Rheumatoid arthritis dynamic prediction method and system based on big data model

Abstract

The present application relates to the technical field of big data model, and particularly relates to a rheumatoid arthritis dynamic prediction method and system based on a big data model. The method comprises the following steps: collecting and preprocessing multi-source heterogeneous time series medical data; constructing a prediction feature set containing static and dynamic time series characteristics; training by using a multi-task time series deep learning model to realize multi-scale dynamic prediction from short-term activity to long-term prognosis; generating a dynamic risk index based on the prediction result and triggering an early warning to provide attribution analysis and treatment simulation. The system comprises a data integration module, a feature calculation module, a model service module and a prediction interaction module. The present application solves the problems of static and single data utilization of existing prediction technology, and realizes continuous and accurate prediction and active intervention support for RA disease.

Description

Technical Field

[0001] This invention relates to the field of big data modeling technology, specifically to a method and system for dynamic prediction of rheumatoid arthritis based on big data models. Background Technology

[0002] Rheumatoid arthritis (RA) is a chronic systemic autoimmune disease characterized by symmetrical polyarthritis, recurrent disease course, and high disability rate. Early diagnosis and accurate assessment are crucial for treatment and prognosis. With the development of medical informatization and artificial intelligence, data-driven RA diagnosis and treatment has become a research direction, but it has significant limitations: Risk assessment methods based on traditional statistical models and static data: Clinicians mainly use statistical models established by cross-sectional studies to assess risk using static data from a single point in time or a limited number of visits. This can roughly stratify the long-term risk of a population, but it cannot capture the dynamic evolution of RA disease course, and the predictive results lack timeliness and individualization.

[0003] The current predictive models rely on a single data source, focusing on structured clinical laboratory indicators. However, RA management involves massive amounts of heterogeneous data from multiple sources. Unstructured data has not been fully explored, and dynamic data has not been systematically integrated, forming "data silos." Integrating multimodal time-series data is a challenge.

[0004] Limitations of machine learning models in time-series prediction and generalization: Machine learning methods have been introduced into RA prediction research, but they are inadequate in handling complex medical time-series data and struggle to model nonlinear interactions and long-term dependencies of clinical indicators. Furthermore, most models are trained and validated on single-center, specific population datasets, resulting in limited generalization ability and a lack of multi-task collaborative prediction frameworks, thus limiting their clinical applicability.

[0005] Lack of real-time dynamic early warning and interpretable decision support loop: Current clinical support systems are mostly retrospective analyses or offline predictions, unable to provide real-time early warnings, making it difficult for doctors to intervene in a timely manner. Existing models are mostly "black boxes," with prediction results lacking interpretability, which undermines doctors' trust, and they do not form a decision support loop, failing to provide scenario simulation tools for adjusting treatment plans.

[0006] In summary, existing rheumatoid arthritis (RA) prediction technologies suffer from staticity and limited data utilization. They also exhibit core shortcomings such as weak dynamic modeling and generalization capabilities, and a lack of real-time, interpretable early warning and decision-making loops. Therefore, a systematic solution is urgently needed that deeply integrates multi-source heterogeneous time-series big data, utilizes advanced time-series deep learning models for accurate dynamic prediction, and provides real-time early warning and interpretable decision support to drive RA diagnosis and treatment towards personalization, precision, and proactive management. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention discloses a dynamic prediction method and system for rheumatoid arthritis based on a big data model, which solves the problems of static prediction, limited data utilization, insufficient model generalization ability, and lack of real-time early warning and decision support in existing technologies.

[0008] This invention is achieved through the following technical solution: On the one hand, this invention provides a dynamic prediction method for rheumatoid arthritis based on a big data model. The method mainly includes the following steps: Step S100: Collection and preprocessing of multi-source heterogeneous medical big data.

[0009] Medical data related to rheumatoid arthritis were collected from multiple data sources. These data sources primarily fall into two categories: first, in-hospital data systems, such as hospital information systems, laboratory information systems, image archiving and communication systems, and electronic medical record systems, from which structured laboratory test sequences, unstructured imaging reports, and text medical records were extracted; second, external dynamic data sources were collected through patient terminal applications or wearable devices, gathering patient self-reported symptom scores, functional status questionnaire results, and objectively monitored joint mobility data. Subsequently, the collected raw data underwent preprocessing: unstructured text data (such as disease descriptions and imaging reports) underwent natural language processing to identify and extract key medical entities and relationships, transforming them into structured features; time-stamped time-series data (such as continuous laboratory indicators and patient self-reported scores) underwent timeline alignment, missing value imputation (using time-series-specific interpolation methods), and data normalization, ultimately constructing a complete longitudinal patient data sequence based on a unified time frame.

