similar degree of association as the known strongly associated nodes). When the connection density score is lower than 0.3, it indicates that the node has a weak association with the core prognostic event and it is not recommended to proceed to the subsequent semantic screening stage), then select the potential new node attribute description text (attribute description text refers to the text information describing the specific clinical characteristics of the node).

[0062] For example, the attribute description text for the "myelosuppression" node might be "On day 7 after chemotherapy, the patient's absolute neutrophil count dropped to 0.5 × 10⁻⁶". 9Keyword vectors are extracted from " / L, accompanied by fever") and analyzed using the TF-IDF algorithm combined with semantic drift (semantic drift specifically refers to the degree of deviation in the semantic representation of the same clinical concept caused by factors such as differences in data sources, different recording habits, and knowledge version updates; its quantitative value is calculated by multiplying the baseline drift value and the time-related drift factor). The semantic relevance score between the keyword vectors and the core topic vectors in the existing knowledge graph (core topic vectors are high-dimensional semantic vectors extracted from the attribute description text of core nodes (such as "relapse", "metastasis", "death") and their associated nodes in the knowledge graph, used to characterize the semantic features of the core prognostic topics focused on by the graph) is obtained to obtain a corrected semantic score. If the corrected semantic score is higher than the preset semantic relevance threshold (generally set to 0.6), the potential new node passes the screening; otherwise, the potential new node is removed.

[0063] Step S40: Integrate the potential new nodes (mainly nodes related to complications that occur during the diagnosis and treatment process) that have passed the screening in Step S30 and their associations with key process nodes (key process nodes refer to clinical treatment records, complications, and prognostic events) into the knowledge graph to form an enhanced knowledge graph with the association between complications and prognosis as the core (the enhanced knowledge graph includes the node set of the original knowledge graph, a node subset consisting of potential new nodes that have passed the screening, and an edge set connecting the node subset with the original core nodes). Based on the enhanced knowledge graph, output the comprehensive risk stratification results of the patients to be analyzed.

[0064] It should be noted that, after completing the screening in steps S10 to S30 of the above-described embodiments of this application, a set of potential new nodes that are semantically related to the core prognostic events and have reliable data timeliness are obtained. At this point, these nodes and their relationships with existing atlas clinical treatment record nodes (such as chemotherapy cycles, radiotherapy sites), complication nodes, and prognostic event nodes are used to form an enhanced knowledge graph (such as...) starting with clinical treatment records, using complication events as the hub, and ending with prognostic risk. Figure 5 (as shown)

[0065] Then, for the patients to be analyzed, the system performs matching and path queries in the enhanced knowledge graph based on their clinical data. For example, if a new node appears for a patient (such as "severe myelosuppression"), the graph will automatically retrieve the associated paths between this node and nodes such as "infection risk," "treatment interruption risk," and "poor prognosis." Combining preset rules such as path length and association strength, a weighted scoring method is used to calculate the patient's comprehensive risk value. Specifically, the weight of each path from the patient's clinical node to the core prognostic node is calculated as: weight = association strength coefficient × (1 / (path length + 1)). The weights of all paths are summed to obtain the comprehensive risk score. Based on the scoring thresholds (e.g., low risk: <0.5, medium risk: 0.5-1.0, high risk: >1.0), stratified results (e.g., low risk, medium risk, or high risk) are output and presented to clinical decision-makers in the form of a structured report or a visual interface.

[0066] The above-described embodiment of the present application first obtains the patient's clinical treatment records and complication data from the database (step S10), and through path analysis (step S20) and semantic and timeliness dual screening (step S30), identifies complication nodes closely related to core prognostic risks. Finally, these nodes are integrated into a knowledge graph to form an enhanced knowledge graph with the correlation between complications and prognosis as the core, and outputs the patient's comprehensive risk stratification results based on the graph (step S40).

[0067] The above-described embodiments of this application first obtain candidate nodes to be evaluated, extract potential new nodes and their association with existing knowledge graphs from a preset database, traverse the paths between core nodes and potential new nodes to form a set of connection paths, and provide structured network data for subsequent quantitative analysis;

[0068] Then, the above path relationships are transformed into comparable quantitative indicators. By calculating the shortest path length and the number of paths, a connection density score is obtained to reflect the degree of correlation between potential new nodes and core prognostic events. The higher the score, the more the node should be included in the graph, thus realizing the transformation from qualitative correlation to quantitative screening. Next, the connection density score is first used as an entry condition to ensure that the nodes to be analyzed have a basic prognostic correlation strength.

[0069] Subsequently, keyword vectors were extracted from the selected nodes, and semantic drift analysis was used to calculate corrected semantic scores. The semantic bias caused by differences in data sources and knowledge version updates was quantified by multiplying the baseline drift value by the timeliness drift factor, ensuring that only nodes with strong semantic relevance and good data timeliness could enter the final integration stage. Finally, the selected nodes and their relationships were integrated into the knowledge graph to form an enhanced knowledge graph with clinical treatment records as the starting point, complication events as the hub, and prognostic risks as the endpoint. Based on this, path matching was performed for the patients to be analyzed, and a weighted scoring method was used to calculate the comprehensive risk value and output stratified results according to thresholds, so that the output could reflect the complete causal relationship from clinical treatment to complications and then to prognosis.

