Clinical data analysis method and system for hematological malignancy patients based on knowledge graph

Abstract

The present invention discloses a clinical data analysis method and system for hematological tumor patients based on knowledge graphs, which relates to the field of medical information processing technology. The present invention utilizes a natural language processing model to parse unstructured text and construct a standardized entity set. Laboratory indicators are aggregated through dynamic windows to construct a timeline data set annotated with treatment stages. Standardized entities are integrated based on a hematological tumor ontology model, and a dynamic global knowledge graph is constructed in combination with clinical causal relationships and temporal paths to achieve chain reaction warnings. Based on real-time patient subgraphs, successful cases with the same molecular characteristics are retrieved in the knowledge graph to generate a set of candidate treatment plans, and a graph neural network trained with an optimized algorithm is used to predict the risk of complications of each plan, and finally the optimal treatment plan is evaluated and selected. The present invention breaks through the barriers of clinical data, realizes dynamic reasoning, advanced risk warning and personalized plan optimization throughout the treatment cycle, and significantly improves the accuracy, safety and efficiency of medical resource utilization in diagnosis and treatment.

Description

Technical Field

[0001] The present invention belongs to the field of medical information processing technology, and more specifically, relates to a method and system for analyzing clinical data of hematological malignancy patients based on a knowledge graph. Background Art

[0002] The current diagnosis and treatment of hematologic malignancies faces a high degree of complexity. Disease subtypes are diverse, treatment plans must be dynamically adjusted based on genetic mutations, and the risk of complications such as bone marrow suppression and infection changes in real time during treatment. Current clinical decision-making relies on physician experience, while multi-source, heterogeneous medical data creates information silos, making unstructured text parsing difficult and key entities challenging to standardize and integrate. Traditional clinical decision support systems, often based on static rule bases, are unable to model treatment timing dynamics and individualized risk warnings, resulting in delayed treatment plan optimization and blind spots in patient safety monitoring.

[0003] Existing technologies have problems such as insufficient data integration, rigid knowledge representation, and delayed decision support. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the problems in the related technology, the present invention provides a clinical data analysis method for blood cancer patients based on knowledge graph to overcome the above-mentioned technical problems existing in the existing related technology.

[0006] (2) Technical solution

[0007] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] S1. Collect and process clinical data of hematological malignancy patients to obtain a standardized entity set and a timeline dataset with annotated treatment stages.

[0009] S2. Based on the blood cancer ontology model file and the standardized entity set, and according to the clinical causal relationship and temporal path, a dynamic global blood cancer knowledge graph is obtained;

[0010] S3. Based on real-time patient data, determine whether to build a warning entity node in the dynamic global hematological cancer knowledge graph; based on the treatment stage dataset marked by the timeline, build a real-time patient stage sub-graph;

[0011] S4. Based on the real-time patient stage subgraph, search the dynamic global blood cancer knowledge graph to obtain the optimal candidate treatment plan set;

[0012] S5. Collect subgraphs of historical patients in the dynamic global hematological tumor knowledge graph to obtain subgraphs of historical patients with complication annotations; train a graph neural network using the subgraphs of historical patients with complication annotations in combination with an optimization algorithm to obtain a final graph neural network;

[0013] S6. The optimal candidate treatment plan set is combined with the real-time patient sub-graph and input into the final graph neural network to obtain the complication warning result set;

[0014] Select the final treatment plan from the optimal candidate treatment plan set based on the complication warning result set;

[0015] The present invention connects to the hospital's multi-source system, parses text using NLP technology to construct a standardized entity set, and aggregates laboratory indicators with the diagnosis date as the origin to form a timeline data set; integrates entities based on the blood tumor ontology model, and constructs a dynamic knowledge graph combining clinical causality and temporal paths to achieve treatment event anchoring and dose-effect warning; generates a set of candidate solutions based on real-time patient sub-graph retrieval of historical successful cases with the same molecular characteristics; uses historical sub-graphs with complication annotations to train graph neural networks to predict the complication risks of each solution; and finally selects the optimal treatment plan based on risk-benefit assessment. This method breaks down data silos, achieves advanced warning and personalized safe treatment, improves efficacy and reduces the risk of complications.

[0016] Preferably, the S1 comprises the following steps:

[0017] S11. Collect disease diagnosis records in electronic medical records, test reports in laboratory information systems, and text descriptions in imaging archiving systems to obtain a multi-source heterogeneous database of the hospital;

[0018] Connecting to the hospital's multi-source heterogeneous database through medical industry standard protocols automatically extracts test indicators and medication records from structured data. Using natural language processing models, it parses unstructured text in the hospital's multi-source heterogeneous database, identifies key medical entities, and generates a standardized entity set. This standardized entity set includes disease classifications, gene mutations, and specific medication regimens.

[0019] S12. Collect the timestamps of drug use and the time series of laboratory test indicators from the hospital's multi-source heterogeneous database to obtain time series data. Using the patient's diagnosis date as the time origin, uniformly calibrate all clinical event time points in the time series data. Aggregate laboratory test indicators using a dynamic window to obtain a timeline dataset annotated with treatment stages.

[0020] This invention automatically extracts structured medication and test data by adopting medical standard protocols, and uses NLP models to accurately parse unstructured text to identify standardized entity sets such as disease typing, gene mutations, and medication regimens; using the diagnosis date as the time origin, it aggregates laboratory indicators through dynamic windows to construct a timeline data set annotated with treatment stages; it breaks the hospital data silos, realizes the deep integration and standardization of key medical entities, and provides a high-quality time series data foundation for subsequent precise analysis.

