An autoimmune disease patient prognosis management system

Abstract

This invention relates to the field of medical management system technology, specifically disclosing a prognostic management system for patients with autoimmune diseases. It adopts an edge-cloud collaborative architecture, comprising a multimodal edge data acquisition and preprocessing module, an edge-end real-time status monitoring module, a cloud-based progressive prognostic assessment module, an intervention intensity matching and adverse reaction early warning module, an intervention effect closed-loop correction and federated learning module, and a doctor-patient shared decision-making and data feedback module, all connected sequentially. This invention simultaneously collects multi-dimensional patient data, achieving a comprehensive understanding of the patient's condition. Simultaneously, it performs preliminary data quality control and real-time status monitoring at the edge, reducing the computational burden on the cloud and enabling timely identification of sudden abnormalities in the patient's condition and triggering local early warnings. By calculating the effective contribution coefficients of each modality of data, the impact of different data on prognosis can be dynamically assessed, improving the effectiveness of data utilization.

Description

Technical Field

[0001] This invention relates to the field of medical management system technology, specifically to a prognosis management system for patients with autoimmune diseases. Background Technology

[0002] Autoimmune diseases are a group of chronic diseases characterized by prolonged course, recurrent flare-ups, and a high risk of irreversible damage to multiple target organs. Their prognosis management directly impacts patients' quality of life and long-term outcomes. Currently, clinical prognosis management for autoimmune diseases largely relies on the subjective judgment of healthcare professionals, lacking systematic and intelligent technological support, and presents numerous unresolved issues.

[0003] In current technologies, disease monitoring is largely limited to clinical diagnosis and treatment data and laboratory test data. It fails to effectively integrate continuous physiological data collected by wearable devices, patient-reported outcome data, environmental exposure data, and medication behavior data. Data from different sources are isolated, forming data silos that cannot comprehensively reflect the patient's true disease status. At the same time, existing prognostic assessments are mostly static, single-assessment methods, which are difficult to dynamically correct for disease activity and cannot accurately predict the risk of disease recurrence and the trend of cumulative damage to target organs. Risk warnings often lag behind changes in the disease, missing the best intervention opportunity.

[0004] In terms of intervention program development, existing systems mostly adopt standardized, general protocols, failing to match the intervention intensity to the individual patient's risk level, which easily leads to under- or over-intervention. Furthermore, current technologies provide a rather coarse assessment of adverse drug reactions, failing to fully consider the patient's genotype, liver and kidney function, and concomitant medications, increasing medication safety risks.

[0005] Regarding model construction and updates, traditional centralized model training methods require the aggregation of raw patient data from multiple centers, posing a serious risk of patient privacy breaches. Furthermore, the limited data volume from a single hospital leads to insufficient model generalization ability. Simultaneously, existing systems lack an effective closed-loop feedback mechanism for intervention effects; changes in patient conditions after intervention cannot be used to correct prognostic assessment models and intervention protocol libraries, hindering continuous performance improvement. In addition, patient participation in prognostic management is low, and the lack of effective decision-making channels between doctors and patients affects adherence to intervention protocols. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a prognostic management system for patients with autoimmune diseases, which solves the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a prognostic management system for patients with autoimmune diseases, comprising the following steps: adopting an edge-cloud collaborative architecture, including sequential communication connections to form a closed-loop process: The multimodal edge data acquisition and preprocessing module is configured to simultaneously collect six types of data—clinical diagnosis and treatment, laboratory testing, wearable physiology, patient-reported outcomes, environmental exposure, and medication behavior—through various intelligent devices deployed on the patient end. It performs edge data noise reduction, time alignment, and preliminary quality control, and calculates the effective contribution coefficient of each modality of data. The edge-end real-time status monitoring module calculates the patient's real-time symptom risk index based on preprocessed high-frequency data, identifies sudden abnormal states and triggers local warnings, while encrypting and uploading standardized data to the cloud. The cloud-based progressive prognostic assessment module constructs a four-level assessment model that integrates multi-scale temporal attention mechanisms, and sequentially outputs the current disease activity correction value, the probability of recurrence risk in the next 4 weeks, the cumulative target organ damage risk index, and the acute complication risk index. The intervention intensity matching and adverse reaction early warning module calculates the overall intervention intensity by comprehensively considering the risk of recurrence, target organ damage and complication, and generates a step-by-step intervention plan. At the same time, it assesses the risk of adverse drug reactions based on the patient's genotype, liver and kidney function and medication history. The intervention effect closed-loop correction and federated learning module tracks the changes in the condition after the intervention is implemented, calculates the intervention effect score and back-corrects the model parameters and the matching rules of the protocol library. The incremental federated learning framework is used to achieve multi-center model collaborative optimization. The doctor-patient shared decision-making and data feedback module synchronously displays prognostic assessment results and intervention plans to both doctors and patients, collects patient feedback data, and sends it back to the multimodal edge data acquisition and preprocessing module.

[0008] Preferably, the multimodal edge data acquisition and preprocessing module includes: The clinical diagnosis and treatment data collection unit automatically synchronizes patients’ SLEDAI-2K, DAS28, BASDAI disease activity scores, affected organ grades, past relapse history and medication history through the electronic medical record interface. The data is updated automatically after each outpatient visit. The laboratory test collection unit connects to the hospital's LIS system and automatically imports data on autoantibody titers, complement C3 / C4, erythrocyte sedimentation rate, C-reactive protein, liver and kidney function, and drug blood concentration. During the active phase, the collection frequency is once every 3-7 days, and during the remission phase, it is once every 1-3 months. The wearable physiological data collection unit communicates with smart bracelets and ECG patches to continuously collect data on heart rate, heart rate variability, sleep stages, daily activity levels, and skin temperature. The sampling interval is no more than 1 minute, and during nighttime sleep, the interval is increased to 10 seconds. The patient report collection unit collects fatigue scores, VAS scores for joint pain, rash area, number of oral ulcers, and gastrointestinal reactions daily via a mobile application, and collects the results of the anxiety and depression self-rating scale once a week. The environmental exposure data collection unit connects to the meteorological platform and patient location data, automatically collecting daily ultraviolet radiation duration, PM2.5 concentration, temperature, humidity, and pollen index at 1-hour intervals. The medication behavior collection unit records the time and dosage of each medication through a smart pillbox, automatically identifies missed doses, late doses, and incorrect doses, and supports patients to manually enter temporary medication adjustment information. A dynamic data acquisition priority adjustment unit is used to automatically adjust the acquisition priority of each modality of data based on the patient's current disease activity, including: S1: Obtain the patient's current disease activity correction value A, and classify it into three levels: low activity A<0.3, medium activity 0.3≤A<0.6, and high activity A≥0.6; S2: For patients with low activity levels, the sampling interval for wearable physiological data will be adjusted to 5 minutes, and the frequency of laboratory test collection will be adjusted to once every 3 months; S3: For patients with moderate activity, the sampling interval for wearable physiological data will be adjusted to 1 minute, and the frequency of laboratory test collection will be adjusted to once a month; S4: For patients with high activity levels, the sampling interval for wearable physiological data will be increased to 10 seconds, the frequency of laboratory test collection will be adjusted to once every 3 days, and a daily symptom-specific collection will be added. S5: When the real-time symptom risk index Q>0.7 is detected, the collection frequency of all modal data will be temporarily increased to the highest level, and will be restored to the default frequency of the corresponding activity level after 24 hours.

[0009] Preferably, the multimodal edge data acquisition and preprocessing module performs the following operations: Step X1: Time alignment processing, align all modal data to a time step of 1 day, calculate the daily mean, maximum value and coefficient of variation for high-frequency wearable data, and use cubic spline interpolation to complete the time series for low-frequency test data; Step X2: Outlier identification. Establish an indicator fluctuation range model related to disease activity. When an indicator exceeds the range but is consistent with the trend of disease activity change, it is determined to be a valid outlier. Otherwise, it is determined to be noise and replaced by the median of the three valid values ​​before and after. Step X3: Standardization processing: Minimum-maximum scaling is used to map clinical scores and test indicators to the [0,1] interval; Z-score standardization is used for wearable physiological data; and logarithmic standardization is used for environmental data. Step X4: Calculate the effective contribution coefficient of each modal data. The formula is: ; in, For the first The effective contribution of modal data to prognostic assessment, with a value ranging from 0 to 1. The larger the value, the higher the quality of the modal data and the greater its contribution to the assessment results. For the first The amount of effective data for a modality within the statistical period; For the first The total amount of data that should be collected for a modality within the statistical period; For the first The number of days since the last valid modal data collection; This is the time decay coefficient, ranging from 0.15 to 0.25, used to control the impact of data timeliness on contribution. This is the decay rate coefficient, with a value of 0.08-0.12, used to adjust the rate of time decay; For the first The basic weights of the modalities are predefined by clinical experts and satisfy the following conditions: Clinical diagnosis and treatment data Laboratory test data Wearable physiological data Patient Reported Data Environmental exposure data Medication behavior data When any mode At that time, a reminder to supplement the data collection is sent to the patient's terminal; The multimodal edge data acquisition and preprocessing module also includes a data quality grading and storage optimization unit, used to grade and store data according to effective contribution coefficients, including: T1: Effective contribution coefficient The data is marked as high-quality data and permanently stored on cloud servers; T2: Effective contribution coefficient The data is labeled as medium quality data and will be stored for 5 years. T3: Effective contribution coefficient The data is marked as low-quality data and stored for one year, after which it will be automatically deleted. T4: When the effective contribution coefficient of low-quality data increases to above 0.65 after supplementary collection, it will be automatically upgraded to medium-quality data and the storage period will be adjusted.