[0070] Specifically, in step S30, the TF-IDF algorithm combined with semantic drift analysis is used to calculate the semantic relevance score between the keyword vector and the existing core topic vector of the graph, and the corrected semantic score is obtained. This includes the following steps:

[0071] Step S31: Obtain the attribute description text of potential new nodes and perform word segmentation to obtain an initial keyword set. The attribute description text includes description text of complication-related risk factors and description text of baseline status-related risk factors. Among them, complication-related risk factors include bone marrow suppression indicators corresponding to chemotherapy cycles and radiotherapy complication indicators (for example, bone marrow suppression indicators corresponding to chemotherapy cycles, such as the absolute neutrophil count dropping to 0.5 × 10⁻⁶ on day 7 after the patient receives the second cycle of chemotherapy). 9 / L, diagnosed as "grade 3 myelosuppression"; post-radiotherapy complication indicators, such as radiation pneumonitis 3 months after the patient completed thoracic radiotherapy, with imaging findings of bilateral interstitial pulmonary infiltrates, accompanied by dry cough and low-grade fever; baseline status-related risk factors include performance status score and abnormal laboratory test indicators (performance status score, such as the patient's ECOG score of 3, daytime bed rest time exceeding 50%, only able to perform limited self-care; abnormal laboratory test indicators, such as the patient's creatinine clearance rate of 45 ml / min, below the safe threshold of 60 ml / min for chemotherapy drugs, indicating mild renal impairment);

[0072] Step S32: Use the TF-IDF algorithm to vectorize the initial keyword set to obtain keyword vectors;

[0073] Step S33: Obtain the existing knowledge graph core topic vector (the existing knowledge graph core topic vector is constructed as follows: first, obtain the attribute description text of all core nodes (such as "death", "relapse", "metastasis", "serious complications") and their related nodes within one hop range in the knowledge graph; then, perform word segmentation on the attribute text of each node and vectorize it using the TF-IDF algorithm to obtain the semantic vector of each node; finally, perform a weighted average of the semantic vectors of all core nodes and related nodes (the weight of core nodes is 0.7 and the weight of related nodes is 0.3) to obtain the core topic vector), and obtain the initial semantic relevance score by calculating the cosine similarity between the keyword vector and the existing knowledge graph core topic vector;

[0074] Step S34: Obtain the data source identifier of potential new nodes, and query the preset semantic drift table based on the data source identifier (the semantic drift table is a comparison table containing a data source identifier field and an initial drift value field; for example, due to differences in doctors' writing habits, the electronic medical record system (EMR) may describe 'bone marrow suppression' as 'low white blood cell count', 'neutropenia', etc., leading to conceptual deviation, and its corresponding initial drift value is set to 0.15; the laboratory information system (LIS) directly outputs the standardized value 'absolute neutrophil count = 0.5 × 10⁻⁶'). 9 / L', with a small offset, corresponding to an initial drift value set to 0.05), yields the baseline drift value corresponding to the data source identifier;

[0075] Step S35: Obtain the data collection timestamp of the potential new node, the medical guideline update timestamp of the risk factor type to which the potential new node belongs, and the record time difference value of the risk factor type to which the potential new node belongs in different data sources. Combine the data collection timestamp, the medical guideline update timestamp, and the record time difference value to perform a composite calculation to obtain the time drift factor.

[0076] ;

[0077] In the formula, For time-lapse factor, Indicates the index of medical guideline versions. This indicates that the index is being updated periodically. This indicates the timestamp of data collection corresponding to the potentially newly added nodes. This represents the weighted average of all key time points (including the release date of this version of the guideline, intermediate revision dates within this period, and the end date of this period) in the i-th update cycle of the a-th version of the medical guideline. The weights are determined based on the citation frequency of each time point in clinical practice; the higher the citation frequency, the greater the weight. This represents the time span of the i-th update cycle for the a-th version of the medical guideline. Preset weighting coefficients are used to differentiate the importance of different guideline versions and update cycles; newer versions and more recent update cycles are assigned higher weights. For example, for the same guideline version A (2021-2024), its first update cycle (2021-2022) has a higher weight. Set to 0.3, weight for the second update cycle (2023-2024). (Set to 0.7 to reflect the greater impact of more recent update cycles on current clinical practice). Indicates the data source index. This represents the weighted average of all record timestamps (including the timestamp of each record in the batch) in the i-th record batch from the b-th data source. The weights are determined based on the completeness of the record entries; the higher the completeness, the greater the weight. This represents the time span of the i-th record batch from the b-th data source. Assign pre-defined weighting coefficients (to differentiate the importance of different data sources, assigning higher weights to data sources with standardized records and reliable quality. In the example, the electronic medical record system is weighted at 0.5, the laboratory information system at 0.3, and the image archiving system at 0.2).

[0078] Time-drift factor The introduction of this method aims to address the issue of inconsistent knowledge timeliness in risk stratification for cancer patients. Specifically, cancer treatment guidelines (such as NCCN and CSCO) are updated regularly. The definition, threshold, or clinical significance of the same risk factor (such as chemotherapy regimen intensity) may change in different guideline versions and even within the same version's update cycle. For example, the definition of high-intensity chemotherapy in the 2020 guidelines may differ from that in the 2024 guidelines. Furthermore, different medical data sources (electronic medical record systems, laboratory information systems) exhibit varying accuracy in timestamps and time offsets in recording the same risk factor. For instance, records in imaging archives may be date-based, while laboratory data may be accurate to the minute. Moreover, simply calculating the difference between the collection timestamp and the guideline update timestamp only reflects a single dimension of impact, ignoring the differentiated importance of different time points (such as start, middle, and end) within the same update cycle, and failing to reflect the differences in recording habits among different data sources. Therefore, a composite calculation method is needed that simultaneously considers the iteration cycle of medical guidelines and the recording characteristics of data sources to accurately quantify the drift of the data collection time point relative to the knowledge update node.