[0021] Preferably, said S2 comprises the following steps:

[0022] S21. Define semantic rules and classification systems for the field of hematological tumors, ensure that the knowledge graph conforms to medical logic, and obtain a hematological tumor ontology model file; the hematological tumor ontology model includes an entity layer, a relationship layer, and a rule layer; the entity layer includes disease subtypes, gene mutations, and treatment options; the relationship layer includes induction therapy-includes-treatment options and FLT3-induces-targeted drug sensitivity; the rule layer includes recurrence warning rules;

[0023] Extract the classes, attributes, and relationship constraints from the blood tumor ontology model file to obtain the semantic definition of the blood tumor ontology;

[0024] S22. Based on the standardized entity set and the semantic definition of the blood tumor ontology, the standardized entities are associated with the ontology classes, and the relationships between the entities are defined according to the ontology attributes to obtain a static blood tumor knowledge graph framework;

[0025] S23. Based on a timeline dataset with annotated treatment stages and a static hematologic malignancy knowledge graph framework, a rule engine is used to detect clinical causal relationships and generate temporal pathways based on treatment responses to obtain a dynamic global hematologic malignancy knowledge graph.

[0026] Based on the blood tumor ontology model, the present invention maps standardized entities into knowledge nodes to construct a static graph framework; combined with the timeline dataset, the rule engine detects clinical causal relationships and generates temporal paths to form a dynamic global knowledge graph; it realizes treatment logic deduction and dose-effect warning, providing medical logic guarantee for accurate decision-making.

[0027] Preferably, the step S3 includes the following steps:

[0028] S31. Collect real-time laboratory flow data and current medication records to obtain real-time patient data;

[0029] S32. Based on real-time patient data and a set of time-sensitive rules, construct warning entity nodes and associated relationships in a dynamic global hematological tumor knowledge graph; issue a warning when a warning entity node is identified, and respond based on the relationship;

[0030] S33. Based on real-time patient data and a timeline-marked treatment stage dataset, discrete stages are divided according to clinical pathways and stage transitions are determined based on medical logic to obtain a stage sub-map.

[0031] The present invention dynamically creates early warning nodes in the knowledge graph by matching timeliness rules with real-time streaming data; at the same time, it automatically divides the treatment stages based on medical logic and generates stage sub-graphs containing current medications and indicator trends; thus achieving proactive intervention of complications and precise management of the entire treatment process.

[0032] Preferably, the S32 includes the following steps:

[0033] S321. Accurately align drug exposure time with changes in laboratory indicators; calculate the dose-effect relationship between drug dose and blood cell decline rate; establish a chain reaction model based on the dose-effect relationship;

[0034] S322. Constructing a timeliness rule set based on the chain reaction model;

[0035] S323. Match the real-time patient data with the timeliness rules in the timeliness rule set. If any data in the real-time patient data exceeds the threshold in the timeliness rule, automatically create a clinical event node in the dynamic hematologic malignancy knowledge graph and associate it with the current treatment cycle, thereby obtaining a dynamic hematologic malignancy knowledge graph containing warning entities and associated relationships. Otherwise, no warning entity is created.

[0036] This invention establishes a dose-effect model linking drug dose to blood cell decline rate, constructing a chained warning rule set, such as "drug → bone marrow suppression → infection," to achieve millisecond-level matching of real-time streaming data with rules. When thresholds are exceeded, a warning node is automatically created in the knowledge graph and associated with the treatment cycle, forming a dynamic risk map. This improves the response speed of clinical validation and reduces the incidence of serious infections.

[0037] Preferably, the S4 comprises the following steps:

[0038] S41. Based on the real-time patient stage subgraph, extract the patient's current gene mutation characteristics to obtain the real-time patient molecular characteristics; search the knowledge graph for historical patient groups with the same molecular characteristics, screen successful cases that have achieved complete remission treatment goals, and obtain a successful case set;

[0039] S42. Analyze the treatment regimens used by successful cases in the successful case set, and score the efficacy of the treatment regimens used by successful cases based on the duration of complete remission, depth of minimal residual disease clearance, and incidence of treatment-related toxicity to obtain the optimal candidate treatment regimen set;

[0040] The present invention uses real-time patient genetic characteristics to accurately match historical successful cases in the knowledge graph, and comprehensively integrates three dimensions: duration of complete remission, depth of clearance of minimal residual lesions, and incidence of treatment toxicity to quantify the efficacy score and generate an optimal set of candidate regimens; it improves the patient's first induced remission rate and reduces the number of regimen-related serious toxic events.

[0041] Preferably, the S5 comprises the following steps:

[0042] S51. Collect sub-graphs of historical patients from the dynamic global hematological tumor knowledge graph to obtain sub-graphs of historical patients with complication annotations;

[0043] S52. Construct a graph neural network; use the historical patient subgraph with complication annotations to train the graph neural network, combine the optimization algorithm to find the learning rate of the graph neural network during the training process, and obtain the optimal solution; use the optimal solution as the learning rate of the graph neural network to obtain the final graph neural network;

[0044] The present invention uses historical patient sub-graphs with complication annotations to train graph neural networks and automatically determines the optimal learning rate through an optimization algorithm; it improves the accuracy of complication prediction, shortens the prediction time, and provides core support for the quantification of treatment plan risks.