[0010] Preferably, the edge-end real-time status monitoring module performs the following operations: Step Y1: Extract the symptom data and physiological data for the day, and calculate the real-time symptom risk index. The formula is: ; in, This is the patient's real-time symptom risk index for the day, ranging from 0 to 1. A higher value indicates a higher risk of sudden deterioration of the condition. For the first The severity of symptoms is scored on a scale of 0 to 3 (0 points for no symptoms, 1 point for mild, 2 points for moderate, and 3 points for severe). Risk weights are assigned to corresponding symptoms, including joint pain. ,fever ,rash fatigue Gastrointestinal reactions ; This represents the average heart rate for the day. The baseline heart rate during the patient's remission period is obtained by the system automatically calculating the average heart rate during the patient's lowest activity level over the past 3 months; Step Y2: When When this occurs, a local audible and visual alarm is triggered, and the data for the day is immediately uploaded to the cloud. Step Y3: When At that time, the standardized data for the day is encrypted and uploaded to the cloud every day at midnight; The edge-end real-time status monitoring module also includes a local anomaly data tracing and verification unit, used to trace and verify the anomaly data that triggers warnings, avoiding false warnings, including: U1: Extract the abnormal data segment that triggered the warning and check the operating status of the corresponding acquisition device and the data transmission link; U2: Compare the changing trends of other modal data of the patient during the same period to determine whether the abnormal data is correlated with other modal data; U3: Sends an abnormal data confirmation request to the patient's terminal, allowing the patient to manually confirm whether there are corresponding symptoms or equipment malfunctions; U4: If the problem is confirmed to be a device malfunction or data transmission error, cancel the alert and mark the data as invalid; if the problem is confirmed to be a genuine change in the patient's condition, retain the alert and immediately upload it to the cloud.

[0011] Preferably, the cloud-based progressive prognostic assessment module includes: The multi-scale temporal feature encoding unit uses a three-layer bidirectional long short-term memory network to encode multimodal data from the past 7 days, 30 days, and 90 days, respectively, and outputs a hidden state sequence. , , ; The cross-scale attention fusion unit calculates attention weights at different time scales, and the steps are as follows: Step Z1: Calculate attention scores at each time scale ,in , For learnable weight matrix, For bias terms; Step Z2: Perform Softmax normalization on the attention scores to obtain cross-scale attention weights. ; Step Z3: Fuse to obtain the comprehensive temporal feature vector ; The Level 4 risk assessment unit calculates the following four related indicators in sequence: (1) Current disease activity correction value The formula is: ; in, This represents the corrected current disease activity level for the patient, ranging from 0 to 1. A higher value indicates higher disease activity. For the first The mapping function of modal features to disease activity was obtained by training a random forest model containing data from more than 10,000 patients. (2) Risk of recurrence in the next 4 weeks The formula is: ; in, This represents the probability of disease recurrence within the next 4 weeks, with a value ranging from 0 to 1. This is the Sigmoid activation function, used to map the model output to the 0-1 range; , These are the learnable parameters for the recurrence risk prediction model; To ensure patient adherence to medication over a 30-day rolling period; The environmental risk correction coefficient is calculated by weighting the duration of ultraviolet radiation, PM2.5 concentration, and pollen index. The value ranges from 0.5 to 1.5. The larger the value, the stronger the promoting effect of environmental factors on recurrence. (3) Target organ cumulative injury risk index The formula is: ; in, This is the cumulative risk index for damage to various target organs of the patient, with a value ranging from 0 to 1. The higher the value, the higher the risk of damage to the target organ. The patient's disease duration (in years); For the patient in the course of the disease The average recurrence risk probability per year; For the first Damage weighting coefficients for each affected organ, including the kidneys. ,heart ,lung ,joint ,skin ; (4) Risk index of acute complications The formula is: ; in, It is a risk index for patients to develop acute complications (such as lupus crisis, severe infection), with a value range of 0-1, which combines the risk of long-term relapse and the risk of short-term symptoms; The cloud-based progressive prognostic assessment module also includes a cross-patient prognostic similarity matching unit, used to match historical patients with similar prognostic characteristics to the current patient, assisting in prognostic assessment, including: V1: Extract the comprehensive temporal feature vector of the current patient. Information on disease type, course of disease, and affected organs; V2: In the cloud-based historical patient database, cosine similarity is used to calculate the feature similarity with the current patient; V3: Select the top 10 historical patients with the highest similarity and extract their actual prognostic results and the effectiveness of the intervention plan; V4: Uses the prognostic statistics of similar patients as a reference and supplements them in the doctor's prognostic assessment report.

[0012] Preferably, the cloud-based progressive prognostic assessment module further includes a dynamic risk grading and source tracing unit, which performs the following steps: Step AA1: Based on the patient's individual historical data and data from the same disease population, a Bayesian optimization algorithm is used to determine three risk thresholds. , , ; Step AA2: Classification of recurrence risk levels: Low risk (Level 1) It is classified as low to medium risk (Level 2). It is classified as medium to high risk (Level 3). High risk (Level 4); Step AA3: Calculate the cumulative recurrence risk value over the past 14 days. The formula is: ; in, The cumulative relapse risk value over the past 14 days is used to assess the overall trend of the recent disease. For the first The probability of recurrence within 1 day; This is the recent risk weighting coefficient, ranging from 0.1 to 0.2, used to assign higher weight to recent data; when Furthermore, if the current risk level is level 2, it will automatically be upgraded to a level 3 warning. Step AA4: Risk factor tracing. The contribution of each feature to the recurrence risk is calculated using multi-scale SHAP values. The top 3 features with the highest absolute values ​​are selected as the main risk factors, and a risk tracing report is generated. Step AA5: When the risk level rises to level 3 or above, or Immediately send early warning information and risk tracing reports to the attending physician's terminal; The cloud-based progressive prognostic assessment module also includes a risk warning and graded response scheduling unit, which automatically schedules corresponding medical resources according to different warning levels, including: W1: Level 1 Warning (Low Risk): Health reminders are sent only to patients, without notifying medical staff; W2: Level 2 Warning (Low to Medium Risk): Push intervention suggestions to the patient's terminal and send a reminder to the responsible nurse's terminal at the same time; W3: Level 3 Warning (Medium-high risk): Push emergency intervention suggestions to the patient's terminal, and send warning information to the terminal of the attending physician and the responsible nurse, requiring follow-up within 24 hours; W4: Level 4 Warning (High Risk): Immediately activate the emergency green channel, send warning information to the emergency department, attending physician and patient's family simultaneously, and push pre-hospitalization notice.

[0013] Preferably, the intervention intensity matching and adverse reaction early warning module includes: Intervention intensity calculation unit, calculates the overall intervention intensity The formula is: ; in, The overall intervention intensity, ranging from 0 to 1, considers the risks of recurrence, target organ damage, and acute complications; based on... The value of is used to divide the intervention intensity into four levels: Basic level, For the enhanced level, It is an enhanced level. Level 3: Emergency; The medication adjustment recommendation unit generates tiered suggestions based on the intensity of the intervention: Basic level: Maintain the current medication dosage and have a follow-up examination every 3 months; Enhanced level: Increase the immunosuppressant dose by 10%-15%, or shorten the dosing interval of biologics by 10%; Enhanced level: Add low-dose glucocorticoids (prednisone 10-20mg / day), and adjust the type of immunosuppressant; Emergency level: Immediately activate the emergency green channel, send a pre-hospitalization notice, and recommend high-dose hormone pulse therapy; Adverse reaction risk assessment unit calculates the adverse drug reaction risk index. The formula is: ; in, The adverse reaction risk index for patients using the current medication regimen, with a value ranging from 0 to 1; For the first The basic adverse reaction risk of this drug is obtained from the drug's instructions and clinical data statistics; The individual risk correction factor is calculated based on the patient's liver and kidney function, genotype, age, and concomitant medications; when If the drug is deemed high-risk, an alternative treatment should be recommended. The Lifestyle and Follow-up Planning section generates personalized recommendations: Exercise duration: ,in The baseline exercise duration is set by the doctor based on the patient's physical condition. Follow-up examination interval: (Unit: days, minimum 7 days); The follow-up examination items are automatically matched according to the affected organs. For example, if the kidneys are involved, a 24-hour urine protein quantification test will be added. The intervention intensity matching and adverse reaction early warning module also includes an intervention protocol personalization adaptation unit, used to adjust the intervention protocol according to the patient's lifestyle and preferences, including: X1: Collect information on the patient's daily routine, dietary habits, exercise preferences, and work schedule; X2: Adjust the medication reminder time according to the patient's daily routine to avoid conflicts with work or rest; X3: Recommend corresponding exercise methods based on the patient's exercise preferences, such as recommending swimming instead of brisk walking if the patient prefers swimming; X4: Adjust dietary recommendations according to the patient's eating habits, and avoid recommending foods that the patient dislikes or cannot obtain; X5: Generates the final personalized intervention plan, which is then implemented after patient confirmation.