[0079] In the above calculation process, through and The two-dimensional separation balances the impact of knowledge updates and data collection habits by decomposing drift into guideline-dimensional drift and data source-dimensional drift, calculating them separately and then summing them to avoid any one dimension dominating the result. A weighted average is then applied. Taking into account all key time points within the cycle, each update cycle has multiple key time points such as start, middle, and end. The weighted average can reflect the combined impact of these points on the knowledge definition, rather than relying solely on endpoint normalization (divided by). To make drift over different time spans comparable, a 3-month update cycle and a 2-year update cycle produce different drift effects from the same time skew. Normalization eliminates the influence of differences in time spans by weighted summation. Distinguishing the importance of different guideline versions and update cycles: The latest guideline version and the most recent update cycle should be given higher weight. To achieve this differentiated processing, the data source dimension is similarly designed to quantify the impact of time offset on batch records. Considering the time offset characteristics of all records within this batch, Normalization makes the drift of different record batches comparable;

[0080] The calculations in the above embodiments of this application incorporate multiple factors, such as the version iteration of medical guidelines, key time points within the update cycle, and batch offsets of data sources, into a unified framework to obtain a continuous and comparable drift value. This value reflects that the further the data collection time is from the latest guideline update node, the greater the drift; the more irregular the data source recording habits, the greater the drift; and the greater the differences between different guideline versions, the greater the drift. The semantic drift is multiplied by the baseline drift value as a decay factor to obtain the final semantic drift amount. Then, the initial semantic relevance score is attenuated and corrected. This means that the semantic relevance of outdated data (out of touch with guideline knowledge) will be appropriately reduced, and the semantic relevance of data from unreliable sources will also be appropriately reduced, avoiding incorrect determination of node relevance due to the use of outdated or biased data. By explicitly introducing interpretable variables such as guideline version index a, update cycle index i, and data source index b, the calculation results of the drift factor have clear clinical and technical meaning.

[0081] Weighting coefficient and The method can be flexibly adjusted according to the actual application scenario (such as different cancer types and different data source quality) to improve its applicability in different medical environments. The exclusion or retention of a certain node is partly due to its data timeliness drift being too large or acceptable.

[0082] By precisely quantifying the timeliness drift, it is ensured that only nodes with strong semantic relevance and good data timeliness can pass the screening in S30. The screened nodes are then integrated into the knowledge graph in the subsequent step S40 to output the comprehensive risk stratification results. The final output stratification results have a data foundation that is both semantically relevant and timely.

[0083] For example, suppose the data collection timestamp of a certain node As of May 2024, the medical guidelines have two versions: Version A has an update cycle spanning 3 years (2021-2024), with a weighted average time point T* of June 2023; Version B has an update cycle spanning 2 years (2019-2021), with a weighted average time point T* of August 2020. If λ is 0.7 and 0.3 respectively, then...

[0084] Guideline dimensional drift = 0.7 × |2024-2023| / 3 + 0.3 × |2024-2020| / 2 = 0.7 × 0.33 + 0.3 × 2 = 0.231 + 0.6 = 0.831.

[0085] The weight of each key time point in the weighted average is determined as follows: the annual citation frequency of that time point in the preset clinical guideline literature database is obtained, and this frequency is divided by the total citation frequency of all key time points to obtain the normalized weight value. For example, if the publication time has a citation frequency of 100 times, the intermediate revision time has a citation frequency of 20 times, and the end time has a citation frequency of 30 times, then the weights are calculated as 0.67, 0.13, and 0.20, respectively.

[0086] In the weighted average of the above recorded time points, the weight is determined as follows: the weight is determined based on the completeness rate of the required fields in the batch of records. Completeness rate = number of filled fields / total number of required fields, and the weight value is equal to the completeness rate.

[0087] Step S36: Multiply the baseline drift value and the timeliness drift factor to obtain the semantic drift amount (the semantic drift amount comprehensively reflects the superposition effect of the concept deviation risk of the data source itself (baseline drift value) and the data timeliness deviation risk (timeliness drift factor). It is used as a decay factor to correct the semantic relevance score, which means that when the data source is unreliable or the data is outdated, its semantic relevance will be appropriately reduced to avoid incorrect inclusion in the knowledge graph. It represents the degree of confidence decay of the semantics of potential new nodes relative to the core semantics of the knowledge graph).

[0088] Step S37: Using the semantic drift as a decay factor, the initial semantic relevance score is attenuated and corrected to obtain the corrected semantic score (initial semantic relevance score × (1 - baseline drift value × time-related drift factor) where semantic drift = baseline drift value × time-related drift factor. When the semantic drift is close to 0, the corrected semantic score is close to the initial score; when the semantic drift is close to 1, the corrected semantic score approaches 0, indicating that the data no longer has semantic relevance due to source bias or timeliness issues).

[0089] It should be noted that in the above embodiments of this application, firstly, the attribute description text of potential new nodes is obtained and word segmentation is performed to obtain an initial keyword set, transforming unstructured clinical text into computable keyword units; then, the keyword set is vectorized using the TF-IDF algorithm to obtain keyword vectors, transforming keywords into computer-processable numerical vector forms; next, a reference standard is constructed and an initial similarity is calculated. First, the existing graph core theme vector is obtained. From the attribute description text of the core node and its associated nodes within one hop range, word segmentation and TF-IDF vectorization are performed, followed by a weighted average with a core node weight of 0.7 and an associated node weight of 0.3; then, the cosine similarity between the keyword vector and the core theme vector is calculated to obtain an initial semantic relevance score, quantifying the semantic matching degree between potential new nodes and the graph core theme;

[0090] Furthermore, a source reliability factor is introduced to obtain the data source identifiers of potential new nodes. A pre-defined semantic drift scale is then used to obtain the corresponding baseline drift value, which reflects the degree of conceptual shift caused by inconsistent representations from different data sources, providing a basis for subsequent correction based on source dimension shifts. Next, a timeliness factor is introduced to obtain the collection timestamp, medical guideline update timestamp, and recording time difference value. A timeliness drift factor is obtained through composite calculations, comprehensively quantifying the degree of shift in data collection time relative to the knowledge update cycle and the degree of difference in data source recording habits. Then, the above two dimensions are combined, and the baseline drift value and the timeliness drift factor are multiplied to obtain the semantic drift amount, comprehensively reflecting the superimposed effect of source conceptual shift risk and data timeliness shift risk, integrating source reliability and timeliness into a unified attenuation index. Finally, the semantic drift amount is used as an attenuation factor to attenuate and correct the initial semantic relevance score, appropriately reducing the semantic relevance score of nodes with unreliable or outdated data sources, ensuring that only nodes with strong semantic relevance and good data timeliness can pass the screening and enter the subsequent integration stage.