[0010] Step S200: Feature engineering construction for dynamic prediction.

[0011] Based on the preprocessed longitudinal data sequence, a feature set is constructed for model training and prediction. This feature set includes: Static baseline characteristics: Extracting relatively stable characteristics of patients at the start of the study or at the time of initial diagnosis, such as age of onset, sex, initial serum antibody status (anti-CCP, RF), and baseline joint involvement pattern.

[0012] Dynamic temporal features: Extracting features reflecting the changing patterns from time-series data, mainly including: 1) Rolling statistical features: Calculating the statistics of key clinical indicators (such as CRP, patient pain VAS score) within a preset sliding time window (e.g., the most recent 1 month, 3 months, 6 months), such as mean, standard deviation, and linear trend slope; 2) Event sequence features: Encoding important clinical events (such as the activation or replacement of a certain disease-modifying antirheumatic drug, the occurrence of extra-articular complications) as event markers and constructing their time series; 3) Imaging evolution features: Quantifying key indicators of joint injury progression (such as the rate of change of Sharp score, changes in synovial blood flow signal level) from continuous imaging examinations (such as series X-rays or musculoskeletal ultrasound) to form a time-series vector.

[0013] Step S300: Training the multi-task dynamic prediction model.

[0014] A multi-task temporal deep learning prediction model is trained using a dataset containing a large amount of longitudinal patient data. The core architecture of the model is preferably a bidirectional long short-term memory network based on an attention mechanism, or a hybrid architecture combining a temporal convolutional network and a Transformer encoder, to enhance the modeling ability of long-term dependencies and key time points. The model employs a multi-task learning framework, simultaneously learning multiple related prediction tasks. Task settings may include: Task 1 (Short-term prediction): Predict the disease activity level for a predetermined short period of time (e.g., the next 1-3 months) (e.g., based on the DAS28-CRP score).

[0015] Task 2 (Interim Prediction): Predict whether clinically significant radiological progression (e.g., an increase in Sharp score exceeding the minimum detectable change) will occur within a predetermined intermediate period (e.g., the next 6-12 months).

[0016] Task 3 (Long-term prediction): Predict the probability of achieving clinical remission or maintaining low disease activity over a predetermined long-term period (e.g., the next 1-2 years).

[0017] Task 4 (Treatment Response Prediction): Predict the patient’s response level to a specific treatment regimen (such as a biologic) (e.g., based on the EULAR criteria).

[0018] During model training, the loss function is a weighted sum of the loss functions of each of the above tasks, and a time smoothness constraint can be introduced to ensure that the predictions output by the model have reasonable continuity in the time dimension.

[0019] Step S400: Dynamic prediction and early warning output.

[0020] The latest, preprocessed, and feature-engineered time-series data of the target patient is input into the trained multi-task dynamic prediction model. The model outputs prediction results corresponding to multiple time scales. The system further executes: Calculate the dynamic risk index: Combine the predicted outputs of various tasks (such as the probability of disease activity worsening, the risk of radiological progression, etc.) and generate a standardized comprehensive risk index (e.g., 0-100 points) through a predefined algorithm. This index dynamically reflects the patient's overall risk of recent disease deterioration.

[0021] Triggering warnings and providing attribution: When the dynamic risk index exceeds a preset threshold, or its rate of increase exceeds a specific range within a short period of time, the system automatically generates a warning signal. Simultaneously, attribution analysis is performed on the prediction using model interpretability techniques (such as attention weight analysis and SHAP value calculation) to identify and highlight key driving factors affecting the prediction results (e.g., "recent CRP trend continues to rise," "methotrexate adherence was less than 70% in the past month").

[0022] Supports decision simulation: Provides a "hypothesis analysis" function, allowing clinicians to virtually modify patients' treatment plans (such as changing drugs or adjusting dosages) in the interactive interface. The system will re-predict the long-term outcome probability under the modified plan based on the model, providing a quantitative reference for treatment plan selection.

[0023] On the other hand, the present invention provides a dynamic prediction system for rheumatoid arthritis based on a big data model for implementing the above-mentioned method. The system includes the following logical modules: Data Integration and Governance Module: Responsible for connecting with various internal hospital information systems and external data sources to achieve automatic or scheduled data extraction. This module has built-in data cleaning, standardization (such as unified medical terminology coding), de-identification, and privacy protection processes to ensure data quality and security compliance.