[0014] Preferably, the intervention effect closed-loop correction and federated learning module performs the following operations: Step BB1: Collect patient data on days 7, 14, and 28 after intervention implementation, and calculate intervention effectiveness scores. The formula is: ; in, The effectiveness of the intervention program is scored, with a value ranging from -1 to 1. A larger value indicates a better intervention effect. , , The disease activity correction value, relapse risk probability, and real-time symptom risk index before intervention; , , The corresponding value after intervention; Step BB2: Grading of Intervention Effects: For significant effectiveness, For it to be effective, Invalid; Step BB3: Model Correction: When When, keep the parameters unchanged; when When, the matching weight of the corresponding intervention plan is increased proportionally; when At that time, incremental training of the model is triggered; Step BB4: Incremental Federated Learning Update: Each edge node trains a local model based on the latest local data and calculates the incremental gradient. ; The incremental gradient is encrypted using a homomorphic encryption algorithm and then uploaded to the cloud. Federated average aggregation is performed in the cloud to generate a global incremental gradient. ,in The number of edge nodes participating in federated learning; The global model is updated in the cloud and encrypted and distributed to each edge node. The intervention effect closed-loop correction and federated learning module also includes a model incremental update triggering and verification unit, used to control the frequency and quality of model incremental updates and avoid model drift, including: Y1: Set the minimum trigger interval for incremental model updates to 1 month to avoid frequent updates that could lead to model instability; Y2: When incremental training of the model is triggered, use local data from the most recent 3 months for training; Y3: After training is complete, use an independent validation set to calculate the prediction accuracy of the updated model; Y4: If the accuracy of the updated model improves by ≥2%, the new model will be adopted; otherwise, the original model will be retained and the model will be waited for the next update.

[0015] Preferably, the autoimmune disease patient prognosis management system further includes a multimodal loss robustness module, which performs the following operations: Step CC1: When a certain mode When data is missing, renormalize the effective contribution coefficients of other available modes: ; Formula explanation: For the renormalized first The effective modal contribution coefficient ensures that the sum of the contribution coefficients of all available modes is 1. Step CC2: Based on generative adversarial networks, generate virtual features for missing modalities using available modal data. ; Step CC3: Calculate the robust fusion feature vector The formula is: ; in, This is the robust fusion feature vector when modality is missing, with the same dimension as the original fusion feature vector; The feature vectors of the available modes; This is the virtual feature weight coefficient, with a value of 0.2-0.4, used to control the influence of virtual features on the fusion result; Step CC4: Use Replace the original fusion features for prognostic assessment, and send a data missing reminder to the patient's terminal; The intervention effect closed-loop correction and federated learning module also includes a missing modality data completion quality assessment unit, used to evaluate the completion quality of virtual features and dynamically adjust the weights of virtual features, including: Z1: After the missing modal data is completed, calculate the mean square error (MSE) between the virtual features and the real features; Z2: If MSE < 0.1, then the virtual feature weights will be adjusted. Adjusted to 0.4; Z3: If 0.1 ≤ MSE < 0.3, then the virtual feature weights will be adjusted. Adjusted to 0.3; Z4: If MSE ≥ 0.3, then the virtual feature weights will be adjusted. Adjust it to 0.2 and mark the completed data as low confidence.

[0016] Preferably, the doctor-patient shared decision-making and data feedback module includes: The visualization unit simultaneously displays disease activity curves, recurrence risk trend charts, target organ damage radar charts, and adverse reaction risk warnings to both doctors and patients. The prognostic simulation unit supports simulating prognostic changes under different intervention protocols and calculating the simulated risk of recurrence. ,in The effectiveness coefficient of the intervention program is derived from historical data in the intervention program database. To simulate the intensity of intervention; The decision recording unit automatically records the discussion process between doctors and patients, the selection of treatment options, and the patient's informed consent, generating electronic medical record archives. The data feedback unit encrypts and transmits patient symptom feedback, medication administration, and intervention effect data back to the multimodal edge data acquisition and preprocessing module, triggering the next round of prognostic assessment. The patient decision-making preference learning unit is used to learn patients' decision-making preferences and optimize subsequent intervention recommendations, including: AA1: Record the patient's acceptance or rejection of each recommended intervention, and the reason for rejection; AA2: Extract characteristics of patients who refuse treatment plans, such as high risk of drug side effects, high exercise intensity, and high frequency of follow-up examinations; AA3: When recommending interventions in the future, reduce the matching weight of protocols that include patient rejection characteristics; AA4: Update the patient's decision preference model every 3 months to ensure that the recommended plan is more in line with the patient's needs.

[0017] This invention provides a prognostic management system for patients with autoimmune diseases. Employing an edge-cloud collaborative architecture, it effectively addresses the aforementioned problems in existing technologies through multimodal data fusion, progressive dynamic prognostic assessment, intervention intensity matching, and a closed-loop correction mechanism. It offers the following beneficial effects: 1. This invention utilizes a multimodal edge data acquisition and preprocessing module to simultaneously collect multi-dimensional patient data, achieving a comprehensive understanding of the patient's condition. Simultaneously, it performs preliminary quality control and real-time status monitoring of the data at the edge, reducing the computational burden on the cloud while enabling timely identification of sudden abnormalities in the patient's condition and triggering local alerts. By calculating the effective contribution coefficients of each modality of data, the impact of different data on prognosis can be dynamically assessed, improving the effectiveness of data utilization.

[0018] 2. This invention constructs a four-level progressive prognostic assessment model that integrates multi-scale temporal attention mechanisms. It sequentially outputs the current disease activity correction value, the probability of future recurrence, the cumulative target organ damage risk index, and the acute complication risk index, achieving a comprehensive prognostic assessment from current status assessment to trend prediction. Combined with dynamic risk grading and source tracing technology, it can accurately locate key risk factors leading to disease deterioration and automatically allocate corresponding medical resources according to the risk level, enabling early detection, early warning, and early intervention of risks.

[0019] 3. This invention, through an intervention intensity matching and adverse reaction early warning module, calculates the overall intervention intensity by comprehensively considering the patient's recurrence risk, target organ damage risk, and acute complication risk, generating a tiered, personalized intervention plan. Simultaneously, it assesses the risk of adverse drug reactions based on the patient's individual characteristics, effectively avoiding insufficient or excessive intervention and improving medication safety. Furthermore, the system can adjust the intervention plan according to the patient's lifestyle and preferences, further improving patient compliance.

[0020] 4. This invention employs an incremental federated learning framework to achieve collaborative optimization of multi-center models. Each edge node only uploads encrypted model gradients, without transmitting original patient data. While strictly protecting patient privacy, it integrates clinical data from multiple centers to improve the model's generalization ability. Simultaneously, through an intervention effect closed-loop correction module, it can track changes in the patient's condition after intervention, calculate intervention effect scores, and reverse-correct model parameters and protocol library matching rules, achieving continuous iterative improvement in system performance.

[0021] 5. This invention incorporates a multimodal missing data robustness module. When data for a particular modality is missing, robust fusion features are generated by renormalizing the weights of available modalities and generating virtual features. This ensures the system continues to operate normally even with incomplete data, improving system stability and reliability. Furthermore, through the doctor-patient shared decision-making and data feedback module, bidirectional synchronization of prognostic assessment results and intervention plans is achieved, supporting joint decision-making by doctors and patients and enhancing patient participation and satisfaction in the prognostic management process. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the prognosis management system for patients with autoimmune diseases as described in this invention. Figure 2 This is a block diagram illustrating the principle of the multimodal edge data acquisition and preprocessing module of the present invention. Figure 3 This is a block diagram illustrating the principle of the cloud-based progressive prognostic assessment module of the present invention. Figure 4 This is a block diagram illustrating the principle of the intervention intensity matching and adverse reaction early warning module of the present invention. Figure 5 This is a block diagram illustrating the principle of the shared decision-making and data feedback module for doctors and patients in this invention. Figure 6 This is a block diagram illustrating the principle of the closed-loop correction and federated learning module for intervention effects in this invention. Detailed Implementation

[0023] 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, and 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.