[0045] Preferably, in the training process in S52, finding the learning rate of the graph neural network in combination with an optimization algorithm to obtain the optimal solution includes the following steps:

[0046] S521, set the parameter search space, the maximum number of random searches, and the cross-validation fold C ;

[0047] S522: Search space according to the parameters ， Define the learning rate set;

[0048] Initialize the optimal learning rate to F , initialize the best performance to G , randomly select a learning rate from the candidate learning rate set H , the historical patient sub-graphs with complications are divided into C Fold, for each fold c , ; Use except c All historical patient subgraphs with complications outside the fold are used to train the graph neural network. c The historical patient sub-graph with complication annotations is used to evaluate the model and obtain the performance index set; the mean of the performance indexes of all folds is calculated to obtain the average performance P ;

[0049] If the average performance P > Best performanceG , then the updated best performance is P , update the optimal learning rate to H ; Otherwise, maintain the original best performance and optimal learning rate;

[0050] S523, repeat S522, and when the maximum number of random searches is reached, stop the iteration to obtain the final graph neural network;

[0051] The present invention uses a random search algorithm to iteratively optimize within a preset learning rate space; cross-validation is performed by dividing historical patient sub-graphs into C folds, randomly selecting a learning rate to train the graph neural network in each round, evaluating the performance with the validation fold and calculating the average performance index P; when P exceeds the current best performance G, the optimal learning rate is updated until the maximum number of searches is reached; adaptive optimization of the learning rate is achieved, which improves the efficiency of model training, enhances the accuracy of complication prediction, and avoids the blindness of manual parameter adjustment.

[0052] Preferably, the S6 comprises the following steps:

[0053] S61. Obtain a real-time patient subgraph using a dynamic global hematologic malignancy knowledge graph; combine candidate treatment plans in the optimal candidate treatment plan set with the real-time patient subgraphs, and sequentially input them into a final graph neural network to obtain a complication warning result set; the complication warning result set includes complication prediction results after each candidate treatment plan is applied to the real-time patient;

[0054] S62, selecting a final treatment plan from the optimal candidate treatment plan set based on the complication warning result set to treat the real-time patient;

[0055] The present invention dynamically predicts the risk of complications under each plan by inputting the optimal candidate treatment plan and the real-time patient sub-graph into a trained graph neural network, and generates a complication warning result set; based on the result set, the risk-benefit ratio is quantitatively evaluated, and the plan with the lowest complication risk and the best efficacy is selected as the final treatment plan; thus, safe optimization of treatment decisions is achieved, clinical risks are significantly reduced, and patient tolerance is improved.

[0056] A knowledge graph-based clinical data analysis system for hematological malignancy patients, designed to implement the aforementioned knowledge graph-based clinical data analysis method for hematological malignancy patients, includes a multi-source medical data integration and standardization module, a dynamic knowledge graph construction module, a real-time patient monitoring and early warning module, an intelligent treatment plan matching module, a graph neural network optimization training module, and a complication prediction and decision-making module.

[0057] The multi-source medical data integration and standardization module is used to connect to the hospital's electronic medical records, laboratory information systems, and imaging archiving systems. It automatically extracts structured data through the HL7 / FHIR protocol and uses natural language processing technology to parse unstructured text, identifying key medical entities such as disease typing, gene mutations, and medication regimens, and constructing a standardized entity set with timestamps. At the same time, it dynamically aggregates laboratory indicators within the window, using the patient's diagnosis date as the time origin, to generate a timeline dataset annotated with treatment stages.

[0058] The dynamic knowledge graph construction module maps standardized entities to ontology classes based on a predefined hematological tumor ontology model to establish a static knowledge graph framework. Clinical causal relationships and treatment response temporal paths are then injected into the rule engine, combined with timeline data to generate a dynamic global knowledge graph that supports logical deduction of disease evolution and execution of early warning rules.

[0059] The real-time patient monitoring and early warning module receives real-time streaming data and performs dynamic risk assessment using a set of time-sensitive rules. When a threshold crossing is detected, an early warning node is automatically created in the knowledge graph and associated with the treatment stage to implement an early warning. At the same time, the patient's current treatment stage is divided according to clinical pathway rules, and a stage-specific sub-graph is constructed.

[0060] The intelligent treatment plan matching module searches for historical successful cases with the same characteristics in the knowledge graph based on real-time patient molecular characteristics, extracts their treatment plans, and scores them based on comprehensive efficacy indicators to generate the optimal candidate treatment plan set, providing a data-driven decision-making basis for personalized treatment;

[0061] The graph neural network optimization training module collects historical patient subgraphs (including complication annotations), constructs a graph neural network model, and uses a random search optimization algorithm to iteratively search for the optimal learning rate within a preset parameter space. Model performance is evaluated through cross-validation, and the learning rate is dynamically updated based on average performance, ultimately training a graph neural network that can accurately model treatment-complication associations.

[0062] The complication prediction and decision-making module inputs candidate treatment plans and real-time patient sub-graphs into the trained graph neural network to predict the complications that may be caused by each plan and generate a complication warning result set; clinicians then combine the efficacy score and complication probability to select the final treatment plan from the candidate plans, thereby achieving risk-controlled individualized treatment decisions.