[0091] Specifically, in step S34, the data source identifier of the potential new node is obtained, and a preset semantic drift table is queried based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier, including the following operation steps:

[0092] Step S341: Extract the data source field from the metadata of the potential new nodes as the data source identifier. The data source identifier includes the electronic medical record system identifier, the laboratory information system identifier, or the image archiving system identifier.

[0093] Step S342: Query the preset semantic drift scale to obtain the initial drift value corresponding to the data source identifier. The semantic drift scale pre-stores the degree of conceptual shift caused by inconsistencies in the expression of the same medical term among different medical data sources.

[0094] Step S343: Obtain the qualification level of the data entry personnel for potential new nodes, and query the preset personnel confidence table according to the qualification level to obtain the personnel confidence coefficient;

[0095] Step S345: Weight the initial drift value with the personnel confidence coefficient to obtain the baseline drift value (the calculation formula is: baseline drift value = initial drift value × w1 + personnel confidence coefficient × w2, where w1 and w2 are preset weight coefficients. This weight allocation is based on statistical analysis of 1000 clinical records: the initial drift value of the data source itself has the greatest contribution to the final semantic drift (explaining approximately 50% of the variance), while personnel qualifications have a smaller contribution (approximately 30%). Specific weights can be adjusted locally based on the data quality assessment results of the medical institution).

[0096] It should be noted that the above-described embodiments of this application first obtain the source identifier, extract the data source field from the metadata of the potential new nodes as the data source identifier, and provide a clear index basis for subsequent querying of the semantic drift table; then, obtain the initial drift value according to the source identifier, query the preset semantic drift table, obtain the initial drift value corresponding to the source identifier, and quantify the inherent offset risk of the data source into a basic value.

[0097] Next, the qualification levels of data entry personnel for potential new nodes are obtained, and a preset personnel confidence table is queried to obtain personnel confidence coefficients. The qualification levels of the data entry personnel are converted into coefficients that can be used in the calculation to correct the recording bias caused by personnel differences. Finally, the factors of the above two dimensions are combined, and the initial drift value is weighted and summed with the personnel confidence coefficients to obtain the baseline drift value. The inherent bias of the source and the influence of personnel qualifications are integrated into a unified baseline drift value, so that this value can comprehensively reflect the overall conceptual bias risk of the node data in the source dimension, and provide a more accurate attenuation basis for the subsequent calculation of semantic drift.

[0098] In another embodiment, in the comprehensive risk stratification of cancer patients, the same risk factor (such as "myelosuppression index") for the same patient may appear simultaneously in multiple data sources. For example, the Electronic Medical Record (EMR) system records the doctor's diagnosis text, the Laboratory Information System (LIS) records blood routine values, and the Picture Archiving System (PACS) records imaging examination results. The description methods, recording times, and numerical ranges of this index may differ among different sources, leading to conflicts in the baseline drift values ​​corresponding to each source in the pre-set semantic drift scale (for example, the baseline drift value of EMR is low, but the actual text description is ambiguous; the baseline drift value of LIS is high, but the value is precise). In this case, directly weighting and summing the initial drift values ​​may introduce errors and biases. Therefore, it is necessary to introduce a logical analysis and feedback mechanism that does not rely on numerical calculations.

[0099] Therefore, in another embodiment, in step S34, the data source identifier of the potential new node is obtained, and a preset semantic drift table is queried according to the data source identifier to obtain the baseline drift value corresponding to the data source identifier, including the following operation steps:

[0100] Step S3401: Obtain all data source identifiers corresponding to potential new nodes, form a source identifier set, and query the preset semantic drift table for each source identifier to obtain the initial drift value corresponding to each source;

[0101] Step S3402: If the source identifier set contains only one source identifier, then the initial drift value of that source is directly used as the baseline drift value; if it contains multiple source identifiers, then each source identifier is sorted from high to low according to its corresponding initial drift value to form a first sorted list, and the first source with the highest initial drift value and the second source with the lowest initial drift value are extracted.

[0102] Step S3403: Obtain the record fragments (including the first record fragment and the second record fragment) from the first source and the second source in the attribute description text of the potential new node, respectively, and extract the record features of the record fragments using natural language processing tools (the record features include keyword set, numerical unit, record time granularity, and recorder role (correspondingly, extract the first keyword set, first numerical unit, first record time granularity, and first recorder role of the first record fragment and the second keyword set, second numerical unit, second record time granularity, and second recorder role of the second record fragment)).

[0103] Step S3405: ​​Conflict determination is performed on the recording features of the first source and the second source, and the baseline drift value is obtained by combining correction and adjustment processing.

[0104] It should be noted that the above-described embodiments of this application first identify the multi-source situation, obtain all data source identifiers corresponding to potential new nodes, forming a source identifier set, and query a preset semantic drift table for each source identifier to obtain the initial drift value corresponding to each source, explicitly identifying the coexistence of multiple sources and providing a complete list of data sources for subsequent conflict handling; then, branching processing is performed according to the number of sources. If the source identifier set contains only one source identifier, the initial drift value of that source is directly used as the baseline drift value; if it contains multiple source identifiers, the source identifiers are sorted from high to low according to their corresponding initial drift values ​​to form a first sorting list, and the first source with the highest initial drift value and the second source with the lowest initial drift value are extracted, simplifying the multi-source situation to a focused processing of sources with extreme differences. By locking the two sources with the highest and lowest values, a basis is provided for subsequent... The subsequent conflict assessment provides a clear comparison object. Next, record fragments from the first and second sources in the attribute description text of potential new nodes are obtained respectively, and the record features of these record fragments are extracted using natural language processing tools. The text records from the two sources are transformed into comparable structured features, so that subsequent conflict assessments no longer depend on the initial drift value itself, but are based on the feature differences of the actual recorded content. Finally, conflict assessments are performed on the record features of the first and second sources, and combined with correction and adjustment processing, a baseline drift value is obtained. By analyzing the degree of difference between the two sources in dimensions such as keyword set, numerical unit, recording time granularity, and recording personnel role, it is determined whether there is an actual conflict. Based on the conflict assessment results, the baseline drift value is corrected and adjusted so that the final baseline drift value can reflect the comprehensive result after logical analysis and feedback correction in a multi-source conflict scenario.