[0024] like Figures 1-6 As shown, this invention provides a technical solution for a prognostic management system for patients with autoimmune diseases. It adopts an edge-cloud collaborative architecture, including a multimodal edge data acquisition and preprocessing module that forms a closed-loop process through sequential communication connections; an edge-end real-time status monitoring module; a cloud-based progressive prognostic assessment module; an intervention intensity matching and adverse reaction early warning module; an intervention effect closed-loop correction and federated learning module; and a doctor-patient shared decision-making and data feedback module. This system is suitable for long-term outpatient prognostic management of patients with various autoimmune diseases such as systemic lupus erythematosus (SLE), rheumatoid arthritis (RA), and ankylosing spondylitis (AS). Through multimodal data fusion, progressive dynamic prognostic assessment, personalized intervention matching, and closed-loop iterative optimization, it achieves intelligent management of the entire lifecycle of autoimmune diseases.

[0025] The multimodal edge data acquisition and preprocessing module is deployed on the patient-side edge computing devices (smartphones, edge gateways). It is configured to simultaneously collect six types of data—clinical diagnosis and treatment, laboratory testing, wearable physiology, patient-reported outcomes, environmental exposure, and medication behavior—through multiple smart devices deployed on the patient side. The module performs edge data noise reduction, time alignment, and preliminary quality control, and calculates the effective contribution coefficient of each modality of data.

[0026] The multimodal edge data acquisition and preprocessing module includes a clinical diagnosis and treatment acquisition unit, a laboratory testing acquisition unit, a wearable physiological acquisition unit, a patient report acquisition unit, an environmental exposure acquisition unit, a medication behavior acquisition unit, and a data acquisition priority dynamic adjustment unit. The specific implementation methods of each unit are as follows: Clinical diagnosis and treatment data collection unit: Through the standardized interface of the hospital's electronic medical record system, it automatically synchronizes the patient's disease activity scores, affected organ grades, past relapse history and medication history for corresponding diseases such as SLEDAI-2K, DAS28, and BASDAI. The data is updated automatically after each outpatient / inpatient visit, and the original data is maintained when there is no diagnosis or treatment.

[0027] Laboratory test data collection unit: Connects to the hospital's laboratory information system (LIS) to automatically import patients' autoantibody titers, complement C3 / C4, erythrocyte sedimentation rate, C-reactive protein, liver and kidney function, and drug blood concentration data; for patients in the active phase of the disease, the collection frequency is set to once every 3-7 days, and for patients in the remission phase of the disease, the collection frequency is set to once every 1-3 months.

[0028] Wearable physiological data acquisition unit: It establishes a Bluetooth / wireless communication connection with wearable devices such as smart bracelets and ECG patches worn by patients to continuously collect data on patients' heart rate, heart rate variability, sleep stages, daily activity levels, and skin temperature; the regular sampling interval is no more than 1 minute, and the sampling interval during nighttime sleep is increased to 10 seconds to achieve continuous monitoring of physiological indicators.

[0029] Patient Report Collection Unit: Through a mobile application on the patient's end, subjective symptom data such as fatigue score, VAS score for joint pain, rash area, number of oral ulcers, and gastrointestinal reactions are collected daily. The results of the patient's anxiety and depression self-rating scale (PHQ-9, GAD-7) are collected weekly to build a patient subjective outcome database.

[0030] Environmental exposure collection unit: Connects to the public meteorological platform and combines patient terminal location information to automatically collect daily ultraviolet radiation duration, PM2.5 concentration, ambient temperature, relative humidity and pollen index, with a sampling interval of 1 hour, to achieve real-time tracking of environmental factors.

[0031] Medication behavior data collection unit: Through the linkage between the smart pillbox and the patient's mobile application, it records the time and dosage of each medication, automatically identifies missed doses, late doses, and incorrect doses, and also supports patients to manually enter temporary medication adjustment information to accurately calculate patient medication adherence.

[0032] The data acquisition priority dynamic adjustment unit is used to automatically adjust the acquisition priority of each modality of data based on the patient's current disease activity. Specifically, it executes the following steps: S1: Obtain the patient's current disease activity correction value A, and classify it into three levels: low activity (A<0.3), medium activity (0.3≤A<0.6), and high activity (A≥0.6); S2: For patients with low activity levels, the sampling interval for wearable physiological data will be adjusted to 5 minutes, and the frequency of laboratory test collection will be adjusted to once every 3 months; S3: For patients with moderate activity, the sampling interval for wearable physiological data will be adjusted to 1 minute, and the frequency of laboratory test collection will be adjusted to once a month; S4: For patients with high activity levels, the sampling interval for wearable physiological data will be increased to 10 seconds, the frequency of laboratory test collection will be adjusted to once every 3 days, and a daily symptom-specific collection will be added. S5: When the real-time symptom risk index Q>0.7 is detected, the collection frequency of all modal data will be temporarily increased to the highest level, and will be restored to the default frequency of the corresponding activity level after 24 hours.

[0033] The multimodal edge data acquisition and preprocessing module performs the following preprocessing operations on the six types of acquired data: Step X1: Time alignment processing, aligning all modal data to a 1-day time step, calculating the daily mean, maximum value and coefficient of variation for high-frequency wearable data, and using cubic spline interpolation to complete the time series for low-frequency test data, thus solving the problem of time dimension mismatch in multimodal data; Step X2: Outlier identification. Establish an indicator fluctuation range model related to the patient's disease activity. When an indicator exceeds the preset range but is consistent with the trend of disease activity change, it is determined to be a valid outlier. Otherwise, it is determined to be noise and replaced by the median of the three valid values ​​before and after, and invalid data is removed. Step X3: Standardization processing: Minimum-maximum scaling is used to map clinical scores and test indicators to the [0,1] interval, Z-score standardization is used for wearable physiological data, and logarithmic standardization is used for environmental data to eliminate the influence of different dimensional data on subsequent model evaluation. Step X4: Calculate the effective contribution coefficient of each modal data. The formula is: ; in, For the first The effective contribution of modal data to prognostic assessment, with a value ranging from 0 to 1. The larger the value, the higher the quality of the modal data and the greater its contribution to the assessment results. For the first The amount of effective data for a modality within the statistical period; For the first The total amount of data that should be collected for a modality within the statistical period; For the first The number of days since the last valid modal data collection; This is the time decay coefficient, ranging from 0.15 to 0.25, used to control the impact of data timeliness on contribution. This is the decay rate coefficient, with a value of 0.08-0.12, used to adjust the rate of time decay; For the first The basic weights of the modalities are predefined by rheumatology and immunology clinical experts to meet the following requirements. Clinical diagnosis and treatment data Laboratory test data Wearable physiological data Patient Reported Data Environmental exposure data Medication behavior data When any mode At that time, a reminder to supplement the data collection is sent to the patient's terminal.

[0034] The multimodal edge data acquisition and preprocessing module also includes a data quality grading and storage optimization unit, which is used to grade and store data according to the effective contribution coefficient, specifically by performing the following steps: T1: Effective contribution coefficient The data is marked as high-quality data and permanently stored on cloud servers; T2: Effective contribution coefficient The data is labeled as medium quality data and the storage period is set to 5 years. T3: Effective contribution coefficient The data is marked as low-quality data, and the storage period is set to 1 year, after which it will be automatically deleted. T4: When the effective contribution coefficient of low-quality data increases to above 0.65 after supplementary collection, it will be automatically upgraded to medium-quality data and the storage period will be adjusted.

[0035] Through the coordinated configuration of clinical diagnosis and treatment data collection units, laboratory test data collection units, wearable physiological data collection units, patient report data collection units, environmental exposure data collection units, and medication behavior data collection units, the system achieves automated, continuous, and synchronous collection of data across six dimensions: clinical diagnosis and treatment, laboratory tests, wearable physiological data, patient reported outcomes, environmental exposure, and medication behavior. This breaks down the information silos caused by the isolation of data from different sources in traditional prognostic management, providing a comprehensive and three-dimensional reflection of the patient's true condition. Furthermore, standardized interfaces enable seamless integration with hospital electronic medical record systems, LIS systems, and various wearable devices, ensuring the standardization and completeness of data collection. The dynamic data collection priority adjustment unit automatically classifies the patient's current disease activity level and adjusts the collection frequency and strategy for each modality accordingly, balancing the monitoring accuracy during high disease activity with device power consumption and patient burden during remission. This achieves refined dynamic data collection management and allows for temporary escalation of collection priorities in case of sudden abnormalities in the patient's condition. The system ensures comprehensive data capture even in abnormal conditions. By performing time alignment, outlier identification, and standardization preprocessing at the edge, it effectively addresses issues such as time dimension mismatch, noise interference, and inconsistent indicator dimensions in multimodal data. Simultaneously, it decentralizes initial data quality control to the edge, significantly reducing the computational and data transmission burden on cloud servers. Through the quantitative calculation of effective contribution coefficients, it enables dynamic evaluation of the data quality and prognostic contribution of each modality. This not only ensures data integrity through supplementary collection reminders but also provides data quality weighting support for subsequent cloud-based prognostic assessments, effectively improving the accuracy of prognostic results. The accompanying data quality grading and storage optimization unit classifies and differentiates data based on effective contribution coefficients. While ensuring the permanent retention of core high-quality clinical data, it sets corresponding storage periods for medium- and low-quality data, significantly reducing redundant cloud storage resource usage and improving the management efficiency and ease of access to patient prognostic data throughout the entire lifecycle.