[0063] (3) Beneficial effects

[0064] The present invention has the following beneficial effects:

[0065] The present invention has built a closed-loop, precise, and intelligent clinical decision support system for blood tumors by deeply integrating multi-source heterogeneous data, constructing a dynamic knowledge graph with temporal logic and causal reasoning capabilities, realizing intelligent recommendations for personalized plans based on molecular characteristics and complication risk prediction, and automating diagnosis and treatment process management. It significantly improves the efficiency and standardization of diagnosis and treatment, realizes advanced and accurate risk warning, optimizes the selection of treatment plans to maximize efficacy while minimizing complication risks, and ultimately improves the quality of life and prognosis of patients, and effectively improves the efficiency of medical resource utilization.

[0066] This invention breaks the data silos within the hospital by connecting to heterogeneous systems within the hospital and using NLP models to accurately parse unstructured text; it automatically extracts structured data such as medication records and diagnostic codes, and accurately identifies and standardizes key medical entities from free text; this deep integration and standardization processing lays a solid and unified data foundation for the subsequent construction of high-quality knowledge graphs and precise analysis, significantly improving the availability and intrinsic value density of clinical data.

[0067] The present invention constructs a dynamic global knowledge graph based on a blood tumor ontology model; this graph not only statically organizes disease classification, gene mutations, treatment plans and their relationships, but also incorporates a powerful temporal dimension and logical reasoning capabilities; it accurately anchors treatment events with the patient's diagnosis date as the origin, and aggregates laboratory indicators through a dynamic window to truly depict the dynamic changes in the treatment process; it uses a rule engine to detect clinical causal relationships in real time and generate treatment response paths, giving the graph clinical deduction capabilities; based on the dose-effect relationship model and chain reaction model, the system can automatically match timeliness rules according to real-time streaming data, trigger graded warnings in advance, improve the timeliness and effectiveness of intervention, and effectively prevent serious complications.

[0068] The present invention achieves an optimized balance between high personalization and safety in recommending treatment options. Based on the real-time molecular characteristics of patients, the system retrieves historical successful cases with the same molecular characteristics and achieving complete remission in the knowledge graph to ensure that the recommended options have a solid evidence-based basis. The candidate options are then scored for multi-dimensional efficacy, taking into account the duration of complete remission, the depth of clearance of minimal residual lesions, and the incidence of treatment-related toxicity, taking into account both long-term survival benefits and treatment safety. The graph neural network model trained with the optimization algorithm is combined with a real-time sub-graph containing information such as the patient's current stage, medication, and indicator trends to dynamically predict the risk of specific complications such as severe infection or organ toxicity in patients under each candidate option. The final treatment option is based on a precise quantitative assessment of the individualized risk-benefit ratio, significantly improving the safety of treatment and patient tolerance.

[0069] This invention significantly improves the intelligence and standardization of the diagnosis and treatment process and optimizes the allocation of medical resources; the system automatically completes the complex closed-loop process from multi-source data integration, time series analysis, knowledge reasoning, risk assessment to plan recommendation; based on preset strict medical logic, it ensures that patients follow the best clinical path and reduces human delays; by creating an independent sub-graph for each treatment stage that includes key events, medication plans, indicator change trends and complications, it achieves refined management and review of the treatment process; the advanced warning mechanism enables medical staff to deploy resources in advance to respond to high-risk events, optimizes human resource allocation, focuses limited energy on the most critical patients, and improves the efficiency and quality of medical services as a whole.

[0070] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, they can also obtain drawings based on these drawings without paying any creative work.

[0072] Figure 1 This is a flow chart of the clinical data analysis method for hematological malignancy patients based on the knowledge graph of the present invention;

[0073] Figure 2 Schematic diagram of the blood tumor ontology model in the knowledge graph-based clinical data analysis method for blood tumor patients of the present invention;

[0074] Figure 3 This is a module diagram of the clinical data analysis system for blood cancer patients based on the knowledge graph of the present invention. DETAILED DESCRIPTION

[0075] The following will clearly and completely describe the technical solutions in the embodiments of the invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0076] In the description of the present invention, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inside" and the like indicating orientation or positional relationship are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the invention.

[0077] Example 1:

[0078] See also Figure 1 、 Figure 2 The present invention discloses a method for analyzing clinical data of hematological malignancy patients based on a knowledge graph, comprising the following steps:

[0079] S1. Collect and process clinical data of hematological malignancy patients to obtain a standardized entity set and a timeline dataset with annotated treatment stages.

[0080] Said S1 comprises the following steps:

[0081] S11. Collect disease diagnosis records in electronic medical records, test reports in laboratory information systems, and text descriptions in imaging archiving systems to obtain a multi-source heterogeneous database of the hospital;

[0082] By connecting to the hospital's multi-source heterogeneous database through standard medical industry protocols, the system automatically extracts test indicators and medication records from structured data. It then uses natural language processing models to parse unstructured text within the hospital's multi-source heterogeneous database, identifying key medical entities and generating a standardized entity set. This standardized entity set includes disease classifications (e.g., acute myeloid leukemia), gene mutations (e.g., FLT3-ITD positive), and specific medication regimens (e.g., cytarabine dosage).