[0105] Specifically, in step S3405, conflict determination is performed on the recording features of the first source and the second source, and a correction and adjustment process is combined to obtain the reference drift value, including the following operation steps:

[0106] Step S34051: When there is a conflict between the recording features of the first source and the second source (the judgment of conflict is that (1) the recording time granularity is inconsistent (e.g., the first source is accurate to the minute, while the second source is only accurate to the day); (2) the numerical units are inconsistent (e.g., the first source uses "mg / dL", while the second source uses "mmol / L"); (3) the recording personnel roles are inconsistent (e.g., the first source is a doctor's diagnosis, while the second source is a doctor's assistant's record); (4) the number of overlapping keyword sets is lower than the preset overlap threshold; any one of these conditions is used to determine that there is a conflict), the first source obtains the batches of past records in the historical records that are the same as the risk factor type of the potential new node, forming the first historical record set, and calculates the consistency score of the recording features of each batch of record segments in the first historical record set to obtain the consistency score within the first source; and obtains the second historical record set of the second source, and calculates the consistency score within the second source (the method of obtaining the consistency score within the first source is consistent and will not be repeated).

[0107] It should be noted that, in the embodiments of this application described above, the calculation method for the intra-source consistency score is as follows: For multiple historical record batches from the same source, the record feature vector of each batch is extracted (including four dimensions: keyword set, numerical unit, record time granularity, and recorder role). After converting the keyword set into a vector representation, the cosine similarity between the record feature vectors of adjacent batches is calculated, and the average of the similarities of all adjacent batches is taken as the intra-source consistency score of that source. The score ranges from 0 to 1; a higher score indicates that the record features of that source are more consistent across different batches, and the credibility is higher.

[0108] Intra-source consistency score is used to evaluate the stability of recorded features from the same data source across different historical batches. If the recorded features from the same source are highly consistent across different batches (e.g., stable keyword set, uniform units, and fixed time granularity), the intra-source consistency score is high, indicating that the source records are standardized and highly reliable. Conversely, if the recorded features fluctuate significantly, the intra-source consistency score is low, indicating that the source records are arbitrary and lowly reliable.

[0109] For example, a data source has three historical batches. The keyword set of the first batch is {"bone marrow suppression", "neutrophils", "fever"}, the second batch is {"bone marrow suppression", "neutrophils", "infection"}, and the third batch is {"bone marrow suppression", "white blood cells", "fever"}. The similarity between the first and second batches is 0.67, and the similarity between the second and third batches is 0.5. Therefore, the intra-source consistency score = (0.67 + 0.5) / 2 = 0.585.

[0110] In this system, adjacent batches are defined chronologically. The keyword set is converted into a vector by constructing a global dictionary based on the keywords appearing in all historical batches. The keyword set of each batch is mapped to a binary vector with the same dimension as the dictionary, where 1 indicates the appearance of the corresponding keyword and 0 indicates the absence of the keyword. Cosine similarity is calculated based on this binary vector.

[0111] Step S34052: Compare the consistency score within the first source with a preset high consistency threshold (this threshold is generally set to 0.8; when the consistency score within the source is higher than 0.8, it is marked as a high-confidence source). If the consistency score within the first source is higher than the high consistency threshold, the first source is marked as a high-confidence source, and its initial drift value is retained as the first candidate drift value. If the consistency score within the first source is lower than a preset low consistency threshold (this threshold is generally set to 0.4; when the consistency score within the source is lower than 0.4, it is marked as a low-confidence source), the first source is marked as a low-confidence source, and the initial drift value of the first source is automatically corrected based on the record features that appear most frequently in its historical record set. (Automatic correction means statistically analyzing the set of keywords, numerical units, record time granularity, and recorder roles that appear most frequently in the historical record set of the source, using these as the dominant feature pattern, comparing the current record segment features with the dominant feature pattern, and calculating the feature matching degree M, which is the proportion of the number of matching features between the current record segment features and the dominant feature pattern to the total number of features.) The corrected candidate drift value is: candidate drift value × (2-M). When the matching degree is low (M<0.5), the drift value increases; when the matching degree is high, the drift value decreases, thus obtaining the first candidate drift value; if the consistency score within the first source is between the low consistency score threshold and the high consistency score threshold, then the first source is marked as a medium-credibility source;

[0112] Step S34053: For the first source marked as a medium-credibility source, retrieve the three historical batches in the first historical record set that are most similar to the recorded features in the attribute description text of the potential new node (i.e., calculate the similarity between the recorded features in the attribute description text of the potential new node and each data in the first historical record set of the first source (the similarity can be calculated using cosine similarity, etc., which will not be elaborated in this application), and then select the three historical batches with the highest similarity. Obtain the historical drift values ​​corresponding to these three historical batches to form a first drift value reference set. Use the mode or median of the first drift value reference set as the reference drift value of the first source, and compare the reference drift value with the initial drift value. If the deviation exceeds the preset deviation threshold (this threshold is generally set to 0.2. When the deviation between the reference drift value and the initial drift value exceeds 20%, the reference drift value is used), then the reference drift value is used as the first candidate drift value. Otherwise, the initial drift value is retained as the first candidate drift value (and at the same time, the second candidate drift value is obtained for the second source intra-source consistency score of the second source in the same way as above).