[0036] The edge-end real-time status monitoring module is deployed on the patient's edge computing device. Based on preprocessed high-frequency data, it calculates the patient's real-time symptom risk index, identifies sudden abnormal states and triggers local warnings, and simultaneously encrypts and uploads standardized data to the cloud.

[0037] The edge real-time status monitoring module performs the following operations: Step Y1: Extract patient symptom data and physiological data after pretreatment on the same day, and calculate the real-time symptom risk index. The formula is: ; in, This is the patient's real-time symptom risk index for the day, ranging from 0 to 1. A higher value indicates a higher risk of sudden deterioration of the condition. For the first The severity of symptoms is scored on a scale of 0 to 3 (0 points for no symptoms, 1 point for mild, 2 points for moderate, and 3 points for severe). Risk weights are assigned to corresponding symptoms, including joint pain. ,fever ,rash fatigue Gastrointestinal reactions ; This represents the average heart rate for the day. The baseline heart rate during the patient's remission period is obtained by the system automatically calculating the average heart rate during the patient's lowest activity level over the past 3 months; Step Y2: When When this occurs, a local audio-visual alarm is triggered on the patient's terminal, and all standardized data for the day is immediately encrypted and uploaded to the cloud. Step Y3: When At a fixed time every morning, the standardized data for the day is encrypted using the national cryptographic SM4 algorithm and uploaded to the cloud.

[0038] The edge-end real-time status monitoring module also includes a local anomaly data tracing and verification unit, used to trace and verify the anomaly data that triggers the warning, thus avoiding false warnings. The specific steps are as follows: U1: Extract the abnormal data fragments that trigger the warning, check the operating status of the corresponding acquisition device and the data transmission link, and troubleshoot device failures and transmission errors; U2: Compare the changing trends of other modal data of the patient during the same period to determine whether the abnormal data is correlated with other modal data and verify the consistency of the data abnormalities; U3: Sends an abnormal data confirmation request to the patient's terminal, allowing the patient to manually confirm whether there are corresponding symptoms or equipment malfunctions; U4: If the problem is confirmed to be a device malfunction or data transmission error, cancel the alert and mark the data as invalid; if the problem is confirmed to be a real change in the patient's condition, retain the alert and immediately upload all data to the cloud.

[0039] The edge-end real-time status monitoring module is deployed on the patient's edge computing device. Based on preprocessed multimodal high-frequency data, it completes real-time monitoring of the patient's condition and secure data transmission. By extracting and integrating the patient's daily symptom data and physiological data, it quantifies the real-time symptom risk index. Combining the risk weights of different symptoms with changes in the patient's heart rate, it accurately quantifies the risk of a sudden deterioration in the patient's condition that day, providing an objective and quantifiable basis for early warning. Based on the calculation results of the real-time symptom risk index, a differentiated early warning triggering and data upload mechanism is set up. When the risk index exceeds the threshold, a local audio-visual warning is immediately triggered on the patient's terminal, allowing the patient to be aware of abnormal changes in their condition as soon as possible. At the same time, the full set of standardized data for the day is uploaded urgently, ensuring that the cloud can obtain complete information about the patient's condition in a timely manner. When the risk index is within a safe range, the standardized data for the day is encrypted using the national cryptographic SM4 algorithm at a fixed time every morning before being uploaded. This reduces the device power consumption and cloud computing pressure caused by high-frequency data transmission. Encryption algorithms are used to fully ensure the security of the transmission of sensitive patient medical data. The accompanying local abnormal data tracing and verification unit can complete the full-process tracing and verification of abnormal data that triggers warnings. It verifies the authenticity of abnormal data from three dimensions: the operating status of the corresponding collection device and the transmission link, the correlation of multimodal data change trends, and manual confirmation by the patient. This effectively avoids false warnings caused by factors such as equipment failure and data transmission errors, significantly improving the accuracy and reliability of disease warnings, while also ensuring the validity of data uploaded to the cloud. This module pushes real-time disease monitoring, local rapid warning, data quality control verification, and secure transmission to the edge, enabling rapid response and pre-emptive treatment of abnormal patient conditions. At the same time, it forms efficient collaboration with the cloud-based progressive prognostic assessment module, providing timely, accurate, and secure data support for subsequent cloud-based full-dimensional prognostic assessment and personalized intervention plan generation, comprehensively improving the real-time performance, security, and reliability of outpatient prognostic management for patients with autoimmune diseases.

[0040] The cloud-based progressive prognostic assessment module is deployed on a cloud server and constructs a four-level assessment model that integrates multi-scale temporal attention mechanisms. It sequentially outputs the current disease activity correction value, the probability of recurrence in the next 4 weeks, the cumulative target organ damage risk index, and the acute complication risk index.

[0041] The cloud-based progressive prognostic assessment module includes a multi-scale temporal feature encoding unit, a cross-scale attention fusion unit, a four-level risk assessment unit, a dynamic risk grading and source tracing unit, a cross-patient prognostic similarity matching unit, and a risk warning grading response scheduling unit. The specific implementation methods of each unit are as follows: Multi-scale temporal feature encoding unit: A three-layer bidirectional long short-term memory network (BiLSTM) is used to encode temporal features of preprocessed multimodal data from the past 7 days, 30 days, and 90 days, respectively, and the corresponding output hidden state sequences are generated. , , They were used to capture the short-term fluctuations, medium-term changes, and long-term patterns of the patient's condition.

[0042] Cross-scale attention fusion unit: Calculates attention weights at different time scales to achieve adaptive fusion of multi-scale features. Specifically, it performs the following steps: Step Z1: Calculate attention scores at each time scale ,in , For learnable weight matrix, For bias terms; Step Z2: Perform Softmax normalization on the attention scores to obtain cross-scale attention weights. ; Step Z3: Fuse to obtain the comprehensive temporal feature vector .

[0043] Level 4 Risk Assessment Unit: Based on the fused comprehensive time-series feature vector, the following four associated prognostic assessment indicators are calculated sequentially: (1) Current disease activity correction value The formula is: ; in, This represents the corrected current disease activity level for the patient, ranging from 0 to 1. A higher value indicates higher disease activity. For the first The mapping function of modal features to disease activity was obtained by training a random forest model containing clinical data from more than 10,000 patients with autoimmune diseases. (2) Risk of recurrence in the next 4 weeks The formula is: ; in, This represents the probability of disease recurrence within the next 4 weeks, with a value ranging from 0 to 1. This is the Sigmoid activation function, used to map the model output to the 0-1 range; , These are the learnable parameters for the recurrence risk prediction model; To ensure patient adherence to medication over a 30-day rolling period; The environmental risk correction coefficient is calculated by weighting the duration of ultraviolet radiation, PM2.5 concentration, and pollen index. The value ranges from 0.5 to 1.5. The larger the value, the stronger the promoting effect of environmental factors on recurrence. (3) Target organ cumulative injury risk index The formula is: ; in, This is the cumulative risk index for damage to various target organs of the patient, with a value ranging from 0 to 1. The higher the value, the higher the risk of damage to the target organ. The patient's disease duration (in years); For the patient in the course of the disease The average recurrence risk probability per year; For the first Damage weighting coefficients for each affected organ, including the kidneys. ,heart ,lung ,joint ,skin ; (4) Risk index of acute complications The formula is: ; in, It is a risk index for patients to develop acute complications (such as lupus crisis, severe infection, and severe joint injury), with a value range of 0-1, which combines the risk of long-term relapse and the risk of short-term symptoms.

[0044] The dynamic risk classification and traceability unit specifically implements the following steps: Step AA1: Based on the patient's individual historical data and data from the same disease population, a Bayesian optimization algorithm is used to determine three risk thresholds. , , ; Step AA2: Classification of recurrence risk levels: Low risk (Level 1) It is classified as low to medium risk (Level 2). It is classified as medium to high risk (Level 3). High risk (Level 4); Step AA3: Calculate the cumulative recurrence risk value over the past 14 days. The formula is: ; in, The cumulative relapse risk value over the past 14 days is used to assess the overall trend of the recent disease. For the first The probability of recurrence within 1 day; This is the recent risk weighting coefficient, ranging from 0.1 to 0.2, used to assign higher weight to recent data; when Furthermore, if the current risk level is level 2, it will automatically be upgraded to a level 3 warning. Step AA4: Risk factor tracing. The contribution of each feature to the recurrence risk is calculated using multi-scale SHAP values. The top 3 features with the highest absolute values ​​are selected as the main risk factors, and a risk tracing report is generated. Step AA5: When the risk level rises to level 3 or above, or In such cases, an early warning message and a risk tracing report will be immediately sent to the attending physician's terminal.