[0024] Feature computation and storage module: Connected to the data integration and governance module, this module performs computations on the cleaned data according to predefined feature engineering rules, generating static baseline features and dynamic time-series features. The computed features are persistently stored in a time-series database or feature repository for efficient model access.

[0025] The model service and management module encapsulates pre-trained multi-task dynamic prediction models, providing high-performance, low-latency online prediction services via an application programming interface (API). This module also manages the model's lifecycle, including version control, performance metric monitoring, drift detection, and scheduling regular or triggered incremental retraining.

[0026] Dynamic prediction and interactive application module: Serves as the front end for system-user interaction. It includes: Physician-side application: It is usually integrated into the physician's workstation in the form of a web dashboard, which visualizes the patient's dynamic risk index trend line, multi-time scale prediction results, triggered warning list and detailed attribution analysis report, and integrates the aforementioned "hypothesis analysis" simulation tool.

[0027] Management / Research Applications: Providing tools for department managers or clinical researchers to perform risk stratification of patient cohorts, screen subjects who meet specific research criteria, or conduct population-level statistical analysis based on model predictions.

[0028] The beneficial effects of this invention are as follows: This invention proposes a systematic multimodal temporal data fusion framework specifically designed for the field of rheumatoid arthritis. It innovatively quantifies musculoskeletal ultrasound imaging features into high-dimensional temporal vectors, performing precise spatiotemporal alignment and correlation modeling with biomarkers and self-reported symptom temporal data. This overcomes the challenge of data from different medical sources, laying a data foundation for dynamic prediction. Simultaneously, a hybrid temporal deep learning model for multi-task dynamic prediction of rheumatoid arthritis is constructed, introducing an attention mechanism to enhance the model's ability to capture key patterns and turning points in disease progression. Furthermore, a dynamic risk index and intelligent threshold early warning mechanism based on the fusion calculation of multi-timescale prediction results are established, achieving "real-time monitoring" and "trend early warning." Deep integration of model interpretability technology provides key driving factor analysis and treatment decision simulation functions, forming a closed loop from risk prediction to intervention recommendations. Finally, the designed system achieves deep integration of prediction technology with clinical workflows, innovating application models and constructing a digital twin experimental environment for optimizing treatment strategies, forming an intelligent closed-loop system to assist in personalized treatment decision-making. Attached Figure Description

[0029] 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.

[0030] Figure 1 This is a schematic diagram illustrating the principle and steps of a dynamic prediction method for rheumatoid arthritis based on a big data model. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0032] Implementation Scenario 1: Clinical Decision Support in the Rheumatology and Immunology Department of a Tertiary Hospital Reference Figure 1 The diagram illustrates the principle steps of a dynamic prediction method for rheumatoid arthritis based on a big data model. This scenario uses the rheumatology and immunology department of a large tertiary hospital as an example to explain how the system of this invention can be integrated into the existing clinical workflow to serve daily diagnosis and treatment and chronic disease management.

[0033] 1. System Deployment and Data Initialization The hospital's information technology department collaborated with the rheumatology department to deploy a localized server version of the system (or a hybrid cloud architecture compliant with data security regulations). First, the system was integrated with the Hospital Information System (HIS), Laboratory Information System (LIS), and Picture Archiving and Communication System (PACS) through standard interfaces of the hospital information platform (such as HL7, FHIR) or by establishing a data middleware library. For historical data, electronic medical records (EMRs) of patients diagnosed with RA in the department over the past five years were batch-de-identified, extracted, and transformed before being imported into the system database. Simultaneously, the department promoted the use of the accompanying officially certified patient-facing app. After newly enrolled patients signed informed consent forms, their historical data were imported, and they began reporting joint pain and morning stiffness visual analog scale (VAS) daily via the app, and completing a simplified health assessment questionnaire (HAQ) weekly.

[0034] 2. Feature Construction and Model Localization Adaptation The "Feature Calculation and Storage Module" in the system's backend runs automatically. For example, for a patient with an 8-year disease course, the system extracts their static features: age at diagnosis (45 years old) and positive anti-CCP antibody. Simultaneously, it extracts the time series of 32 CRP tests over the past 8 years from their historical LIS data to calculate dynamic features, such as the "average CRP value in the last 6 months" and the "linear trend slope (increasing) of CRP values ​​in the last 3 months." It retrieves the patient's annual bi-hand X-ray images from the PACS and automatically calculates the annual Sharp score using the integrated image analysis submodule (or by calling a third-party AI tool), generating the "annual growth rate of Sharp score" as an image evolution feature. Initially, the system uses a general model pre-trained on multi-center data from across the country. After 3 months of operation, a certain amount of new patient data from the hospital is accumulated, and the incremental learning function of the "Model Service and Management Module" is activated. This allows for fine-tuning of the model using the hospital's data, making its predictions more closely aligned with the characteristics of the hospital's patient population.