[0113] Step S34054: Cross-compare the first candidate drift value with the second candidate drift value. If the deviation between the two is within a preset acceptable deviation range (generally set to ±0.15; when the deviation between the two candidate drift values ​​is within 15%, it is considered acceptable; this value is based on historical statistical analysis of the differences in baseline drift values ​​from multiple data sources (EMR, LIS, PACS). The analysis shows that for the same clinical concept, the 95% confidence interval for the difference in baseline drift values ​​from different data sources is approximately 0.1-0.2, therefore 0.15 is set as the threshold for determining whether multi-source voting is necessary), then calculate the weighted average of the two as the provisional baseline drift value, where the weighting coefficient is determined based on the consistency within the first source. The ratio of the first source's consistency score to the second source's consistency score is determined. If the deviation exceeds the acceptable deviation range, a third source with the same risk factor type as the potential new node is obtained. The third source is any source other than the first and second sources in the source identifier set. The third record fragment of the third source is extracted from the attribute description text of the potential new node, and the third source's consistency score and third candidate drift value are obtained. The first candidate drift value, the second candidate drift value, and the third candidate drift value are voted on by multiple sources, and the drift value with the highest number of votes is selected as the temporary benchmark drift value. If none of the three sources receive more than half of the votes, the candidate drift value corresponding to the source with the highest consistency score is selected as the temporary benchmark drift value.

[0114] For example, the candidate drift values ​​for the three sources are: EMR 0.2, LIS 0.5, and PACS 0.2. After voting, the drift value 0.2 appears twice and receives the most votes, so the provisional baseline drift value is 0.2. If the three values ​​are 0.2, 0.3, and 0.4 respectively, none of which receive a majority, then the candidate drift value corresponding to the source with the highest intra-source consistency score is selected as the provisional baseline drift value.

[0115] Step S34056: Compare the temporary reference drift value with the candidate drift values ​​of each source in the source identifier set, excluding the selected reference source. If the deviation of any candidate drift value from the temporary reference drift value exceeds a preset deviation tolerance threshold (this threshold is generally set to 0.1; when the deviation of the candidate drift value from the temporary reference drift value exceeds 10%, it is marked as a source to be calibrated), then the source is marked as a source to be calibrated, and the historical drift value change trend of the source is retrieved. If the change trend shows a monotonically increasing or monotonically decreasing regular change, then the candidate drift value of the source is extrapolated and corrected according to the slope of the change trend to obtain the calibrated drift value, and the calibrated drift value is then used. Replace the original candidate drift value and repeat the above operations (S34053 to S34054) until the deviations of all candidate drift values ​​from the current temporary baseline drift value are lower than the deviation tolerance threshold, or the number of iterations reaches the preset maximum number of iterations (this value is generally set to 5 times to avoid infinite loops. To avoid infinite loops, after the number of iterations reaches the preset maximum number of iterations (default 5 times), even if there are still sources with deviations exceeding the tolerance threshold, the iteration will be terminated, and the current temporary baseline drift value will be determined as the final baseline drift value. At this time, the sources that have not converged will be marked as pending manual review and will not be automatically included in the map). The temporary baseline drift value at this time will be determined as the final baseline drift value.

[0116] It should be noted that, in the above embodiments of this application, firstly, when there is a conflict between the record features of the first source and the second source, the historical record batches of the two sources that are the same as the risk factor type of the potential new node are obtained in the historical records, forming their respective historical record sets, and calculating their respective intra-source consistency scores. The conflict identification is combined with the quantification of the stability of each source, providing a basis for subsequent credibility judgment. Then, the sources are classified according to the intra-source consistency scores. The intra-source consistency scores of the first source are compared with the preset high and low consistency score thresholds to obtain the first candidate drift value and low credibility sources. The initial drift value is automatically corrected according to the record features that appear most frequently in the historical record set. The method of determining the candidate drift value is differentiated according to the historical stability of the source, so that the source with standardized records and high credibility retains the original value, and the source with arbitrary records and low credibility reduces its impact through correction.

[0117] Next, reference drift value calibration is performed on the trusted source. The three historical batches most similar to the current record features in the historical record set of the source are retrieved, and the historical drift values ​​corresponding to these three batches are obtained to form a reference set. The mode or median is taken as the reference drift value. This value is compared with the initial drift value. If the deviation exceeds the preset deviation threshold, the reference drift value is used as the candidate drift value. Otherwise, the initial drift value is retained. By introducing the drift values ​​of similar historical cases, the initial drift value of the trusted source is corrected, so that the determination of the candidate drift value not only refers to historical experience but also maintains a reasonable correlation with the initial value.

[0118] Furthermore, the candidate drift values ​​from the first and second sources are cross-compared. If the deviation between the two is within an acceptable range, a weighted average is calculated as a temporary baseline drift value. If the deviation exceeds the acceptable range, a third source is introduced, and the candidate drift values ​​from the three sources are voted on by multiple sources. The one with the most votes is selected as the temporary baseline drift value. If none of them receive more than half the votes, the candidate drift value corresponding to the source with the highest consistency score within the source is selected. The cross-comparison and voting mechanism resolves the conflict between the two extreme sources, ensuring that the determination of the temporary baseline drift value can reflect the majority consensus or optimal credibility in the multi-source information.

[0119] Finally, the temporary reference drift value is compared with the candidate drift values ​​of other sources in the source identifier set. If any deviation exceeds the deviation tolerance threshold, the source is marked as a source to be calibrated. The historical drift value change trend is retrieved. If it shows a monotonically increasing or decreasing regular change, the drift value is extrapolated and corrected according to the slope to obtain the calibrated drift value and replace the original value. The iterative process is repeated until the deviation of the candidate drift value of all sources from the temporary reference drift value is lower than the tolerance threshold or the number of iterations reaches the preset maximum number of iterations. The temporary reference drift value at this time is determined as the final reference drift value. Through the iterative calibration mechanism, the drift values ​​of all sources are gradually converged to the reference value, so that the final determined reference drift value can take into account the historical change patterns of each source. At the same time, the setting of the maximum number of iterations and the manual review mark avoids infinite loops and ensures the termination of the calculation process.

[0120] Example 2

[0121] On the other hand, such as Figure 6 As shown, this second embodiment is based on the comprehensive risk stratification method for tumor patients based on data analysis provided in the first embodiment of the invention, and also provides a comprehensive risk stratification system for tumor patients based on data analysis, including a path extraction module 10, a connection density analysis module 20, a screening module 30 and an enhanced stratification module 40;

[0122] Among them, the path extraction module 10 is used to extract potential new nodes and the association data of potential new nodes from the preset database, and to traverse the paths between core nodes and potential new nodes in the existing knowledge graph corresponding to the preset database according to the association data, so as to obtain a set of connection paths between potential new nodes and core nodes.