[0045] The cross-patient prognostic similarity matching unit is used to match historical patients with similar prognostic characteristics to the current patient, assisting in prognostic assessment. Specifically, it performs the following steps: V1: Extract the comprehensive temporal feature vector of the current patient. Information on disease type, course of disease, and affected organs; V2: In the cloud-based historical patient database, cosine similarity is used to calculate the feature similarity with the current patient; V3: Select the top 10 historical patients with the highest similarity and extract their actual prognostic results and the effectiveness of the intervention plan; V4: The prognostic statistics of similar patients are used as a clinical reference and supplemented in the prognostic assessment report of the physician.

[0046] The risk warning and graded response scheduling unit is used to automatically schedule corresponding medical resources according to different warning levels, and specifically executes the following steps: W1: Level 1 Warning (Low Risk): Health reminders are sent only to patients, without notifying medical staff; W2: Level 2 Warning (Low to Medium Risk): Push intervention suggestions to the patient's terminal and send a reminder to the responsible nurse's terminal at the same time; W3: Level 3 Warning (Medium-high risk): Push emergency intervention suggestions to the patient's terminal, and send warning information to the terminal of the attending physician and the responsible nurse, requiring the patient to complete the follow-up within 24 hours; W4: Level 4 Warning (High Risk): Immediately activate the emergency green channel, send warning information to the hospital emergency department, attending physician and patient's family at the same time, and push pre-hospitalization notice.

[0047] A four-level assessment model integrating multi-scale temporal attention mechanisms was constructed. A three-layer bidirectional long short-term memory network was used in the multi-scale temporal feature encoding unit to encode temporal features from patients' multimodal data over the past 7, 30, and 90 days. This simultaneously captures short-term fluctuations, medium-term changes, and long-term disease progression patterns, solving the problem that traditional prognostic models cannot simultaneously consider disease characteristics across different time dimensions, and providing a comprehensive temporal feature foundation for subsequent prognostic assessment. An adaptive fusion of features from different time scales was achieved through a cross-scale attention fusion unit. Based on the calculation and normalization of attention scores, feature weights for each time scale were dynamically allocated, automatically strengthening time-dimensional features with greater impact on patient prognosis, effectively improving the expression of temporal features. The accuracy and sensitivity of prognostic assessment are enhanced. Based on the fused comprehensive temporal feature vector, a four-level risk assessment unit sequentially completes the progressive quantitative calculation of the current disease activity correction value, the probability of recurrence risk in the next 4 weeks, the target organ cumulative damage risk index, and the acute complication risk index. This achieves a comprehensive and multi-level prognostic analysis from judging the patient's current condition, predicting short-term recurrence risk, assessing long-term target organ damage, to warning of acute complication risk. This provides a comprehensive, objective, and quantifiable core decision-making basis for the formulation of subsequent personalized intervention plans. The accompanying dynamic risk grading and traceability unit determines the risk threshold based on individual patient data and data from the same disease population through a Bayesian optimization algorithm, completing the accurate classification of recurrence risk levels. The system dynamically verifies risk levels by calculating cumulative relapse risk values, avoiding the limitations of static risk grading. It then uses multi-scale SHAP values ​​to pinpoint core risk factors leading to disease deterioration and generates source tracing reports. For high-risk situations, it can promptly push early warning information to healthcare terminals, achieving precise risk stratification and clarifying the core driving factors of disease changes, significantly improving the targeting and accuracy of prognostic warnings. Through cross-patient prognostic similarity matching units, based on cosine similarity matching with historical patients whose prognostic characteristics are similar to the current patient, it extracts their actual prognostic outcomes and intervention effects as clinical references, providing doctors with real clinical data support for developing intervention plans and further improving the rationality of clinical decisions. Through risk warnings… The tiered response scheduling unit sets up a differentiated medical resource scheduling mechanism for different risk levels, achieving precise matching between patient risk levels and medical resources. This avoids the waste of medical resources in low-risk situations and ensures that high-risk patients can receive timely and efficient medical intervention, effectively reducing the risk of acute crises in autoimmune diseases. This module as a whole realizes intelligent, precise, and multi-dimensional dynamic assessment of the prognosis of patients with autoimmune diseases. It forms efficient collaboration with the edge real-time status monitoring module, providing core decision support for matching the intensity of subsequent interventions and generating personalized intervention plans. At the same time, through the tiered early warning and resource scheduling mechanism, it comprehensively improves the risk control capabilities and clinical intervention effects of outpatient prognosis management for patients with autoimmune diseases.

[0048] The intervention intensity matching and adverse reaction early warning module is deployed on a cloud server. It calculates the overall intervention intensity by comprehensively considering the risk of recurrence, target organ damage and complication, and generates a tiered intervention plan. At the same time, it assesses the risk of adverse drug reactions based on the patient's genotype, liver and kidney function and medication history.

[0049] The intervention intensity matching and adverse reaction early warning module includes an intervention intensity calculation unit, a medication adjustment recommendation unit, an adverse reaction risk assessment unit, a lifestyle and follow-up plan unit, and a personalized intervention plan adaptation unit. The specific implementation methods of each unit are as follows: Intervention intensity calculation unit: Calculates the overall intervention intensity The formula is: ; in, The overall intervention intensity, ranging from 0 to 1, considers the risks of recurrence, target organ damage, and acute complications; based on... The value of is used to divide the intervention intensity into four levels: Basic level, For the enhanced level, It is an enhanced level. It is classified as an emergency.

[0050] Medication Adjustment Recommendation Unit: Generates step-by-step medication treatment recommendations based on the overall intervention intensity. Basic level: Maintain the patient's current medication dosage and have a follow-up examination every 3 months; Enhanced level: Increase the immunosuppressant dose by 10%-15%, or shorten the dosing interval of biologics by 10%; Enhanced level: Add low-dose glucocorticoids (prednisone 10-20mg / day), and adjust the type of immunosuppressant; Emergency Level: Immediately activate the emergency green channel, send a pre-hospitalization notice, and recommend high-dose hormone pulse therapy.

[0051] Adverse Reaction Risk Assessment Unit: Calculates the adverse drug reaction risk index The formula is: ; in, The adverse reaction risk index for patients using the current medication regimen, with a value ranging from 0 to 1; For the first The basic adverse reaction risk of this drug is obtained from the drug instructions and large-sample clinical data statistics; The individual risk correction factor is calculated based on the patient's liver and kidney function, genotype, age, and concomitant medications; when If the drug is deemed high-risk, a clinically equivalent low-risk alternative is recommended to the patient.

[0052] Lifestyle and Follow-up Planning Unit: Based on the patient's overall intervention intensity, personalized lifestyle and follow-up planning recommendations are generated. Exercise duration: ,in The baseline exercise duration is set by the doctor based on the patient's physical condition. Follow-up examination interval: (Unit: days, minimum 7 days); The follow-up examination items are automatically matched according to the patient's affected organs. For example, if the kidneys are involved, a 24-hour urine protein quantification test will be added, and if the lungs are involved, lung function tests and chest CT scans will be added.

[0053] Personalized intervention program adaptation unit: Used to adjust the intervention program according to the patient's lifestyle and preferences, specifically by implementing the following steps: X1: Collect patient information on their daily routine, dietary habits, exercise preferences, and work schedule through a mobile application. X2: Adjust the medication reminder time according to the patient's daily routine to avoid conflicts with the patient's work or rest time; X3: Recommend corresponding exercise methods based on the patient's exercise preferences. For example, if the patient prefers swimming, recommend swimming instead of brisk walking; if the patient prefers yoga, recommend low-intensity yoga instead of regular rehabilitation exercises. X4: Adjust dietary recommendations based on the patient's eating habits, avoid recommending foods that the patient dislikes or cannot obtain, and develop personalized recipes in combination with the patient's dietary restrictions; X5: Generates the final personalized intervention plan and pushes it to both the doctor and the patient, who then confirms and implements it.

[0054] The intervention intensity calculation unit quantifies the overall intervention intensity based on the patient's recurrence risk, target organ damage risk, and acute complication risk, and completes a four-level classification, providing a precise quantitative basis for the development of tiered intervention plans. The medication adjustment recommendation unit matches the corresponding intervention intensity to generate tiered drug treatment suggestions, achieving precise matching of intervention intensity with the patient's disease risk and effectively avoiding clinical problems of insufficient or excessive intervention. The adverse reaction risk assessment unit quantifies the adverse drug reaction risk index, accurately identifying high-risk drugs and recommending equivalent alternatives, significantly improving the safety of long-term medication for autoimmune diseases. The lifestyle and follow-up plan unit generates personalized treatment plans based on the intervention intensity. The exercise program and follow-up plan, combined with the automatic matching of follow-up examination items for the patient's affected organs, achieve synergistic cooperation between clinical treatment, lifestyle intervention, and follow-up monitoring. The personalized adaptation unit for intervention plans optimizes the details of the intervention plan based on the patient's lifestyle habits and preferences, effectively improving patient acceptance and adherence. This module as a whole realizes intelligent decision-making throughout the entire process, from quantifying intervention intensity, generating personalized plans, providing medication safety warnings, to optimizing plan adaptation. It forms efficient collaboration with the cloud-based progressive prognostic assessment module, providing patients with accurate, safe, and implementable personalized intervention plans, significantly improving the clinical effectiveness and patient management efficiency of outpatient prognostic interventions for autoimmune diseases.

