Hematologic tumor cord blood transplantation treatment prognosis index evaluation decision generation method and system based on prediction modeling, medium and electronic equipment

Abstract

The invention provides a hematologic tumor cord blood transplantation treatment prognosis index evaluation decision generation method and system based on prediction modeling, a medium and electronic equipment, and relates to the technical field of medical artificial intelligence. The method comprises the following steps: acquiring multi-source data of a key time point in a whole process of blood tumor umbilical cord blood transplantation treatment of a patient; integrating expert knowledge to form an expert knowledge network, taking the expert knowledge network as priori knowledge of multi-target prediction, and performing multi-dimensional prognosis prediction by using multi-source data; periodically acquiring time sequence multi-source data, capturing time evolution characteristics of patient states, and automatically triggering an updating mechanism at a key clinical time point to dynamically adjust a prediction result; generating a risk assessment grade, a feature importance analysis result and an individualized interpretation report based on the prediction result, and performing visualization; and generating a clinical decision using the large language model. According to the technology, accurate evaluation and individualized clinical decision support for hematologic tumor cord blood transplantation prognosis are achieved.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a method, system, medium, and electronic device for generating decision-making on prognostic indicators for hematologic malignancy umbilical cord blood transplantation based on predictive modeling. Background Technology

[0002] Umbilical cord blood hematopoietic stem cell transplantation is an important means of treating malignant hematological diseases. Compared with traditional bone marrow and peripheral blood stem cell transplantation, it has advantages such as less stringent HLA matching requirements, non-invasive collection, low donor risk, and low incidence of graft-versus-host disease, providing treatment opportunities for patients who cannot find a matching donor.

[0003] However, umbilical cord blood transplantation currently faces numerous challenges in clinical application: prognostic assessment relies heavily on physician experience, lacks a systematic and standardized model, and is highly subjective with limited accuracy; existing predictive models primarily target single prognostic indicators, failing to comprehensively assess multidimensional indicators and their interrelationships; integration of heterogeneous clinical data from multiple sources is difficult; models are based on pre-transplant static data, making it difficult to track patients' dynamic states and adjust treatment strategies; machine learning models are often "black boxes," exhibiting poor interpretability and limiting practical application; it is difficult to integrate the experience and knowledge of clinical experts to achieve a data-driven and knowledge-based approach; data gaps affect model training and application, and so on.

[0004] Therefore, there is an urgent need to develop a technology that can integrate multi-source heterogeneous data, achieve multi-objective prediction, have time series analysis capabilities, provide interpretable results, and support individualized decision-making, so as to provide comprehensive prognostic assessment and decision support for umbilical cord blood transplantation for hematological malignancies. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method, system, medium, and electronic device for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling. This invention at least solves the problem that existing umbilical cord blood transplantation prognostic assessments mainly rely on the experience judgment of clinicians, resulting in highly subjective and limited accuracy of the prognostic assessment results.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, this application discloses a method for generating decision-making methods for assessing prognostic indicators in umbilical cord blood transplantation for hematological malignancies based on predictive modeling, the method comprising:

[0010] Acquire multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for patients with hematologic malignancies;

[0011] Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. This expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0012] The system periodically acquires multi-source time-series data and combines it with a recurrent neural network structure to capture the temporal evolution characteristics of patient status. It also automatically triggers an update mechanism at key clinical time points to dynamically adjust the prediction results.

[0013] Based on the prediction results, risk assessment classification, feature importance analysis results, and individualized interpretation reports are generated and visualized.

[0014] Based on the prediction results and expert knowledge graph, clinical decisions are generated using a large language model.

[0015] In one embodiment, the multi-source data includes: patient demographics, disease classification, umbilical cord blood unit characteristics, pretreatment protocols, and genomic data.

[0016] In one embodiment, the integration of expert knowledge based on knowledge graph technology to form an expert knowledge network, and the use of the expert knowledge network as prior knowledge for multi-objective prediction, includes:

[0017] Experienced knowledge was acquired from experts through semi-structured questionnaires and clinical case analysis, and a knowledge graph was constructed in the form of entity-relationship-entity triples; where entities include disease characteristics, treatment plans and prognostic outcomes; and relationships include influencing factors, treatment responses and risk associations.

[0018] By integrating Bayesian network structures with data-driven models, expert knowledge networks are transformed into prior probability distributions and rule sets usable by the model after knowledge conflict detection and consistency evaluation.

[0019] In one embodiment, the key clinical time points include: when hematopoietic reconstitution is completed, when immunosuppressants are adjusted, when complications occur, and during routine follow-up.

[0020] In one embodiment, the feature importance analysis includes using the SHAP value calculation method to identify key factors affecting prognosis, generating a feature contribution waterfall plot and feature interaction analysis;

[0021] The individualized interpretation report includes clinical interpretation and intervention recommendations.

[0022] In one embodiment, the visualization includes:

[0023] The overall risk assessment is presented using an overview layer;

[0024] The details layer displays the specific risks and key influencing factors of each forecast target;

[0025] Use trend layers to display the time trajectory of risk changes.

[0026] In one embodiment, the method further includes:

[0027] Automatic optimization is achieved based on self-learning and adaptation mechanisms.

[0028] Secondly, this application also discloses a predictive modeling-based system for evaluating and generating prognostic indicators in umbilical cord blood transplantation for hematological malignancies, the system comprising:

[0029] The data acquisition module is configured to acquire multi-source data at key time points throughout the entire process of umbilical cord blood transplantation for patients with hematologic malignancies.

[0030] The multi-objective prediction module is configured to integrate expert knowledge based on knowledge graph technology to form an expert knowledge network, and use the expert knowledge network as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, it uses the multi-source data to perform multi-dimensional prognostic prediction. The multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0031] The time series analysis dynamic update module is configured to periodically acquire the multi-source time series data, combine it with a recurrent neural network structure to capture the temporal evolution characteristics of the patient's state, and automatically trigger an update mechanism at key clinical time points to dynamically adjust the prediction results.