[0083] The S11 includes the following steps:

[0084] Deploy a medical data interface adapter to connect to the hospital's EMR system via the HL7 / FHIR standard protocol, automatically extracting structured data such as medication records and diagnosis codes daily;

[0085] Connect to the LIS (Laboratory Information System) to obtain the time series of test indicators, including key parameters such as white blood cell count, hemoglobin, platelets, etc. in blood routine tests;

[0086] Integrate with PACS (Picture Archiving System) API to extract raw image data and structured diagnostic conclusions from bone marrow biopsy reports;

[0087] The test report text is segmented into character sequences and fed into the BERT model to obtain context-dependent vector representations. Long-range semantic dependencies are captured through a bidirectional LSTM. The CRF layer decodes and outputs entity labels (e.g., "FLT3:c.2503T>G" is identified as a "gene mutation" entity). Key entity extraction: Core attributes such as the mutant gene name, variant type (SNV / INDEL), and variant frequency are parsed from the variant description.

[0088] S12. Collect drug usage timestamps and laboratory test index time series from the hospital's multi-source heterogeneous database to obtain time series data. Using the patient's diagnosis date as the time origin, uniformly calibrate all clinical event time points in the time series data, and aggregate laboratory test indicators according to a 72-hour dynamic window (e.g., taking the lowest platelet value within 72 hours after chemotherapy) to obtain a timeline dataset annotated with treatment stages.

[0089] Constructing a timeline of treatment phases involves the following steps:

[0090] The main timeline was established with the patient's first diagnosis date as the origin (t0). Key events were anchored: the start date of induction chemotherapy was marked as t0 + 3 days (median time from diagnosis to treatment initiation); the completion date of genetic testing was marked as t0 + 10 days (according to clinical pathway specifications); the start date of consolidation therapy was set within 7 days after the first efficacy assessment (when bone marrow puncture showed CR);

[0091] All event timestamps are accurate to the minute, ensuring the chronological order of medication records and laboratory tests is traceable. A 72-hour rolling time window is defined; the starting point of each window is a key treatment event (such as the time of chemotherapy administration). Laboratory indicator aggregation rules include: taking the minimum platelet count of multiple tests within the window (reflecting the trough of bone marrow suppression); the absolute neutrophil count is weighted by exponential decay (with more recent data having a higher weight); and special scenario processing: hemoglobin values ​​within 24 hours after a blood transfusion are marked as post-intervention data and excluded from aggregation.

[0092] S2. Based on the blood cancer ontology model file and the standardized entity set, and according to the clinical causal relationship and temporal path, a dynamic global blood cancer knowledge graph is obtained;

[0093] The S2 comprises the following steps:

[0094] S21. Define semantic rules and classification systems for hematological malignancies (e.g., disease classification hierarchy, drug attribute definitions, treatment regimens, etc.), ensure that the knowledge graph conforms to medical logic, and obtain a hematological malignancy ontology model file; the hematological malignancy ontology model includes an entity layer, a relationship layer, and a rule layer; the entity layer includes disease subtypes, gene mutations, and treatment regimens; the relationship layer includes induction therapy-includes-treatment regimens and FLT3-induces-targeted drug sensitivity; and the rule layer includes recurrence warning rules;

[0095] Extract the classes (such as Chemotherapy and DiseaseSubtype), attributes (such as causes and treats), and relationship constraints from the blood tumor ontology model file to obtain the semantic definition of the blood tumor ontology;

[0096] S22. Based on the standardized entity set and the semantic definition of the hematological cancer ontology, associate the standardized entities with the ontology classes (for example, the clinical term "DA regimen" is mapped to the ChemotherapyClass in the ontology), and define the relationships between entities based on the ontology attributes (for example, "daunorubicin" belongsTo "DA regimen") to obtain a static hematological cancer knowledge graph framework;

[0097] S23, based on the timeline dataset with annotated treatment stages and the static hematological tumor knowledge graph framework, the rule engine is used to detect clinical causal relationships (e.g., when platelets are <50×10 9 When the patient is receiving cytarabine (CART), a "drug-induced thrombocytopenia" relationship is established, and a temporal path (e.g., "induction therapy → complete remission → consolidation therapy") is generated based on the treatment response, resulting in a dynamic global hematologic malignancy knowledge graph.

[0098] S3. Based on real-time patient data, determine whether to build a warning entity node in the dynamic global hematological cancer knowledge graph; based on the treatment stage dataset marked by the timeline, build a real-time patient stage sub-graph;

[0099] The S3 includes the following steps:

[0100] S31. Collect real-time laboratory flow data (such as neutrophil counts updated every minute) and current medication records to obtain real-time patient data;

[0101] S32. Based on real-time patient data and a set of time-sensitive rules, construct warning entity nodes and association relationships in a dynamic global blood cancer knowledge graph;

[0102] The S32 includes the following steps:

[0103] S321. Accurately align drug exposure time with changes in laboratory indicators; calculate the dose-effect relationship between drug dose and rate of blood cell decline; establish a chain reaction model based on the dose-effect relationship (e.g., "drug → bone marrow suppression → secondary infection");

[0104] S322. Constructing a timeliness rule set based on the chain reaction model;

[0105] S323. Match the real-time patient data with the timeliness rules in the timeliness rule set. If any data in the real-time patient data exceeds the threshold in the timeliness rule (for example, when the neutrophil count decreases by more than 60% within 48 hours and a specific chemotherapy drug is used), automatically create a clinical event node (such as "bone marrow suppression") in the dynamic hematologic malignancy knowledge graph and associate it with the current treatment cycle, thereby obtaining a dynamic hematologic malignancy knowledge graph containing warning entities and associated relationships (such as "cyclophosphamide causes bone marrow suppression"). Otherwise, no warning entity is created.