[0035] 3. Clinical usage process example Routine ward rounds and early warning management: Dr. Li opened his workstation in the morning. The system plugin panel displayed the list of patients he managed, with a yellow exclamation mark next to the name of patient "Zhao" as a warning. Dr. Li clicked to enter the patient's personalized prediction dashboard. The panel clearly showed that Zhao's dynamic risk index had rapidly climbed from 40 to 75 in the past two weeks. The prediction results showed that the probability of disease activity progressing from mild to moderate within the next month was 85%. The key attribution analysis highlighted: "The system detected that the patient's most recent (7 days ago) musculoskeletal ultrasound showed a 30% increase in the thickness of the synovial membrane of the right wrist joint, and the energy Doppler signal increased from level 1 to level 2." At the same time, "APP records show that the average daily morning stiffness time has exceeded 60 minutes in the past week, and the frequency of analgesic use has increased." Based on this, Dr. Li clicked the "hypothesis analysis" button in the system to simulate adding a biological agent to the patient's treatment. The system prediction showed that after taking this approach, the probability of the patient achieving low disease activity in 3 months could increase from the current 20% to 65%. Combining this quantitative reference with clinical judgment, Dr. Li decided to start intensive treatment after communicating with the patient.

[0036] Regular follow-up assessments: During patient Sun's follow-up visit, Dr. Li not only reviewed the current test results but also retrieved the dynamic risk index trend chart from the system. The chart revealed that Sun's index had steadily declined and remained within the low-risk green zone for four months after the most recent medication change. The system predicted a <10% risk of radiographic progression within the next 12 months. This provided the doctor with data to maintain the current effective treatment plan and allowed them to communicate this optimistic prediction to the patient, enhancing treatment adherence.

[0037] 4. Effect Retrospective Analysis One year after the system went into operation, the department was able to conduct a retrospective analysis. By comparing the time before and after the system's implementation, the average time from warning to doctor intervention for patients with moderate to severe RA activity was shortened from approximately 4 weeks to 1.5 weeks. During the same period, the average DAS28-CRP score of the department's RA patient population decreased by 0.5, and the rate of unplanned hospitalizations due to acute exacerbations decreased by 15%. This preliminarily validates the positive impact of the system on improving clinical management efficiency and patient prognosis.

[0038] Implementation Scenario 2: Hierarchical diagnosis and referral support under regional medical consortia This scenario applies to a rheumatology medical consortium led by a municipal central hospital and composed of multiple community hospitals, aiming to achieve the downward flow of high-quality medical resources and the hierarchical management of patients.

[0039] 1. System cloud deployment and data interconnection This system will be deployed on a regional healthcare cloud platform. The lead hospital will act as the data center, responsible for the unified maintenance and upgrades of the model. Each member community hospital will access the system via a secure dedicated medical network. The data integration module must be compatible with information systems at different hospital levels. For community hospitals with lower levels of IT infrastructure, standardized web forms will be supported for entering core data (such as the number of swollen joints, the number of tender joints, and the patient's overall assessment VAS). Patients will be encouraged to use a unified app to report data, ensuring the availability of basic time-series data.

[0040] 2. Application Mode Community hospital initial screening and monitoring: Community doctors conduct preliminary examinations of suspected arthritis patients and enter the data into the system. The system uses the patient's clinical characteristics (such as joint distribution and antibody status) and short-term symptom time-series data to preliminarily assess the risk of developing typical RA, providing a reference for whether to refer the patient to a specialist hospital for definitive diagnosis. For diagnosed stable RA patients, community doctors primarily rely on the system for monitoring: patients regularly report data via the app, and the system automatically generates monitoring reports. Once a patient's dynamic risk index exceeds the "low-risk threshold" set for the community hospital, the system automatically alerts the community doctor and suggests "initiating an online consultation or referring the patient to the lead hospital."