[0032] The results interpretation and visualization module is configured to generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and then visualize them.

[0033] The clinical decision generation module is configured to generate clinical decisions using a large language model based on the prediction results and expert knowledge graph.

[0034] Thirdly, this application further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the predictive modeling-based umbilical cord blood transplantation prognostic indicator assessment decision generation method for hematologic malignancies as described in any of the preceding claims.

[0035] Fourthly, this application finally discloses a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for generating prognostic indicators for umbilical cord blood transplantation treatment of hematologic malignancies based on predictive modeling as described in any of the preceding claims.

[0036] (III) Beneficial Effects

[0037] This invention provides a method, system, medium, and electronic device for generating decision-making methods for assessing prognostic indicators in umbilical cord blood transplantation for hematological malignancies based on predictive modeling. Compared with existing technologies, it has the following advantages:

[0038] This application proposes a method for predictive modeling-based assessment and decision generation of prognostic indicators for umbilical cord blood transplantation in hematological malignancies. The method includes: acquiring multi-source data at key time points throughout the entire umbilical cord blood transplantation process for hematological malignancies; integrating expert knowledge using knowledge graph technology to form an expert knowledge network, using this network as prior knowledge for multi-objective prediction; and performing multi-dimensional prognostic predictions using the multi-source data based on an integrated learning architecture of gradient boosting decision trees and deep neural networks. The multi-objective predictions include predictions of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution. The method also involves periodically acquiring the time-series multi-source data and combining it with a recurrent neural network structure to capture the temporal evolution characteristics of the patient's condition, automatically triggering an update mechanism at key clinical time points to dynamically adjust the prediction results; generating risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualizing these results; and generating clinical decisions using a large language model based on the prediction results and the expert knowledge graph. The method in this application achieves a deep integration of expert knowledge and machine learning. By using knowledge graph technology, the experience and judgment rules of clinical experts are formalized and incorporated into the prediction model, which significantly improves the reliability and clinical relevance of the prediction model. It not only utilizes the data mining capabilities of machine learning but also integrates the clinical guidance value of expert knowledge, providing decision support that is more in line with clinical practice. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a method for generating prognostic indicators for umbilical cord blood transplantation for hematological malignancies based on predictive modeling, as described in an embodiment of the present invention.

[0041] Figure 2This is a system principle diagram of the prognostic indicator assessment and decision generation system for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling in an embodiment of the present invention.

[0042] Figure 3 This is a flowchart of multi-source data acquisition and preprocessing in an embodiment of the present invention;

[0043] Figure 4 This is a flowchart of the multi-objective prediction modeling process in an embodiment of the present invention;

[0044] Figure 5 This is a flowchart of the time-series analysis dynamic update process in an embodiment of the present invention;

[0045] Figure 6 This is a flowchart illustrating the visualization process for interpreting results in embodiments of the present invention;

[0046] Figure 7 This is a flowchart of the clinical decision support process in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] With advancements in hematological malignancy treatment technologies, umbilical cord blood hematopoietic stem cell transplantation has become a crucial treatment method for malignant hematological diseases. Compared to traditional bone marrow transplantation and peripheral blood stem cell transplantation, umbilical cord blood transplantation offers advantages such as relatively relaxed HLA matching requirements, non-invasive collection procedures, low donor risk, and a low incidence of graft-versus-host disease (GVHD), providing treatment opportunities for many patients who cannot find a matching donor.

[0049] However, umbilical cord blood transplantation still faces the following key issues and technical bottlenecks in clinical applications:

[0050] Current prognostic assessments for umbilical cord blood transplantation rely heavily on the experience of clinicians, lacking systematic and standardized predictive models. Physicians typically make empirical judgments based on a few indicators (such as HLA matching, CD34+ cell count, and disease status), which suffers from high subjectivity, inconsistent standards, and difficulty in quantification. Furthermore, assessments based on a single indicator cannot fully reflect the complexity of the transplantation process, resulting in limited accuracy in prognostic assessments.

[0051] Umbilical cord blood transplantation involves multiple prognostic indicators, including overall survival, risk of graft-versus-host disease, risk of disease relapse, and immune reconstitution status. Existing predictive models often focus on a single indicator and lack the ability to comprehensively assess multidimensional prognostic indicators. This fragmented predictive approach fails to capture the interrelationships between different prognostic indicators, making it difficult to provide a comprehensive basis for clinical decision-making.

[0052] Clinical data for patients with hematologic malignancies comes from a complex and diverse range of sources, including demographic data, laboratory test results, genomic data, and imaging data. These heterogeneous data differ significantly in format, acquisition standards, and storage methods, and effectively integrating this multi-source heterogeneous data has always been a challenge for clinical decision support systems.

[0053] During transplantation, the patient's condition changes dynamically, and the corresponding prognostic risks also change over time. Most existing predictive models are based on static data before transplantation and lack the ability to track the dynamic changes in the patient's condition during transplantation, making it impossible to adjust the predictive results and treatment strategies in a timely manner.

[0054] Most current machine learning models are "black box" models. Although they may have a certain level of predictive accuracy, they cannot provide clinical interpretation of the prediction results. The interpretability of clinical decision support systems is crucial for physician acceptance, which greatly limits their application value in actual clinical settings.

[0055] Clinical decision-making often involves considering a variety of complex factors, including medical evidence, expert experience, and individual patient differences. The key to improving the usability of a decision support system lies in effectively integrating the experience and knowledge of clinical experts into it, achieving an organic combination of data-driven and knowledge-driven approaches.

[0056] Data gaps are inevitable during clinical data collection, posing a challenge to model training and application. Maintaining stable system operation despite incomplete data is a crucial factor determining the system's practicality.