[0106] Take bone marrow suppression as an example: the early warning mechanism is as follows,

[0107] When the absolute neutrophil count drops by more than 60% within 48 hours, the warning mechanism is automatically triggered; Level 1 warning: when the neutrophil count drops to 1.5×10 9 / L, a clinical attention reminder is generated; Level 2 warning: when the neutrophil count drops by >60% within 48 hours and is accompanied by fever, an emergency treatment process is triggered; Level 3 warning: when the neutrophil count is <0.5×10 9 / L automatically notifies the attending physician;

[0108] S33. Based on real-time patient data and a timeline-labeled treatment phase dataset, discrete phases are divided according to the clinical pathway (e.g., initial diagnosis → induction therapy → response assessment) and phase transitions are determined based on medical logic (e.g., consolidation therapy within 7 days after complete response) to generate a phase sub-map.

[0109] Take bone marrow suppression as an example:

[0110] Starting from the first pathological diagnosis, including core events such as disease typing and molecular testing; starting from the first chemotherapy administration and continuing until the first efficacy evaluation, within 7 days after the bone marrow puncture examination, the complete remission (CR) and partial remission (PR) status are confirmed. After CR, intensive treatment is started. Intensive treatment usually consists of 2-4 cycles. After the completion of intensive treatment, long-term low-intensity treatment is maintained to prevent relapse;

[0111] When bone marrow blasts are <5% and peripheral blood cells recover, the system automatically marks the transition from "induction therapy to remission assessment"

[0112] When CR status is confirmed and organ function reaches the target (such as left ventricular ejection fraction > 50%), consolidation therapy is initiated; the interval between each stage should not exceed 14 days (except for special treatment plans);

[0113] Create an independent sub-atlas for each treatment stage, which includes the stage-specific medication regimen and dosage, the trend of changes in key laboratory indicators, and the occurrence of complications within the stage;

[0114] S4. Based on the real-time patient stage subgraph, search the dynamic global blood cancer knowledge graph to obtain the optimal candidate treatment plan set;

[0115] The S4 comprises the following steps:

[0116] S41. Based on the real-time patient stage subgraph, extract the patient's current gene mutation characteristics to obtain the real-time patient molecular characteristics; search the knowledge graph for historical patient groups with the same molecular characteristics, screen successful cases that achieved complete remission (CR) treatment goals, and obtain a successful case set;

[0117] S42. Analyze the treatment regimens used by successful cases in the successful case set, and score the efficacy of the treatment regimens used by successful cases based on the duration of complete remission, depth of minimal residual disease clearance, and incidence of treatment-related toxicity to obtain the optimal candidate treatment regimen set;

[0118] S5. Collect subgraphs of historical patients in the dynamic global hematological tumor knowledge graph to obtain subgraphs of historical patients with complication annotations; train a graph neural network using the subgraphs of historical patients with complication annotations in combination with an optimization algorithm to obtain a final graph neural network;

[0119] The S5 comprises the following steps:

[0120] S51. Collect sub-graphs of historical patients from the dynamic global hematological tumor knowledge graph to obtain sub-graphs of historical patients with complication annotations;

[0121] S52. Construct a graph neural network; use the historical patient subgraph with complication annotations to train the graph neural network, combine the optimization algorithm to find the learning rate of the graph neural network during the training process, and obtain the optimal solution; use the optimal solution as the learning rate of the graph neural network to obtain the final graph neural network;

[0122] In the training process in S52, the optimization algorithm is combined to find the learning rate of the graph neural network to obtain the optimal solution, which includes the following steps:

[0123] S521, set the parameter search space to A , the maximum number of random searches is M , set the cross validation fold to C ;

[0124] S522, searching the space according to the parameters A， Define the learning rate set d i represents the first i A random learning rate, e Indicates the total number of candidate learning rates;

[0125] Initialize the optimal learning rate to F , initialize the best performance to G , from the candidate learning rate set The learning rate is randomly selected in H , the historical patient sub-graphs with complications are divided into C Fold, for each fold c , Use except cAll historical patient subgraphs with complications outside the fold are used to train the graph neural network. c The historical patient subgraph with complication annotations is used to evaluate the model and obtain the performance indicator set in, K c Indicates the c performance indicators, C Indicates the total number of performance indicators; calculate the mean of all fold performance indicators to get the average performance P , the calculation formula is as follows,

[0126]

[0127] If the average performance R > Best performance G , then the updated best performance is R , update the optimal learning rate to H ; Otherwise, maintain the original best performance and optimal learning rate;

[0128] S523, repeat S522, when the maximum number of random searches is reached B , stop iteration and get the final graph neural network;

[0129] S6. The optimal candidate treatment plan set is combined with the real-time patient sub-graph and input into the final graph neural network to obtain the complication warning result set;

[0130] Select the final treatment plan from the optimal candidate treatment plan set based on the complication warning result set;

[0131] The S6 comprises the following steps:

[0132] S61. Obtain a real-time patient subgraph using a dynamic global hematologic malignancy knowledge graph; combine candidate treatment plans in the optimal candidate treatment plan set with the real-time patient subgraphs, and sequentially input them into a final graph neural network to obtain a complication warning result set; the complication warning result set includes complication prediction results after each candidate treatment plan is applied to the real-time patient;

[0133] S62. Select a final treatment plan from the optimal candidate treatment plan set based on the complication warning result set to treat the real-time patient.