[0041] Leading hospitals provide remote guidance and admission of high-risk patients: Specialists at the leading hospital can view dashboards of data for all patients within the medical consortium who have triggered alerts through the system platform. Through the online consultation module, combined with the system's attribution analysis and prediction results, they can remotely guide community doctors to adjust treatment plans. For complex patients predicted by the system as having "high risk of radiological progression" or "high probability of poor response to multiple traditional treatments," the leading hospital can reserve appointment slots in advance, achieving accurate referral and admission.

[0042] 3. Implementation Results This model enables continuous and dynamic management of RA patients in the region, allowing high-risk patients to be identified and referred in a timely manner, while patients in stable condition are managed in the community. The outpatient resources of the lead hospital are more concentrated on patients with complex and severe illnesses, improving the efficiency of the overall healthcare system.

[0043] Implementation Scenario 3: Subject Screening and Efficacy Prediction in New Drug Clinical Trials This scenario applies to a Phase III clinical trial conducted by a pharmaceutical company for the treatment of moderate to severe RA with a novel JAK inhibitor.

[0044] 1. Customized system deployment Subject to compliance with GCP guidelines and data privacy protection, the sponsor (pharmaceutical company) may deploy this system in the backend of the electronic data collection (EDC) system for clinical trials, or as a standalone system integrated with the EDC. The system accesses historical data of patients in the screening period already collected by each research center.

[0045] 2. Application in various stages of clinical trials Precise Enrollment Screening: Traditional enrollment criteria are typically based on disease activity at baseline (e.g., DAS28 > 3.2). This invention's system offers more dynamic screening dimensions. Researchers can set screening criteria such as: "Patients whose dynamic risk index has consistently been at intermediate-high risk (>50) for the past 3 months, and whose model predicts poor response to traditional DMARDs (probability > 70%)." The system can automatically and quickly identify potential subjects meeting these criteria from all screened individuals, significantly improving the homogeneity of enrolled patients and the sensitivity of the trial.

[0046] Early prediction of treatment efficacy: During a double-blind trial, the system can generate an early response probability for each subject based on dynamic data (such as symptom trends and early laboratory indicator responses) from the first few weeks or two months after patient enrollment, using its "Treatment Response Prediction" task module. While this prediction is not unblinded to the research team, it can be used for adaptive design analyses during the trial's mid-term. For example, the data monitoring committee can review the prediction results to assess the continued value of the trial or whether sample size adjustments are necessary.

[0047] In-depth analysis after the trial: After the trial is unblinded, the system's attribution analysis and feature importance ranking functions can be used to explore which types of dynamic features (e.g., the rate of CRP decline within 1 month after treatment initiation, the pattern of morning stiffness relief) are most correlated with the final excellent efficacy, thereby discovering new biomarkers or efficacy predictors and providing academic basis for the precise use of drugs after they are marketed.

[0048] 4. Implementation Value This system significantly improves the accuracy and efficiency of clinical trial subject screening and has the potential to provide data support for adaptive clinical trial design through early efficacy prediction, thereby reducing R&D costs and risks.

[0049] In summary, this invention proposes a systematic multimodal temporal data fusion framework specifically designed for the field of rheumatoid arthritis. Its innovation lies particularly in transforming continuously acquired musculoskeletal ultrasound image features into high-dimensional temporal vectors through quantitative scoring, and then performing precise spatiotemporal alignment and correlation modeling of this image evolution temporal data with biomarker temporal data from laboratory information systems and self-reported symptom temporal data from patient terminals. This method overcomes the challenges of inconsistent temporal granularity, complex missing patterns, and semantic heterogeneity in medical data from different sources, laying a high-quality, high-information-density data foundation for subsequent dynamic prediction.

[0050] This invention constructs a hybrid temporal deep learning model for multi-task dynamic prediction of rheumatoid arthritis. Instead of simply applying a general neural network, this model introduces an attention mechanism, enabling it to adaptively focus on the most important time points and data features in the patient's historical data that predict future outcomes (e.g., the model may focus more on changes in indicators before and after the most recent relapse, rather than data from stable periods). This design significantly enhances the model's ability to capture key dynamic patterns and turning points in disease progression.

[0051] This invention establishes a dynamic risk index and intelligent threshold early warning mechanism based on the fusion calculation of multi-timescale prediction results. The core of this mechanism lies not only in providing discrete probabilities of future events, but also in achieving "real-time monitoring" and "trend early warning" of the patient's condition through a continuously changing comprehensive risk index. Furthermore, this invention deeply integrates model interpretability technology, providing visualized and attribution-based analysis of key driving factors, transforming "black box" predictions into medical insights that clinicians can directly understand (such as "which indicator abnormality led to increased risk"). It also innovatively provides treatment decision simulation and deduction functions, realizing a closed loop from risk prediction to intervention recommendations.