[0057] In view of the above problems, there is an urgent need to develop an intelligent system that can integrate multi-source heterogeneous data, achieve multi-objective prediction, have time series analysis capabilities, provide interpretable results, and support individualized clinical decision-making, so as to provide comprehensive prognostic assessment and decision support for umbilical cord blood transplantation for hematological malignancies.

[0058] This application provides a method, system, medium, and electronic device for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling. It at least solves the problem that existing umbilical cord blood transplantation prognostic assessments mainly rely on the experience judgment of clinicians, resulting in highly subjective and limited accuracy of the prognostic assessment results, and achieves the goal of deep integration of expert knowledge and machine learning.

[0059] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0060] Example 1:

[0061] Firstly, this invention proposes a method for generating decision-making methods based on predictive modeling for prognostic indicators in umbilical cord blood transplantation for hematological malignancies, see [link to relevant documentation]. Figure 1 The method includes:

[0062] S1. Obtain multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for patients with hematologic malignancies;

[0063] S2. Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. The expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes the prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0064] S3. Periodically acquire the multi-source data in the time series, and combine it with the recurrent neural network structure to capture the time evolution characteristics of the patient's state, and automatically trigger the update mechanism at key clinical time points to dynamically adjust the prediction results.

[0065] S4. Generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualize them;

[0066] S5. Based on the prediction results and expert knowledge graph, generate clinical decisions using a large language model.

[0067] The following is in conjunction with the appendix Figure 1-7 The following details the implementation process of an embodiment of the present invention, including explanations of the specific steps S1-S5.

[0068] This embodiment proposes a predictive modeling-based method for evaluating and generating prognostic indicators for umbilical cord blood transplantation in hematological malignancies. This method is implemented using a pre-built predictive modeling-based system for evaluating and generating prognostic indicators for umbilical cord blood transplantation in hematological malignancies. This system comprises five functional modules: a data acquisition module, a multi-objective prediction module, a time-series analysis and dynamic update module, a results interpretation and visualization module, and a clinical decision support module. For details, please refer to [link to relevant documentation]. Figure 2 These functional modules are interconnected through data interaction interfaces to jointly complete all steps of the prognostic indicator assessment, decision generation, and support for umbilical cord blood transplantation for hematological malignancies.

[0069] First, a predictive modeling-based decision-making system for assessing prognostic indicators in umbilical cord blood transplantation for hematological malignancies was deployed in the hematology department of a tertiary-level hospital. A data interface was established with the hospital's existing electronic medical record system to automatically acquire patient basic information, test results, and treatment records. Simultaneously, a dedicated server was built to run the predictive model and knowledge graph. All functional modules exchange data through a secure internal network to ensure the collaborative operation of all system components. The specific execution steps are as follows:

[0070] S1. Obtain multi-source data at key time points throughout the entire process of umbilical cord blood transplantation for patients with hematologic malignancies.

[0071] The data acquisition module integrates patient clinical data, collecting patient demographics, disease classification, umbilical cord blood unit characteristics, preprocessing protocols, and genomic data. Then, a standardized processing algorithm eliminates heterogeneity from different data sources, achieving comprehensive collection and structured storage of data at key time points before, during, and after treatment. For example... Figure 3 As shown.

[0072] In one embodiment, the aforementioned multi-source data is collected using the patient basic information collection unit, disease characteristic analysis unit, umbilical cord blood characteristic assessment unit, treatment plan recording unit, genomics data analysis unit, and data standardization processing unit within the data acquisition module. For example, for newly enrolled acute myeloid leukemia patients, the patient basic information collection unit first records information such as age, gender, height, weight, and underlying diseases; the disease characteristic analysis unit integrates FAB typing, cytogenetic examination, and molecular biological examination results to assess disease risk level; the umbilical cord blood characteristic assessment unit records parameters such as HLA typing, total cell count, and CD34+ cell count of the selected umbilical cord blood unit; the treatment plan recording unit records the drugs and dosages used in the pretreatment protocol; and the genomics data analysis unit integrates gene sequencing results to identify prognostic-related gene mutations. Finally, the data standardization processing unit performs uniform format conversion, missing value handling, and outlier detection on all collected data to generate a standardized patient dataset. Specifically:

[0073] The patient baseline information collection unit gathers demographic data such as age, sex, body mass index, comorbidities, and past treatment history to establish a patient baseline characteristic model. This patient demographic data, including age, sex, body mass index, comorbidities, and past treatment history, is used to assess the patient's baseline condition and transplant tolerance. The establishment of this baseline characteristic model provides a basis for individualized adjustments to the transplant plan.

[0074] The disease characteristic analysis unit integrates information such as disease type, stage, cell morphology characteristics, surface markers, and chromosome karyotype for assessing disease severity and treatment sensitivity. Disease subtyping information, including disease type, stage, cell morphology characteristics, surface markers, and chromosome karyotype, is used to determine disease severity and treatment sensitivity, and corresponding transplantation strategies and monitoring priorities are formulated based on different subtyping characteristics.

[0075] The umbilical cord blood characterization unit analyzes parameters such as HLA typing, total cell count, CD34+ cell count, cryopreservation time, and collection method to predict implantation capability and long-term survival. Umbilical cord blood unit characteristics are used to assess graft quality and implantation capability, including HLA typing, total cell count, CD34+ cell count, cryopreservation time, and collection method, and these characteristics are used to predict implantation speed and long-term survival.

[0076] The treatment protocol recording unit collects information such as pretreatment protocol type, drug type, dosage, administration time, and radiotherapy dose to assess treatment intensity and toxicity. Pretreatment protocol information, including myeloablative or non-myeloablative protocols, drug type, dosage, administration time, and radiotherapy dose, is used to optimize pretreatment intensity based on patient condition to balance treatment efficacy and complication risk.