[0123] The connection density analysis module 20 is used to calculate the shortest path length from the potential new node to the core node based on the set of connection paths, and to determine the connection density score of the potential new node based on the shortest path length and the number of paths.

[0124] The filtering module 30 is used to extract keyword vectors from the attribute description text of potential new nodes if the connection density score is higher than a preset threshold, and to calculate the semantic relevance score between the keyword vectors and the core topic vectors in the existing graph using the TF-IDF algorithm combined with semantic drift analysis to obtain a corrected semantic score; if the corrected semantic score is higher than the preset semantic relevance threshold, the potential new node passes the filtering.

[0125] The enhanced stratification module 40 is used to integrate the potential new nodes that have passed the screening and the relationship between the potential new nodes and the key nodes in the process into the knowledge graph, forming an enhanced knowledge graph with the correlation between complications and prognosis as the core. Based on the enhanced knowledge graph, the comprehensive risk stratification results of the patients to be analyzed are output.

[0126] Example 3

[0127] On the other hand, this embodiment three, based on the data analysis-based comprehensive risk stratification method for tumor patients provided in embodiment one, also provides a storage medium, the storage medium including a stored program, wherein, when the program is running, it controls the device where the storage medium is located to execute the above-mentioned data analysis-based comprehensive risk stratification method for tumor patients.

[0128] In summary, the comprehensive risk stratification method and system for cancer patients based on data analysis proposed in this invention uses knowledge graph path analysis to calculate connection density scores for the first quantitative screening, and then combines semantic drift analysis for the second semantic and time-sensitive dual screening. After integrating the screened nodes into an enhanced knowledge graph, the risk stratification results are output. This constructs a complete technical chain from path quantitative screening to semantic and time-sensitive refinement to graph enhancement applications, optimizing the overall deficiency of lacking a systematic screening and integration mechanism, and ensuring that the final output risk stratification results are based on data that has been verified in multiple dimensions.

[0129] In the further processing, based on the TF-IDF semantic similarity calculation, the product of the baseline drift value and the timeliness drift factor is introduced as the semantic drift amount to attenuate and correct the initial semantic relevance score. The baseline drift value reflects the conceptual deviation risk of the data source, and the timeliness drift factor reflects the degree of deviation of the data collection time relative to the iteration cycle of the medical guideline. At the same time, the reliability differences of the data source and the timeliness of medical knowledge are ignored, so that the corrected semantic score can effectively filter nodes with unreliable sources or outdated data.

[0130] In the further processing, based on the initial drift value of the data source, the qualification level of the data entry personnel is further obtained, and the personnel confidence coefficient is queried separately. The two are weighted and summed to obtain the benchmark drift value, which optimizes the problem of the single dimension of the benchmark drift value calculation, so that the benchmark drift value can comprehensively reflect the overall conceptual shift risk of the two dimensions of inherent source bias and personnel qualification influence.

[0131] In the further processing, when there are multiple data sources for the same risk factor, the record features of each source are extracted for conflict judgment. The sources are graded according to the consistency score within the source. The final benchmark drift value is determined through cross-comparison, multi-source voting and iterative calibration mechanism. This optimizes the problem of error bias introduced by direct weighted summation in multi-source conflict scenarios, so that the determination of the benchmark drift value can reflect the majority consensus or optimal credibility in multi-source information. Iterative calibration ensures the termination of the calculation process.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A comprehensive risk stratification method for cancer patients based on data analysis, characterized in that, The following steps are included: Extract potential new nodes and their associated relationship data from the preset database. Based on the associated relationship data, traverse the paths between core nodes and potential new nodes in the existing knowledge graph corresponding to the preset database to obtain a set of connection paths between potential new nodes and core nodes. The shortest path length from the potential new node to the core node is calculated based on the set of connection paths, and the connection density score of the potential new node is determined based on the shortest path length and the number of paths. If the connection density score is higher than a preset threshold, keyword vectors are extracted from the attribute description text of the potential new node, and the semantic relevance score between the keyword vectors and the core topic vectors in the existing graph is calculated using the TF-IDF algorithm combined with semantic drift analysis to obtain a corrected semantic score; if the corrected semantic score is higher than the preset semantic relevance threshold, the potential new node passes the screening. The potential new nodes selected through screening and the relationships between these nodes and key nodes in the process are integrated into the knowledge graph to form an enhanced knowledge graph centered on the correlation between complications and prognosis. Based on the enhanced knowledge graph, the comprehensive risk stratification results of the patients to be analyzed are output.

2. The comprehensive risk stratification method for cancer patients based on data analysis according to claim 1, characterized in that, The TF-IDF algorithm combined with semantic drift analysis is used to calculate the semantic relevance score between keyword vectors and existing core topic vectors in the geographic graph, thus obtaining the corrected semantic score. The steps include the following: Obtain the attribute description text of potential new nodes and perform word segmentation to obtain an initial keyword set; The TF-IDF algorithm is used to vectorize the initial keyword set to obtain keyword vectors. Obtain the attribute description text of the core topic vector of the existing graph and the related nodes within one hop range of the core topic vector of the existing graph; then, perform word segmentation on the attribute text of each node and vectorize it using the TF-IDF algorithm to obtain the semantic vector of each node; finally, perform a weighted average of the semantic vectors of all core nodes and related nodes, and obtain the initial semantic relevance score by calculating the cosine similarity between the keyword vector and the core topic vector of the existing graph. Obtain the data source identifier of potential new nodes, and query the preset semantic drift table based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier; The data collection timestamps for potential new nodes, the medical guideline update timestamps for the risk factor types to which potential new nodes belong, and the record time difference values ​​for the risk factor types to which potential new nodes belong in different data sources are obtained. The time-lapse factor is calculated by combining the collection timestamps, medical guideline update timestamps, and record time difference values. Multiply the baseline drift value by the time-sensitivity drift factor to obtain the semantic drift amount; The initial semantic relevance score is attenuated and corrected using the semantic drift, resulting in the corrected semantic score.