[0052] The system designed in this invention achieves deep integration of predictive technology and clinical workflow. Its innovative application model is reflected in the design of a clinical decision-making simulation interface that supports "hypothesis analysis," allowing physicians to virtually adjust treatment parameters within the system and instantly obtain predictive feedback on the model's impact on long-term outcomes. This essentially constructs a digital twin trial environment for treatment strategy optimization, transcending the scope of traditional risk prediction systems that merely provide information prompts, and forming an intelligent closed-loop system to assist in personalized treatment decision-making.

[0053] 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 dynamic prediction method for rheumatoid arthritis based on a big data model, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous medical big data and perform preprocessing. The multi-source heterogeneous medical big data includes in-hospital data and out-of-hospital dynamic data. The preprocessing includes natural language processing of unstructured text to convert it into structured features, and alignment, interpolation and normalization of time-series data to construct a longitudinal data sequence of patients under a unified time axis. S2: Based on the preprocessed data, construct feature engineering for dynamic prediction, including extracting static baseline features and dynamic time series features; S3: Using preprocessed time-series data and constructed features, train a multi-task dynamic prediction model, which adopts an attention-based time-series deep learning architecture to perform prediction tasks at least two different time scales. S4: Input the latest time-series data of the target patient into the trained multi-task dynamic prediction model to obtain prediction results at multiple time scales, and generate a dynamic risk index and early warning information based on the prediction results.

2. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1, characterized in that, In step S1, the in-hospital data includes at least laboratory test results and imaging data extracted from the Hospital Information System (HIS), Laboratory Information System (LIS), Picture Archiving and Communication System (PACS), and Electronic Medical Record (EMR); the out-of-hospital dynamic data includes at least patient self-assessment data and joint range of motion monitoring data collected through patient-side applications or wearable devices.

3. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 2, characterized in that, The series of imaging data includes continuous musculoskeletal ultrasound images, and the preprocessing in step S1 further includes quantizing the features of the continuous musculoskeletal ultrasound images into a temporal vector.

4. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1, characterized in that, In step S2, the dynamic temporal features include: rolling statistical features of key laboratory indicators and patient self-assessment indicators within a sliding time window, event sequence features of changes in treatment drugs or acute exacerbations, and image evolution features quantified from continuous imaging examinations.

5. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1, characterized in that, In step S3, the architecture of the multi-task dynamic prediction model is a combination of a bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism, or a hybrid model combining a temporal convolutional network (TCN) and a Transformer encoder.

6. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1 or 5, characterized in that, In step S3, the prediction tasks at least two different time scales include: predicting the disease activity score level for the next 1-3 months, and predicting whether radiological progression will occur within the next 6-12 months.

7. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 6, characterized in that, The predictive tasks also include: predicting the probability of achieving clinical remission or low disease activity in the next 1-2 years, and / or predicting the patient's treatment response level to a specific targeted drug.

8. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1, characterized in that, In step S4, generating the dynamic risk index includes: calculating a comprehensive risk score by integrating prediction results from multiple time scales; and triggering the early warning information when the dynamic risk index exceeds a preset threshold or rises sharply.

9. The method for dynamic prediction of rheumatoid arthritis based on a big data model according to claim 1 or 8, characterized in that, Step S4 also includes: providing attribution analysis of key factors affecting prediction results, and hypothesis analysis functions to support simulation of treatment adjustments and prediction of changes in long-term outcomes.

10. A dynamic prediction system for rheumatoid arthritis based on a big data model, used to implement the dynamic prediction method for rheumatoid arthritis based on a big data model as described in any one of claims 1-9, characterized in that, include: The data integration and governance module is used to connect to multi-source heterogeneous data sources and perform data extraction, cleaning, standardization, and privacy anonymization. The feature calculation and storage module is used to calculate and store static baseline features and dynamic time series features based on preprocessed data according to predefined rules. The model service and management module is used to encapsulate pre-trained multi-task dynamic prediction models, provide prediction application programming interfaces (APIs), and manage model versions, monitor performance, and perform periodic retraining. The dynamic prediction and interactive application module is used to provide dynamic risk trend display, early warning information, attribution analysis and decision support functions to doctors and / or patients based on the prediction results provided by the model service and management module.

Images