[0077] The genomics data analysis unit integrates key gene polymorphisms, mutation status, and expression profile data for molecular-level disease risk assessment. Genomics data is used to assess individual characteristics and molecular-level disease risk, including key gene polymorphisms, mutation status, and expression profiles, and to identify specific gene biomarkers and their prognostic relevance to guide precision medicine.

[0078] The data standardization processing unit ensures the comparability and integrity of data from different sources through format unification conversion, unstructured text extraction, outlier detection, unit normalization, categorical variable coding, and missing data processing.

[0079] It should be noted that data collection and structured storage are carried out at key time points before, during, and after treatment, using a distributed database system for storage, and improving query efficiency through a spatiotemporal index structure. At the same time, data encryption and access control are implemented to protect patient privacy.

[0080] S2. Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. The expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes the prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0081] This study utilizes a multi-objective prediction module, employing knowledge graph technology to integrate the clinical experience and judgment rules of multiple hematology experts, forming an expert knowledge network. This network serves as the model's prior knowledge. Simultaneously, dedicated prediction units are constructed for four key prognostic indicators: survival rate, graft-versus-host disease, disease relapse, and immune reconstitution. A multi-dimensional prognostic prediction is achieved based on an ensemble learning architecture combining gradient boosting decision trees and deep neural networks. For details, please refer to [link to details]. Figure 4 .

[0082] In practical implementation, the multi-objective prediction module includes an expert knowledge extraction and graph construction unit, a knowledge-driven model training unit, and a multi-task learning framework.

[0083] The expert knowledge extraction and graph construction unit acquires experiential knowledge from experts through semi-structured questionnaires and clinical case analysis, constructing a knowledge graph in the form of entity-relationship-entity triples. Entities include disease characteristics, treatment plans, and prognostic outcomes, while relationships include influencing factors, treatment responses, and risk associations. In one embodiment, clinical experience and decision-making rules are collected from five blood transplantation experts through semi-structured questionnaires. The expert knowledge extraction unit transforms this experiential knowledge into a knowledge graph containing relationship triples such as "disease characteristics-impact-survival rate" and "HLA matching degree-impact-GVHD risk."

[0084] The knowledge-driven model training unit transforms expert knowledge into prior probability distributions and rule sets, which are then integrated with the data-driven model through a Bayesian network structure. After knowledge conflict detection and consistency evaluation, the expert knowledge network is transformed into prior probability distributions and rule sets usable by the model. The Bayesian network structure is then used to integrate expert knowledge with the data-driven model, thereby achieving knowledge-guided model training.

[0085] The multi-task learning framework includes a survival prediction unit, a graft-versus-host disease prediction unit, a disease relapse prediction unit, and an immune reconstitution prediction unit. Different algorithms are used to model each prognostic indicator, and the output results of each unit are integrated through a stacked generalization method to achieve multi-dimensional prognostic assessment. The prediction efficiency and accuracy are improved by sharing underlying features.

[0086] The survival prediction unit uses a Cox proportional hazards model combined with a random forest algorithm, taking patient baseline characteristics and umbilical cord blood characteristics as input variables to predict overall survival and event-free survival at 100 days, 1 year, and 5 years. The graft-versus-host disease (GVHD) prediction unit uses the XGBoost classification algorithm, integrating features such as HLA matching, pretreatment protocols, and immunosuppression strategies to predict the risk and severity of acute and chronic GVHD, respectively. The disease relapse prediction unit uses a deep neural network structure, fusing information such as disease subtype, minimal residual disease, and gene mutation characteristics to predict the risk of early (within 1 year) and late (after 1 year) disease relapse. The immune reconstitution prediction unit uses a multi-task learning framework to simultaneously predict the recovery time of neutrophils, platelets, T cells, and B cells, as well as long-term immune function status, providing a basis for infection risk assessment. The ensemble learning architecture integrates the outputs of the above prediction units through a stacked generalization method, uses an attention-based feature weighting algorithm to highlight key predictive factors, and ensures the reliability of the prediction results through confidence calibration, thereby achieving multi-dimensional prognostic assessment of patient status at specific time points.

[0087] S3. Periodically acquire the multi-source data in the time series, and combine it with a recurrent neural network structure to capture the time evolution characteristics of the patient's state, and automatically trigger the update mechanism at key clinical time points to dynamically adjust the prediction results.

[0088] By using a time-series analysis dynamic update module to collect continuous data at set intervals, and combining it with a recurrent neural network structure to capture the temporal evolution characteristics of patient status, the prediction model is automatically triggered to update at key clinical time points, thereby achieving dynamic adjustment of prediction results and continuously optimizing the knowledge graph structure based on new data.

[0089] See Figure 5As shown, in practical implementation, the time-series analysis dynamic update module includes a multi-source time-series data acquisition unit, a feature extraction and fusion unit, a long short-term memory network modeling unit, and an incremental learning update unit. The multi-source time-series data acquisition unit collects dynamic change data such as blood cell counts, biochemical indicators, immune cell subsets, inflammatory factor levels, and clinical symptom scores at a preset frequency. The feature extraction and fusion unit processes multi-source data using a "feature extraction first, fusion later" strategy. The long short-term memory network modeling unit captures the temporal evolution characteristics of the patient's state through a multi-layer LSTM network, identifying short-term fluctuations and long-term trends. The incremental learning update unit triggers model parameter updates at key clinical time points and optimizes the knowledge graph structure through reinforcement learning algorithms to improve the accuracy of time-series predictions. For example, in one embodiment, for patients who have completed transplantation, daily blood routine and biochemical indicator test results are automatically obtained from the electronic medical record system, weekly immune cell subset data is obtained, and changes in clinical symptom scores are recorded. The feature extraction unit extracts trend features, fluctuation features, and periodic features from these time-series data. The long short-term memory network modeling unit predicts the risk of complications and the progress of immune reconstitution within the next 30 days based on these features. When a patient experiences a new clinical event (such as infection or GVHD), the incremental learning update unit automatically triggers model parameter updates and adjusts the weights of relevant relationships in the knowledge graph.