[0055] The intervention effect closed-loop correction and federated learning module is deployed on cloud servers and edge nodes of various cooperating medical institutions to track changes in the condition after the intervention is implemented, calculate the intervention effect score and reverse correct the model parameters and the matching rules of the protocol library. The incremental federated learning framework is used to achieve multi-center model collaborative optimization.

[0056] The intervention intensity matching and adverse reaction early warning module performs the following operations: Step BB1: Collect multimodal data from patients on days 7, 14, and 28 after the intervention program is implemented, and calculate the intervention effect score. The formula is: ; In the formula, The effectiveness of the intervention program is scored, with a value ranging from -1 to 1. A larger value indicates a better intervention effect. , , The disease activity correction value, relapse risk probability, and real-time symptom risk index before intervention; , , The corresponding value after intervention; Step BB2: Grading of Intervention Effects: For significant effectiveness, For it to be effective, Invalid; Step BB3: Model Correction: When When, keep the model parameters and the matching weights of the solution library unchanged; when When, proportionally increase the matching weight of the corresponding intervention plan in the plan library; when At that time, incremental training of the model is triggered; Step BB4: Incremental federated learning update to achieve collaborative optimization of multi-center models: Each collaborating medical institution's edge nodes train local models based on the latest local patient data and calculate incremental gradients. ; The incremental gradient is encrypted using a homomorphic encryption algorithm and then uploaded to the cloud central server. The cloud-based central server performs federated average aggregation to generate a global incremental gradient. ,in The number of edge nodes participating in federated learning; The global model is updated in the cloud and encrypted and distributed to each edge node, completing one iteration of the global model.

[0057] The intervention intensity matching and adverse reaction early warning module also includes a model incremental update triggering and verification unit to control the frequency and quality of model incremental updates and avoid model drift. Specifically, it executes the following steps: Y1: Set the minimum trigger interval for incremental model updates to 1 month to avoid frequent updates that could lead to model instability; Y2: When incremental training of the model is triggered, local patient data from the most recent 3 months for the corresponding edge node is used for training; Y3: After training is complete, use an independent validation set to calculate the prediction accuracy of the updated model; Y4: If the accuracy of the updated model improves by ≥2%, the new model will be adopted; otherwise, the original model will be retained and the model will be waited for the next update.

[0058] The system also includes a multimodal missing robustness module, deployed on a cloud server, to address the stability issue of prognostic assessment when some modal data is missing. Specifically, it performs the following operations: Step CC1: When a certain mode When data is missing, renormalize the effective contribution coefficients of other available modes: ; in, For the renormalized first The effective modal contribution coefficient ensures that the sum of the contribution coefficients of all available modes is 1. Step CC2: Based on a pre-trained Generative Adversarial Network (GAN), generate virtual features for missing modalities using available modality data. ; Step CC3: Calculate the robust fusion feature vector The formula is: ; in, This is the robust fusion feature vector when modality is missing, with the same dimension as the original fusion feature vector; The feature vectors of the available modes; This is the virtual feature weight coefficient, with a value of 0.2-0.4, used to control the influence of virtual features on the fusion result; Step CC4: Use The original fusion features are used for prognostic assessment, and a data missing alert is sent to the patient's terminal.

[0059] The intervention effect closed-loop correction and federated learning module also includes a missing modality data completion quality evaluation unit, which is used to evaluate the completion quality of virtual features and dynamically adjust the weights of virtual features. The specific steps are as follows: Z1: After the missing modal data is completed, calculate the mean square error (MSE) between the virtual features and the historical true features of the modality; Z2: If MSE < 0.1, then the virtual feature weights will be adjusted. Adjusted to 0.4; Z3: If 0.1 ≤ MSE < 0.3, then the virtual feature weights will be adjusted. Adjusted to 0.3; Z4: If MSE ≥ 0.3, then the virtual feature weights will be adjusted. Adjust it to 0.2 and mark the completed data as low confidence.

[0060] By quantitatively calculating and grading intervention effect scores, the clinical efficacy of intervention programs can be accurately assessed. Based on the effect grading, model parameters and program library matching weights are adjusted in reverse, achieving closed-loop iterative optimization of prognostic assessment and intervention programs. An incremental federated learning framework, combined with homomorphic encryption technology, is used to achieve collaborative optimization of multi-center models. While strictly protecting the privacy of patients' original medical data, this effectively improves the generalization ability and prediction accuracy of the global model. The accompanying model incremental update triggering and verification unit effectively controls the frequency and quality of model updates by setting minimum update intervals, limiting the training data range, and verifying the improvement in model accuracy, avoiding model drift and ensuring the stability and reliability of the prognostic model. Simultaneously, the system's multimodal missing robustness module ensures the normal operation of prognostic assessment when some modal data is missing. This is achieved through effective contribution coefficient renormalization, generative adversarial network virtual feature generation, and robust fusion feature calculation, resolving the assessment interruption problem caused by incomplete outpatient data collection. Furthermore, the missing modal data completion quality assessment unit dynamically adjusts virtual feature weights and marks data credibility, further improving the system's stability, robustness, and the accuracy of assessment results.

[0061] The doctor-patient shared decision-making and data feedback module is deployed on a cloud server. It enables data interaction through doctor and patient terminals, synchronously displays prognostic assessment results and intervention plans to both doctors and patients, collects patient feedback data, and sends it back to the multimodal edge data acquisition and preprocessing module.

[0062] The shared decision-making and data feedback module for doctors and patients includes a visualization unit, a prognosis simulation unit, a decision recording unit, a data feedback unit, and a patient decision preference learning unit. The specific implementation methods of each unit are as follows: Visualization Unit: Simultaneously displays the patient's disease activity curve, recurrence risk trend chart, target organ damage radar chart, and adverse reaction risk warnings to both doctors and patients. The visualization is presented in an easy-to-understand way for patients and provides professional clinical data details for doctors.

[0063] Prognostic simulation unit: Supports doctors and patients in simulating prognostic changes under different intervention programs and calculating the simulated risk of recurrence. ,in The efficacy coefficient of the intervention protocol is derived from historical clinical data in the intervention protocol database. To simulate the intensity of intervention and provide a quantitative reference for joint decision-making by doctors and patients.

[0064] Decision recording unit: Automatically records the discussion process between doctors and patients, the selection of solutions, and the patient's informed consent, generates archived files that comply with electronic medical record standards, and synchronizes them to the hospital's electronic medical record system.

[0065] Data feedback unit: After encrypting the patient's symptom feedback, medication implementation and intervention effect data using the national cryptographic SM4 algorithm, the data is sent back to the multimodal edge data acquisition and preprocessing module to trigger the next round of prognostic assessment, forming a closed loop for the whole process management.

[0066] Patient Decision Preference Learning Unit: Used to learn patients' decision preferences and optimize subsequent intervention recommendations. Specifically, it executes the following steps: AA1: Record the patient's acceptance or rejection of each recommended intervention, and the reasons for rejection indicated by the patient; AA2: Extract the core characteristics of patients who refuse treatment plans, such as high risk of drug side effects, high intensity of exercise, high frequency of follow-up examinations, and high treatment costs; AA3: When recommending interventions in the future, reduce the matching weight of protocols that include patient rejection characteristics; AA4: Update the patient's decision preference model every 3 months to ensure that the recommended plan better meets the patient's actual needs. Specific implementation examples: This embodiment uses the outpatient prognostic management of a 38-year-old female patient with systemic lupus erythematosus (SLE) as an application scenario. It employs the autoimmune disease patient prognostic management system described in this invention to complete intelligent prognostic management and intervention throughout the patient's entire lifecycle. The specific implementation process is as follows: Patient baseline clinical data: The patient was diagnosed with systemic lupus erythematosus 5 years ago, with a history of kidney and skin involvement, a baseline SLEDAI-2K score of 4, and was in remission; the patient had a history of hypertension, no history of drug allergies, and normal liver and kidney function; the baseline medication regimen was hydroxychloroquine 0.2g bid, mycophenolate mofetil 0.5g bid, and prednisone 5mg qd; the patient wore a smart bracelet, was equipped with a smart pillbox, and used a patient-side mobile application to complete data reporting.