[0134] Example 2:

[0135] See also Figure 3, a clinical data analysis system for hematological malignancy patients based on knowledge graph, which is used to implement the above-mentioned clinical data analysis method for hematological malignancy patients based on knowledge graph, including a multi-source medical data integration and standardization module, a dynamic knowledge graph construction module, a real-time patient monitoring and early warning module, a treatment plan intelligent matching module, a graph neural network optimization training module, and a complication prediction and decision-making module;

[0136] The multi-source medical data integration and standardization module is used to connect to the hospital's electronic medical records, laboratory information systems, and imaging archiving systems. It automatically extracts structured data through the HL7 / FHIR protocol and uses natural language processing technology to parse unstructured text, identifying key medical entities such as disease typing, gene mutations, and medication regimens, and constructing a standardized entity set with timestamps. At the same time, it dynamically aggregates laboratory indicators within the window, using the patient's diagnosis date as the time origin, to generate a timeline dataset annotated with treatment stages.

[0137] The dynamic knowledge graph construction module maps standardized entities to ontology classes based on a predefined hematological tumor ontology model to establish a static knowledge graph framework. Clinical causal relationships and treatment response temporal paths are then injected into the rule engine, combined with timeline data to generate a dynamic global knowledge graph that supports logical deduction of disease evolution and execution of early warning rules.

[0138] The real-time patient monitoring and early warning module receives real-time streaming data and performs dynamic risk assessment using a set of time-sensitive rules. When a threshold crossing is detected, an early warning node is automatically created in the knowledge graph and associated with the treatment stage to implement an early warning. At the same time, the patient's current treatment stage is divided according to clinical pathway rules, and a stage-specific sub-graph is constructed.

[0139] The intelligent treatment plan matching module searches for historical successful cases with the same characteristics in the knowledge graph based on real-time patient molecular characteristics, extracts their treatment plans, and scores them based on comprehensive efficacy indicators to generate the optimal candidate treatment plan set, providing a data-driven decision-making basis for personalized treatment;

[0140] The graph neural network optimization training module collects historical patient subgraphs (including complication annotations), constructs a graph neural network model, and uses a random search optimization algorithm to iteratively search for the optimal learning rate within a preset parameter space. Model performance is evaluated through cross-validation, and the learning rate is dynamically updated based on average performance, ultimately training a graph neural network that can accurately model treatment-complication associations.

[0141] The complication prediction and decision-making module inputs candidate treatment plans and real-time patient sub-graphs into the trained graph neural network to predict the complications that may be caused by each plan and generate a complication warning result set; clinicians then combine the efficacy score and complication probability to select the final treatment plan from the candidate plans, thereby achieving risk-controlled individualized treatment decisions.

[0142] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0143] The preferred embodiments of the invention disclosed above are intended only to help illustrate the invention. These preferred embodiments do not exhaust all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A clinical data analysis method for hematological malignancy patients based on knowledge graph, characterized in that: The following steps are involved: S1. Collect and process clinical data of hematological malignancy patients to obtain a standardized entity set and a timeline dataset with annotated treatment stages. S2. Based on the blood cancer ontology model file and the standardized entity set, and according to the clinical causal relationship and temporal path, a dynamic global blood cancer knowledge graph is obtained; S3. Based on real-time patient data, determine whether to build a warning entity node in the dynamic global hematological cancer knowledge graph; based on the treatment stage dataset marked by the timeline, build a real-time patient stage sub-graph; S4. Based on the real-time patient stage subgraph, search the dynamic global blood cancer knowledge graph to obtain the optimal candidate treatment plan set; The S4 comprises the following steps: S41. Based on the real-time patient stage subgraph, extract the patient's current gene mutation characteristics to obtain the real-time patient molecular characteristics; search the knowledge graph for historical patient groups with the same molecular characteristics, screen successful cases that have achieved complete remission treatment goals, and obtain a successful case set; S42. Analyze the treatment regimens used by successful cases in the successful case set, and score the efficacy of the treatment regimens used by successful cases based on the duration of complete remission, depth of minimal residual disease clearance, and incidence of treatment-related toxicity to obtain the optimal candidate treatment regimen set; S5. Collect subgraphs of historical patients in the dynamic global hematological tumor knowledge graph to obtain subgraphs of historical patients with complication annotations; train a graph neural network using the subgraphs of historical patients with complication annotations in combination with an optimization algorithm to obtain a final graph neural network; S6. The optimal candidate treatment plan set is combined with the real-time patient sub-graph and input into the final graph neural network to obtain the complication warning result set; The final treatment plan is selected from the optimal candidate treatment plan set based on the complication warning result set.

2. The method for analyzing clinical data of hematological malignancies patients based on knowledge graph according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect disease diagnosis records in electronic medical records, test reports in laboratory information systems, and text descriptions in imaging archiving systems to obtain a multi-source heterogeneous database of the hospital; Connecting to the hospital's multi-source heterogeneous database through medical industry standard protocols automatically extracts test indicators and medication records from structured data. Using natural language processing models, it parses unstructured text in the hospital's multi-source heterogeneous database, identifies key medical entities, and generates a standardized entity set. This standardized entity set includes disease classifications, gene mutations, and specific medication regimens. S12. Collect the timestamps of drug use and the time series of laboratory test indicators from the hospital's multi-source heterogeneous database to obtain time series data; use the patient's diagnosis date as the time origin, uniformly calibrate all clinical event time points in the time series data, and aggregate laboratory test indicators based on dynamic windows to obtain a timeline dataset marked with treatment stages.