[0090] The time-series analysis includes dynamic changes in data such as blood cell counts, biochemical indicators, immune cell subsets, inflammatory factor levels, and clinical symptom scores. This data is collected periodically at a preset frequency and incorporated into the analysis system. Multi-source data fusion employs a "feature extraction followed by fusion" strategy. First, features are extracted from each type of time-series data separately, and then integrated at the feature level to avoid information distortion caused by direct fusion. The temporal evolution characteristics of patient status are captured using a multilayer long short-term memory (LSTM) network. This network structure includes input, hidden, and output layers, enabling simultaneous identification of short-term fluctuations and long-term trends, and modeling the post-transplant hematopoietic reconstitution process, immune function recovery, and complication development. Key clinical time points include the completion of hematopoietic reconstitution, adjustment of immunosuppressants, occurrence of complications, and routine follow-up. The system automatically triggers prediction model parameter updates at these time points.

[0091] The dynamic adjustment of the prediction results is achieved through incremental learning, which adjusts the model weights based on the latest observation data, so that the prediction results are updated in real time as the patient's condition changes.

[0092] Knowledge graph optimization includes three aspects: entity expansion, relation updating, and weight adjustment. The system identifies inaccurate or missing nodes in the knowledge graph by comparing the prediction results with the actual clinical outcomes, and automatically adjusts the relation weights through reinforcement learning algorithms, so that the knowledge structure can be continuously improved with the accumulation of clinical experience, thereby improving the accuracy of time series prediction.

[0093] S4. Generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualize them.

[0094] The results interpretation and visualization module employs a tiered risk assessment system, presenting prediction results as risk scores categorized into four levels: low risk, medium risk, high risk, and extremely high risk. It also sets multiple warning thresholds and identifies key factors influencing prognosis, generating individualized risk interpretation reports. Specifically, as follows... Figure 6 As shown.

[0095] The results interpretation and visualization module includes a graded risk assessment unit, a feature importance analysis unit, and a personalized interpretation report generation unit. The graded risk assessment unit uses a 0-100 scoring system to classify predicted results into four levels: low risk, moderate risk, high risk, and very high risk. The feature importance analysis unit uses the SHAP value calculation method to identify key factors affecting prognosis, generating a feature contribution waterfall chart and feature interaction analysis. The personalized interpretation report generation unit automatically generates clinical interpretations and intervention direction suggestions based on patient characteristics, providing personalized risk assessment and management recommendations for doctors and patients. For example, in one embodiment, for a specific patient, the system generates a risk score of 72, belonging to the high-risk level. The feature importance analysis unit identifies that the patient's high risk mainly stems from two factors: "HLA incomplete matching" and "high-risk disease grouping," and displays the contribution of each factor through a waterfall chart. The personalized interpretation report generation unit automatically generates a clinical interpretation based on these characteristics: "This patient has an HLA match of only 5 / 6 and belongs to high-risk AML with a complex karyotype. It is recommended to strengthen the GVHD prevention program and consider early interventional preventative measures to reduce the risk of recurrence." In addition, the system also provides intervention recommendations and references for high-risk factors. Specifically:

[0096] The results interpretation and visualization module processes raw data to predict key prognostic indicators for hematologic malignancy umbilical cord blood transplantation, including the probability of acute and chronic graft-versus-host disease, 100-day and 1-year overall survival, disease relapse risk, immune reconstitution rate, infection risk, and graft function assessment. The risk score is based on a 0-100 scale, with 0-40 indicating low risk, 41-70 indicating intermediate risk, 71-90 indicating high risk, and 91-100 indicating very high risk. Different risk levels are distinguished by different colors, and independent scoring and warning thresholds are set for each prediction target.

[0097] Feature importance analysis employs the SHAP value calculation method to identify key factors that significantly impact prognosis, and displays the contribution of each factor in a waterfall plot. Simultaneously, feature interaction analysis is generated to reveal the combined effects of multiple factors. Key features include HLA typing accuracy, CD34+ cell count, pretreatment regimen intensity, disease risk stratification, minimal residual disease status, age, and comorbidity burden. The system automatically generates clinical interpretations based on feature performance.

[0098] The individualized risk interpretation report provides corresponding intervention recommendations based on different characteristics. For example, when HLA typing is not a perfect match and the patient is older, it is recommended to strengthen the prophylactic anti-GVHD regimen; when the disease risk grading is high and minimal residual disease is positive, it is recommended to start prophylactic intervention early after transplantation.

[0099] The visualization interface adopts a layered design, including an overview layer, a details layer, and a trend layer. The overview layer presents the overall risk assessment, the details layer displays the specific risks and key influencing factors of each prediction target, and the trend layer displays the time trajectory of risk changes, supporting the multi-level information needs of doctors and patients.

[0100] S5. Based on the prediction results and expert knowledge graph, generate clinical decisions using a large language model.

[0101] The clinical decision support module integrates a large language model analysis engine. This engine, professionally trained with medical literature and medical records, can understand unstructured medical text and extract key clinical features. Based on the extracted features, it performs matching analysis with an expert knowledge graph to generate personalized treatment recommendations for adjusting immunosuppressive regimens, preventing complications, and following up plans. The module continuously evaluates the intervention's effectiveness during treatment implementation. The clinical decision support module uses a knowledge graph matching process to associate patient characteristics with past cases and expert handling patterns to generate personalized treatment recommendations. Figure 7 As shown.