3. The comprehensive risk stratification method for cancer patients based on data analysis according to claim 2, characterized in that, The attribute description text includes description text of risk factors related to complications and description text of risk factors related to baseline status.

4. The comprehensive risk stratification method for cancer patients based on data analysis according to claim 3, characterized in that, The timeliness drift factor is calculated as follows: For the medical guideline dimension, obtain the medical guideline version index 'a' and update cycle index 'i'. Calculate the absolute difference between the data collection timestamp corresponding to the current potential new node and the weighted average of all key time nodes in the 'i' update cycle of the 'a'-th medical guideline version. Divide this difference by the time span of the update cycle to obtain the normalized time deviation for that version and cycle. Multiply this by the preset weight coefficient corresponding to that version and cycle. Sum the above products for all versions and all cycles to obtain the guideline dimension drift value. For the data source dimension, obtain the data source index 'b' and record batch index 'i'. Calculate the absolute difference between the collection timestamp and the weighted average of all record times in the 'i'-th record batch of the 'b'-th data source. Divide this by the time span of the record batch to obtain the normalized time deviation for that source and batch. Multiply this by the preset weight coefficient corresponding to that source and batch. Sum the above products for all sources and all batches to obtain the data source dimension drift value. Add the guideline dimension drift value to the data source dimension drift value to obtain the timeliness drift factor.

5. The comprehensive risk stratification method for cancer patients based on data analysis according to claim 4, characterized in that, Obtain the data source identifier of potential new nodes, and query the preset semantic drift table based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier. This includes the following steps: Extract the data source field from the metadata of potential new nodes as a data source identifier. The data source identifier includes the electronic medical record system identifier, the laboratory information system identifier, or the image archiving system identifier. Query the preset semantic drift scale to obtain the initial drift value corresponding to the data source identifier. The semantic drift scale pre-stores the degree of conceptual shift caused by inconsistencies in the expression of the same medical term among different medical data sources. Obtain the qualification level of data entry personnel for potential new nodes, and query the preset personnel confidence table based on the qualification level to obtain the personnel confidence coefficient; The baseline drift value is obtained by weighting and summing the initial drift value with the personnel confidence coefficient.

6. The comprehensive risk stratification method for cancer patients based on data analysis according to claim 5, characterized in that, Obtain the data source identifier of potential new nodes, and query the preset semantic drift table based on the data source identifier to obtain the baseline drift value corresponding to the data source identifier. This includes the following steps: Obtain all data source identifiers corresponding to potential new nodes, form a source identifier set, and query the preset semantic drift table for each source identifier to obtain the initial drift value corresponding to each source; If the source identifier set contains only one source identifier, then the initial drift value of that source is directly used as the baseline drift value; if it contains multiple source identifiers, then each source identifier is sorted from high to low according to its corresponding initial drift value to form a first sorted list, and the first source with the highest initial drift value and the second source with the lowest initial drift value are extracted. The record fragments from the first source and the second source in the attribute description text of the potential new node are obtained respectively, and the record features of the record fragments are extracted by natural language processing tools. The record features from the first and second sources are used to determine conflicts, and then a correction and adjustment process is performed to obtain the baseline drift value.

7. A comprehensive risk stratification method for cancer patients based on data analysis according to claim 5, characterized in that, The record features from the first and second sources are used to determine conflicts, and then a correction and adjustment process is performed to obtain the baseline drift value. This includes the following steps: When there is a conflict between the record features of the first source and the second source, the first historical record set of the first source is obtained and the consistency score of the first historical record set within the first source is calculated; and the second historical record set of the second source is obtained and the consistency score of the second source is calculated. The consistency scores within the first and second sources are analyzed and processed using preset high consistency thresholds and preset low consistency thresholds to obtain the first candidate drift value and the second candidate drift value, respectively. The first candidate drift value and the second candidate drift value are cross-compared, the deviation between the two is calculated, and the temporary reference drift value is obtained by analyzing the deviation between the two and the preset acceptable deviation range. The temporary baseline drift value is compared with each candidate drift value in the source identifier set to calculate the deviation; based on the deviation and a preset deviation tolerance threshold, the baseline drift value is obtained through iterative trend analysis.

8. A comprehensive risk stratification method for cancer patients based on data analysis according to claim 7, characterized in that, Conflict detection refers to determining whether the recording time granularity is consistent, whether the numerical units are uniform, whether the recording personnel roles are consistent, or whether the number of overlapping keyword sets is lower than a preset overlap threshold.

9. A comprehensive risk stratification system for cancer patients based on data analysis, characterized in that, It includes a path extraction module, a connectivity density analysis module, a filtering module, and an enhanced hierarchical module; The path extraction module is used to extract potential new nodes and their associated relationship data from a preset database, and to traverse the paths between core nodes and potential new nodes in the existing knowledge graph corresponding to the preset database based on the associated relationship data, thereby obtaining a set of connection paths between potential new nodes and core nodes. The connection density analysis module is used to calculate the shortest path length from the potential new node to the core node based on the set of connection paths, and to determine the connection density score of the potential new node based on the shortest path length and the number of paths. The filtering module is used to extract keyword vectors from the attribute description text of potential new nodes if the connection density score is higher than a preset threshold, and to calculate the semantic relevance score between the keyword vectors and the core topic vectors in the existing graph using the TF-IDF algorithm combined with semantic drift analysis to obtain a corrected semantic score; if the corrected semantic score is higher than the preset semantic relevance threshold, the potential new node passes the filtering. The enhanced stratification module is used to integrate the potential new nodes that have passed the screening and the relationships between the potential new nodes and the key nodes in the process into the knowledge graph, forming an enhanced knowledge graph with the correlation between complications and prognosis as the core. Based on the enhanced knowledge graph, the comprehensive risk stratification results of the patients to be analyzed are output.

10. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to perform a data analysis-based comprehensive risk stratification method for tumor patients as described in any one of claims 1-8.

Images