3. The method for analyzing clinical data of hematological malignancies patients based on knowledge graph according to claim 1, characterized in that: The S2 comprises the following steps: S21. Define semantic rules and classification systems for the field of hematological tumors, ensure that the knowledge graph conforms to medical logic, and obtain a hematological tumor ontology model file; the hematological tumor ontology model includes an entity layer, a relationship layer, and a rule layer; the entity layer includes disease subtypes, gene mutations, and treatment options; the relationship layer includes induction therapy-includes-treatment options and FLT3-induces-targeted drug sensitivity; the rule layer includes recurrence warning rules; Extract the classes, attributes, and relationship constraints from the blood tumor ontology model file to obtain the semantic definition of the blood tumor ontology; S22. Based on the standardized entity set and the semantic definition of the blood tumor ontology, the standardized entities are associated with the ontology classes, and the relationships between the entities are defined according to the ontology attributes to obtain a static blood tumor knowledge graph framework; S23. Based on the timeline dataset with annotated treatment stages and the static hematological cancer knowledge graph framework, the clinical causal relationship is detected through the rule engine, and a temporal path is generated according to the treatment response to obtain a dynamic global hematological cancer knowledge graph.

4. The method for analyzing clinical data of hematological malignancies patients based on knowledge graph according to claim 1, characterized in that: The S3 includes the following steps: S31. Collect real-time laboratory flow data and current medication records to obtain real-time patient data; S32. Based on real-time patient data and a set of time-sensitive rules, construct warning entity nodes and associated relationships in a dynamic global hematological tumor knowledge graph; issue a warning when a warning entity node is identified, and respond based on the relationship; The S32 includes the following steps: S321. Accurately align drug exposure time with changes in laboratory indicators; calculate the dose-effect relationship between drug dose and blood cell decline rate; establish a chain reaction model based on the dose-effect relationship; S322. Constructing a timeliness rule set based on the chain reaction model; S323. Match the real-time patient data with the timeliness rules in the timeliness rule set. If any data in the real-time patient data exceeds the threshold in the timeliness rule, automatically create a clinical event node in the dynamic hematologic malignancy knowledge graph and associate it with the current treatment cycle, thereby obtaining a dynamic hematologic malignancy knowledge graph containing warning entities and associated relationships. Otherwise, no warning entity is created. S33. Based on real-time patient data and the treatment stage dataset marked on the timeline, the discrete stages are divided according to the clinical pathway and the stage transitions are judged based on medical logic to obtain a stage sub-map.

5. The method for analyzing clinical data of hematological malignancy patients based on knowledge graph according to claim 1, characterized in that: The S5 comprises the following steps: S51. Collect sub-graphs of historical patients from the dynamic global hematological tumor knowledge graph to obtain sub-graphs of historical patients with complication annotations; S52. Construct a graph neural network; use the historical patient sub-graph with complication annotations to train the graph neural network. During the training process, combine the optimization algorithm to find the learning rate of the graph neural network and obtain the optimal solution; use the optimal solution as the learning rate of the graph neural network to obtain the final graph neural network.

6. The method for analyzing clinical data of hematological malignant tumor patients based on knowledge graph according to claim 5, characterized in that: In the training process in S52, the optimization algorithm is combined to find the learning rate of the graph neural network to obtain the optimal solution, which includes the following steps: S521, set the parameter search space, the maximum number of random searches, and the cross-validation fold C ; S522, defining a learning rate set according to the parameter search space; Initialize the optimal learning rate to F , initialize the best performance to G , randomly select a learning rate from the candidate learning rate set H , the historical patient sub-graphs with complications are divided into C Fold, for each fold c , ; Use except c All historical patient subgraphs with complications outside the fold are used to train the graph neural network. c The historical patient sub-graph with complication annotations is used to evaluate the model and obtain the performance index set; the mean of the performance indexes of all folds is calculated to obtain the average performance P ; If the average performance P > Best performance G , then the updated best performance is P , update the optimal learning rate to H ; Otherwise, maintain the original best performance and optimal learning rate; S523. Repeat S522. When the maximum number of random searches is reached, stop the iteration and obtain the final graph neural network.

7. The method for analyzing clinical data of hematological malignancies patients based on knowledge graph according to claim 1, characterized in that: The S6 comprises the following steps: S61. Obtain a real-time patient subgraph using a dynamic global hematologic malignancy knowledge graph; combine candidate treatment plans in the optimal candidate treatment plan set with the real-time patient subgraphs, and sequentially input them into a final graph neural network to obtain a complication warning result set; the complication warning result set includes complication prediction results after each candidate treatment plan is applied to the real-time patient; S62. Select a final treatment plan from the optimal candidate treatment plan set based on the complication warning result set to treat the real-time patient.

8. A clinical data analysis system for hematological malignancies patients based on knowledge graph, characterized by: A method for analyzing clinical data of hematological malignancies patients based on a knowledge graph as described in any one of claims 1 to 7 is implemented, wherein the system includes a multi-source medical data integration and standardization module, a dynamic knowledge graph construction module, a real-time patient monitoring and early warning module, a treatment plan intelligent matching module, a graph neural network optimization training module, and a complication prediction and decision-making module.

Images