[0102] In practical implementation, the clinical decision support module includes a medical knowledge training unit, an unstructured medical text understanding unit, a knowledge graph matching analysis unit, and an intervention effect evaluation unit. The medical knowledge training unit fine-tunes the large language model using hematology-specific medical literature and multi-center case data to improve its domain-specific understanding capabilities. The unstructured medical text understanding unit can recognize standard medical abbreviations and clinical colloquial expressions. The knowledge graph matching analysis unit maps patient characteristics to knowledge graph nodes, finding similar cases and treatment patterns. The intervention effect evaluation unit continuously evaluates treatment effectiveness through a "prediction-intervention-observation-adjustment" cycle, adjusting subsequent intervention recommendations based on actual responses. For example, in one embodiment, for a 56-year-old male patient diagnosed with MDS-RAEB2, a 4 / 6 HLA-matched umbilical cord blood transplant is planned. First, the unstructured medical text understanding unit extracts key information from the medical record, including "intermediate-2 IPSS-R score" and "CD34+ cell count of 2.3 × 10^5 / kg," etc. The knowledge graph matching and analysis unit maps these features to corresponding nodes in the graph, and identifies the five most similar historical cases, their treatment plans, and outcomes using an activation diffusion algorithm. Based on these matching results, the system generates the following recommendation: "Consider adopting a non-myeloablative pretreatment regimen containing ATG to strengthen GVHD prevention, and it is recommended to start regular monitoring for minimal residual disease on day 30 post-transplantation, and intervene as early as possible depending on the situation." During the implementation of the regimen, the intervention effect is continuously evaluated, and when the patient develops mild intestinal GVHD, the subsequent immunosuppressant use recommendations are automatically adjusted.

[0103] The medical literature training materials include research on umbilical cord blood transplantation published in authoritative hematology journals, guidelines issued by international blood and bone marrow transplant organizations, and multicenter clinical trial reports, with a focus on transplant indications, pre-treatment regimen selection, prevention and treatment of graft-versus-host disease, and management of post-transplant complications. The case training materials include medical records from multiple transplant centers, covering complete treatment process records for different disease types, risk stratifications, and treatment protocols, with particular emphasis on the criteria for clinical decision points and their correlation with patient follow-up results.

[0104] The training of large language models focuses on understanding the terminology system of hematologic oncology, grasping the key nodes of the transplantation treatment process, identifying the clinical logic of treatment plan adjustments, and predicting the potential effects of interventions. Supervised fine-tuning and contrastive learning methods are used to improve the model's understanding depth in the professional field.

[0105] The medical abbreviation processing module can not only recognize standard medical abbreviations such as CR (complete remission), VOD (hepatic sinusoidal obstruction disease), and CMV (cytomegalovirus), but also understand informal expressions used in daily clinical communication, such as colloquial abbreviations like "hepatic sinusoidal obstruction," "transplant rejection," and "three blood lineages," and achieve accurate conversion through contextual semantic analysis and a professional vocabulary mapping table.

[0106] Knowledge graph matching uses a combination of semantic similarity calculation and rule reasoning to map patient features extracted from clinical records to corresponding nodes in the knowledge graph. The activation diffusion algorithm is used to find the most similar historical cases and expert processing patterns in the graph to the current case.

[0107] The intervention effectiveness evaluation method calculates the actual benefits of the intervention by comparing the difference between the predicted indicators before treatment and the actual observed results, and at the same time, evaluates the treatment satisfaction and quality of life improvement by combining the patient-reported outcomes.

[0108] The continuous evaluation mechanism follows a cyclical process of "prediction-intervention-observation-adjustment". The system will automatically trigger the evaluation process at a preset time point after each intervention, update the prediction results based on the latest clinical data, and adjust subsequent intervention recommendations according to the actual response level to ensure that the treatment plan can adapt to the dynamic changes in the patient's condition.

[0109] S6. Automatic optimization based on self-learning and adaptation mechanisms.

[0110] In a preferred embodiment, in addition to the basic functions described above, the system possesses a self-learning and adaptive mechanism. It can continuously improve the knowledge graph structure from new clinical cases, automatically optimize model parameters when the accuracy falls below a preset threshold by comparing the consistency between predicted results and actual clinical outcomes, and automatically adjust the prediction threshold according to different disease subtypes and patient characteristics. Furthermore, it maintains stable system operation through multiple imputation techniques in the event of data loss, thereby achieving accurate prognostic assessment and individualized clinical decision support for umbilical cord blood transplantation in hematological malignancies. Specifically:

[0111] The method for improving knowledge graphs adopts an incremental learning strategy. By analyzing the differences between newly added clinical cases and existing knowledge structures, potential new entities, new relationships, or knowledge conflicts are identified, and the graph is expanded or revised. Special attention is paid to supplementing knowledge of rare cases and atypical clinical manifestations.

[0112] The knowledge graph improvement report is automatically generated monthly, including the number of new knowledge points, the types of corrected relationships, changes in prediction performance, and analysis of differences in expert consensus. The evolution of the graph structure is presented in a visual format.

[0113] The operator intervention mechanism allows clinical experts to review knowledge graph update suggestions proposed by the system through an interactive interface. They can choose to accept, reject, or modify specific knowledge points and record the basis for adjustments through the annotation function. All human intervention records will serve as important feedback for the system's continuous learning.

[0114] Model parameter optimization is automatically adjusted using a Bayesian optimization algorithm. When the accuracy of any prediction target falls below a preset threshold, the system will explore better parameter combinations while maintaining model stability and generate a performance comparison report before and after optimization.

[0115] The prediction threshold adjustment is based on the results of subgroup analysis. Threshold systems are established for different disease types (such as acute myeloid leukemia, acute lymphoblastic leukemia, and myelodysplastic syndrome), risk stratification, and specific populations (such as children, elderly patients, and patients with comorbidities) to ensure the specificity of the warning.

[0116] Multiple imputation techniques are used to address the unavoidable problem of missing data in clinical practice. A multiple imputation algorithm based on the Markov chain Monte Carlo method is used to infer missing values ​​based on the correlation between known variables, generate multiple complete datasets for model training and prediction, and evaluate the impact of missing data on prediction results through sensitivity analysis.

[0117] The system also features automatic performance monitoring, which regularly assesses the operating status of each module, the utilization rate of computing resources, and predicts deviation trends. When system performance anomalies are detected, it automatically issues maintenance suggestions and supports remote updates and version iterations to ensure that the system can adapt to the ever-evolving clinical practice and medical research progress.

[0118] It's important to note that the self-learning and adaptive mechanisms ensure the system continuously optimizes its performance. After each complete case follow-up cycle, the system automatically compares the predicted results with the actual clinical outcomes, automatically triggering model parameter optimization when a prediction discrepancy is detected. Simultaneously, the system periodically analyzes the prediction accuracy for different disease subtypes and patient characteristics, adjusting prediction thresholds accordingly. For newly emerging clinical patterns, the system automatically extracts them as potential knowledge points, which are then incorporated into the knowledge graph after expert confirmation, achieving continuous improvement of the knowledge structure.

[0119] When dealing with missing data, the system employs a multiple imputation algorithm based on the Markov chain Monte Carlo method to infer missing values ​​based on the correlation between known variables, ensuring stable operation in actual clinical work.

[0120] This completes the entire process of the method for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling, as proposed in the above embodiments.

[0121] Example 2:

[0122] Secondly, this invention also provides a predictive modeling-based system for evaluating and generating decision-making systems for prognostic indicators in umbilical cord blood transplantation for hematological malignancies, see [link to relevant documentation]. Figure 2 The system includes:

[0123] The data acquisition module is configured to acquire multi-source data at key time points throughout the entire process of umbilical cord blood transplantation for patients with hematologic malignancies.

[0124] The multi-objective prediction module is configured to integrate expert knowledge based on knowledge graph technology to form an expert knowledge network, and use the expert knowledge network as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, it uses the multi-source data to perform multi-dimensional prognostic prediction. The multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0125] The time series analysis dynamic update module is configured to periodically acquire the multi-source time series data, combine it with a recurrent neural network structure to capture the temporal evolution characteristics of the patient's state, and automatically trigger an update mechanism at key clinical time points to dynamically adjust the prediction results.

[0126] The results interpretation and visualization module is configured to generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and then visualize them.

[0127] The clinical decision generation module is configured to generate clinical decisions using a large language model based on the prediction results and expert knowledge graph.

[0128] It is understood that the predictive modeling-based prognostic indicator assessment and decision generation system for umbilical cord blood transplantation for hematological malignancies provided in this embodiment of the invention corresponds to the above-mentioned predictive modeling-based prognostic indicator assessment and decision generation method for umbilical cord blood transplantation for hematological malignancies. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the predictive modeling-based prognostic indicator assessment and decision generation method for umbilical cord blood transplantation for hematological malignancies, and will not be repeated here.

[0129] Example 3:

[0130] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method for generating a prognostic indicator assessment decision for hematological malignancy umbilical cord blood transplantation based on predictive modeling as described in any of the above embodiments and their preferred embodiments. The method mainly includes:

[0131] S1. Obtain multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for patients with hematologic malignancies;

[0132] S2. Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. The expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes the prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0133] S3. Periodically acquire the multi-source data in the time series, and combine it with the recurrent neural network structure to capture the time evolution characteristics of the patient's state, and automatically trigger the update mechanism at key clinical time points to dynamically adjust the prediction results.

[0134] S4. Generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualize them;

[0135] S5. Based on the prediction results and expert knowledge graph, generate clinical decisions using a large language model.

[0136] It is understood that the electronic device for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling provided in this embodiment of the invention corresponds to the above-mentioned method and system for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling. The explanation, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the method and system for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling, and will not be repeated here.

[0137] Example 4:

[0138] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for generating a decision-making method for prognostic indicators of hematologic malignancy umbilical cord blood transplantation based on predictive modeling as described in any of the above embodiments and their preferred embodiments, the method comprising:

[0139] S1. Obtain multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for patients with hematologic malignancies;

[0140] S2. Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. The expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes the prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution.

[0141] S3. Periodically acquire the multi-source data in the time series, and combine it with the recurrent neural network structure to capture the time evolution characteristics of the patient's state, and automatically trigger the update mechanism at key clinical time points to dynamically adjust the prediction results.

[0142] S4. Generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualize them;

[0143] S5. Based on the prediction results and expert knowledge graph, generate clinical decisions using a large language model.

[0144] It is understood that the storage medium for generating and storing prognostic indicators for umbilical cord blood transplantation for hematological malignancies based on predictive modeling provided in this embodiment of the invention corresponds to the aforementioned method and system for generating and storing prognostic indicators for umbilical cord blood transplantation for hematological malignancies based on predictive modeling. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the method and system for generating and storing prognostic indicators for umbilical cord blood transplantation for hematological malignancies based on predictive modeling, and will not be repeated here.

[0145] In summary, compared with existing technologies, it has the following beneficial effects:

[0146] 1. This application proposes a method for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling, comprising: acquiring multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for hematological malignancies; integrating expert knowledge based on knowledge graph technology to form an expert knowledge network, using the expert knowledge network as prior knowledge for multi-objective prediction, and using the multi-source data to perform multi-dimensional prognostic prediction based on an integrated learning architecture of gradient boosting decision tree and deep neural network; the multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution; periodically acquiring the time-series multi-source data, and combining a recurrent neural network structure to capture the temporal evolution characteristics of the patient's status, and automatically triggering an update mechanism at key clinical time points to dynamically adjust the prediction results; generating risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and visualizing them; and generating clinical decisions based on the prediction results and expert knowledge graph using a large language model. The method in this application achieves a deep integration of expert knowledge and machine learning. By using knowledge graph technology, the experience and judgment rules of clinical experts are formalized and incorporated into the prediction model, which significantly improves the reliability and clinical relevance of the prediction model. It not only utilizes the data mining capabilities of machine learning but also integrates the clinical guidance value of expert knowledge, providing decision support that is more in line with clinical practice.

[0147] 2. The technology in this application constructs a multi-objective prediction framework that predicts four key prognostic indicators: survival rate, graft-versus-host disease, disease relapse, and immune reconstitution. Through feature sharing and task association, it captures the mutual influence between different prognostic indicators, providing a comprehensive prognostic assessment for clinical decision-making.

[0148] 3. The technology in this application possesses powerful time-series analysis capabilities. By capturing the temporal evolution characteristics of the patient's state through a recurrent neural network structure, it enables dynamic adjustment of the prediction results. The system can update the prediction results and treatment recommendations in a timely manner based on changes in the patient's state during the transplantation process, ensuring the timeliness and relevance of decision support.

[0149] 4. The technology in this application emphasizes the interpretability of prediction results. It identifies key factors affecting prognosis through feature importance analysis and provides clinical interpretations by combining expert knowledge. The risk interpretation report generated by the system clearly identifies high-risk factors and their clinical significance, greatly improving physicians' understanding and acceptance of the prediction results.

[0150] 5. The technology in this application integrates a large language model analysis engine, which, after professional training on medical literature and medical records, can effectively process unstructured medical text, extract key clinical features, and generate personalized treatment suggestions by combining expert knowledge graphs, thus providing intelligent support for clinical decision-making.

[0151] 6. The technology in this application possesses excellent adaptability and robustness, enabling it to continuously improve the knowledge graph structure from new clinical cases, automatically optimize model parameters, and maintain stable system operation through multiple imputation techniques in the event of missing data. This self-learning and adaptive mechanism ensures that the system can continuously improve its performance with the accumulation of clinical experience, adapting to the needs of different patient groups and clinical environments.

[0152] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating decision-making and assessment of prognostic indicators for umbilical cord blood transplantation in hematological malignancies based on predictive modeling, characterized in that, The method includes: Acquire multi-source data at key time points throughout the entire process of umbilical cord blood transplantation treatment for patients with hematologic malignancies; Based on knowledge graph technology, expert knowledge is integrated to form an expert knowledge network. This expert knowledge network is used as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, multi-dimensional prognostic prediction is performed using the multi-source data. The multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution. The system periodically acquires multi-source time-series data and combines it with a recurrent neural network structure to capture the temporal evolution characteristics of patient status. It also automatically triggers an update mechanism at key clinical time points to dynamically adjust the prediction results. Based on the prediction results, risk assessment classification, feature importance analysis results, and individualized interpretation reports are generated and visualized. Based on the prediction results and expert knowledge graph, clinical decisions are generated using a large language model.

2. The method as described in claim 1, characterized in that, The multi-source data includes: patient demographics, disease classification, umbilical cord blood unit characteristics, pretreatment protocols, and genomic data.

3. The method as described in claim 1, characterized in that, The integration of expert knowledge based on knowledge graph technology to form an expert knowledge network, and the use of the expert knowledge network as prior knowledge for multi-objective prediction, includes: Experienced knowledge was acquired from experts through semi-structured questionnaires and clinical case analysis, and a knowledge graph was constructed in the form of entity-relationship-entity triples; where entities include disease characteristics, treatment plans and prognostic outcomes; and relationships include influencing factors, treatment responses and risk associations. By integrating Bayesian network structures with data-driven models, expert knowledge networks are transformed into prior probability distributions and rule sets usable by the model after knowledge conflict detection and consistency evaluation.

4. The method as described in claim 1, characterized in that, The key clinical time points include: when hematopoietic reconstitution is completed, when immunosuppressants are adjusted, when complications occur, and during routine follow-up.

5. The method as described in claim 1, characterized in that, The feature importance analysis includes using the SHAP value calculation method to identify key factors affecting prognosis, generating a feature contribution waterfall plot and feature interaction analysis; The individualized interpretation report includes clinical interpretation and intervention recommendations.

6. The method as described in claim 1, characterized in that, The visualization includes: The overall risk assessment is presented using an overview layer; The details layer displays the specific risks and key influencing factors of each forecast target; Use trend layers to display the time trajectory of risk changes.

7. The method as described in claim 1, characterized in that, The method also includes automatic optimization based on self-learning and adaptation mechanisms.

8. A predictive modeling-based system for evaluating and generating decision-making parameters for prognostic indicators in umbilical cord blood transplantation for hematological malignancies, characterized in that, The system includes: The data acquisition module is configured to acquire multi-source data at key time points throughout the entire process of umbilical cord blood transplantation for patients with hematologic malignancies. The multi-objective prediction module is configured to integrate expert knowledge based on knowledge graph technology to form an expert knowledge network, and use the expert knowledge network as prior knowledge for multi-objective prediction. Based on the integrated learning architecture of gradient boosting decision tree and deep neural network, it uses the multi-source data to perform multi-dimensional prognostic prediction. The multi-objective prediction includes prediction of prognostic indicators such as survival rate, graft-versus-host disease, disease relapse, and immune reconstitution. The time series analysis dynamic update module is configured to periodically acquire the multi-source time series data, combine it with a recurrent neural network structure to capture the temporal evolution characteristics of the patient's state, and automatically trigger an update mechanism at key clinical time points to dynamically adjust the prediction results. The results interpretation and visualization module is configured to generate risk assessment grading, feature importance analysis results, and individualized interpretation reports based on the prediction results, and then visualize them. The clinical decision generation module is configured to generate clinical decisions using a large language model based on the prediction results and expert knowledge graph.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for generating prognostic indicators for umbilical cord blood transplantation treatment of hematological malignancies based on predictive modeling as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for generating prognostic indicators for umbilical cord blood transplantation treatment of hematologic malignancies based on predictive modeling as described in any one of claims 1-7.

Images