[0068] System deployment boundary parameters:

[0069] Basic data acquisition configuration: The system configures six basic data collection rules for patients: Clinical diagnosis and treatment data: Automatically synchronized after each outpatient visit; baseline data synchronized monthly when there is no diagnosis or treatment. Laboratory test data: During the remission period, the sampling frequency was once every 2 months, and during the active period, it was adjusted to once every 7 days; Wearable physiological data: The regular sampling interval is 1 minute, and the interval is increased to 10 seconds during nighttime sleep. Patient report data: Subjective symptom scores were collected daily, and the Anxiety and Depression Self-Rating Scale was collected weekly; Environmental exposure data: Meteorological and environmental data of the patient's location were collected every hour; Medication usage data: The smart pillbox records medication usage in real time and synchronizes it to the system daily.

[0070] Multimodal data acquisition and preprocessing implementation: After the system starts up, the multimodal edge data acquisition and preprocessing module completes the acquisition and preprocessing of patient data. The specific implementation process is as follows: Data Acquisition: The system continuously collects multimodal data from patients for 14 days. Clinical diagnosis and treatment data are synchronized with the patient's most recent outpatient SLEDAI-2K score, affected organ classification, and medication history. Laboratory test data are synchronized with the patient's most recent complement C3 / C4, erythrocyte sedimentation rate, C-reactive protein, and liver and kidney function results. Wearable devices continuously collect data on the patient's heart rate, sleep, and activity level. Patients report symptoms daily. Environmental data are synchronized with the daily ultraviolet radiation and PM2.5 concentration data of the patient's location. The smart pillbox records the patient's daily medication usage.

[0071] Preprocessing operations: Time alignment processing: All modal data are aligned to a time step of 1 day. The daily mean, maximum value and coefficient of variation are calculated for high-frequency wearable data. Cubic spline interpolation is used to complete the time series for low-frequency test data. Outlier identification: Establish a fluctuation range model of indicators related to the patient's disease activity, remove noisy data that exceeds the range and is unrelated to the trend of disease activity, and replace it with the median of the three valid values ​​before and after. Standardization processing: Clinical scores and test indicators were mapped to the [0,1] interval using min-max scaling; wearable physiological data were standardized using Z-score; and environmental data were standardized using logarithmic scaling. Effective contribution coefficient calculation: The effective contribution coefficient of six types of data is calculated according to a preset formula, including clinical diagnosis and treatment data. Laboratory test data Wearable physiological data Patient Reported Data Environmental exposure data Medication use behavior data All modes All values ​​are ≥0.65, no further data collection is required; Data tiered storage: All data with an effective contribution coefficient ≥0.8 are marked as high-quality data and permanently stored on cloud servers.

[0072] Implementation of real-time status monitoring at the edge: The edge-end real-time status monitoring module calculates the patient's real-time symptom risk index Q daily. The specific implementation process is as follows: For the first 13 days, the patient's daily symptom score was 0-1, the average heart rate was 72 beats / min, the baseline heart rate was 70 beats / min, the daily Q value was ≤0.3, there was no warning trigger, and the standardized data of the day was encrypted and uploaded to the cloud every morning. On day 14, the patient reported a VAS score of 3 for joint pain and 2 for fatigue, with an average heart rate of 90 beats per minute. The following calculations were performed using the formula: ; After normalization ; The system immediately triggers a local audio-visual alert on the patient's terminal and simultaneously initiates local abnormal data tracing and verification: the device is checked to ensure it is operating normally, and the laboratory test data from the same period show that the patient's complement C3 level has decreased compared to the previous period, which is consistent with the trend of symptom changes. The patient manually confirms the presence of the corresponding symptoms, confirming the effectiveness of the alert, and immediately uploads all data for the day to the cloud.

[0073] Implementation of cloud-based progressive prognostic assessment: After receiving the data urgently uploaded by the patient, the cloud server completes a comprehensive prognostic assessment through the cloud-based progressive prognostic assessment module. The specific implementation process is as follows: Multi-scale temporal feature encoding: A three-layer BiLSTM network is used to encode the patient's multimodal data over the past 7 days, 30 days, and 90 days, respectively, and output the corresponding hidden state sequences; Cross-scale attention fusion: The attention weights for the 7-day, 30-day, and 90-day timescales were calculated to be 0.6, 0.3, and 0.1, respectively, and fused to obtain a comprehensive temporal feature vector. ; Level 4 Risk Assessment: Four core indicators are calculated sequentially: Current disease activity correction value It was determined to be a high level of activity; Risk of recurrence in the next 4 weeks ; Target organ cumulative damage risk index ; Risk index for acute complications: ; Dynamic risk stratification and source tracing: patients The risk level was determined to be Level 3 (medium-high); the cumulative recurrence risk value over the past 14 days was calculated. Further confirm the risk level; using SHAP value calculation, the core risk factors are found to be decreased complement C3, worsening of joint pain symptoms, and increased duration of ultraviolet exposure, and a risk tracing report is generated; Early warning response dispatch: Immediately push Level 3 early warning information and risk tracing report to the patient's attending physician and responsible nurse's terminals, requiring follow-up to be completed within 24 hours; at the same time, push emergency intervention suggestions to the patient's terminal; Similar patient matching: The top 10 SLE patients with the highest similarity were matched in the cloud database. Among them, 8 patients achieved remission after adjusting the immunosuppressant dosage. This result was added to the doctor's assessment report.

[0074] Intervention plan generation and adverse reaction early warning implementation: The intervention intensity matching and adverse reaction early warning module generates personalized intervention plans based on patients' risk indicators. The specific implementation process is as follows: Intervention intensity calculation: Calculated using the formula: ; Determined to be an intensive intervention; Recommended medication adjustments: For enhanced drug adjustment, add prednisone 15mg / day, adjust mycophenolate mofetil dose to 0.75g bid, and maintain the original dose of hydroxychloroquine; Adverse reaction risk assessment: Calculate the adverse drug reaction risk index for this regimen. There are no high-risk drugs, and no treatment adjustments are needed. Lifestyle and Follow-up Plan: Generating Personalized Recommendations: The patient's baseline exercise duration was 30 minutes per day, and the adjusted exercise duration was: ; Minutes / day, low-intensity slow walking is recommended; follow-up examination interval: ; The follow-up examination will be conducted one month later, and the examination will include 24-hour urine protein quantification, complement and erythrocyte sedimentation rate tests. Personalized adaptation of the plan: Taking into account the patient's 9-to-5 work schedule, the medication reminder time is set to 8 am and 8 pm to avoid conflicts with work hours; taking into account the patient's ultraviolet allergy, a daily ultraviolet protection reminder is added; the final personalized intervention plan is generated and pushed to both the doctor and the patient.

[0075] Intervention effect tracking and model revision implementation: After the intervention plan is implemented, the system completes effect tracking and model optimization through the intervention effect closed-loop correction and federated learning module. The specific implementation process is as follows: Intervention effectiveness score calculation: Patient data were collected on days 7, 14, and 28 after the intervention. The intervention effectiveness score was calculated on day 28. (Previous intervention score was not included in the original text.) , , After intervention , , ; Calculated using the formula, we get: ; It was determined to be significantly effective; Model correction: The intervention was significantly effective. Keeping the model parameters unchanged, the matching weight of this intervention in the SLE patient protocol library was increased. Federated Learning Update: In the current month, all three edge nodes participating in federated learning completed local model training, uploaded encrypted incremental gradients, performed federated average aggregation in the cloud, updated the global model, and improved the accuracy of recurrence risk prediction by 2.8%, completing the global model iteration.

[0076] Implementation of Shared Decision-Making and Closed-Loop Management between Doctors and Patients: Visualization: The disease activity curve and relapse risk trend chart before and after the intervention are displayed simultaneously on both the doctor and patient sides, allowing patients to intuitively view the intervention effect; Prognostic simulation: Doctors simulate the prognostic changes of patients after hormone dosage reduction, simulating the intensity of intervention. The simulated recurrence risk was calculated as follows: ; This will provide a reference for subsequent hormone tapering plans; Decision-making record: Automatically records the discussion process of the treatment plan between doctors and patients, the patient's informed consent, generates electronic archive files, and synchronizes them to the hospital's electronic medical record system; Data feedback: Patient symptom feedback, medication adherence, and intervention effectiveness data are encrypted and transmitted back to the data acquisition module, triggering the next round of prognostic assessment and forming a closed loop for the entire management process; Decision preference learning: Record the patient's acceptance of the intervention plan, update the patient decision preference model, and subsequently recommend treatment plans with lower risk of adverse reactions of the same type.

[0077] In this embodiment, the system, through an edge-cloud collaborative architecture, promptly identifies early warning signals of worsening patient condition, performs accurate prognostic risk assessment, generates a matching personalized intervention plan, and after intervention, the patient's condition rapidly improves. The risk of relapse within the next 4 weeks decreases from 0.72 to 0.18, disease activity returns to a low activity level, and no adverse drug reactions occur. Simultaneously, incremental federated learning enables collaborative optimization of the multi-center model, improving the predictive accuracy of the global model and validating the clinical applicability and effectiveness of this invention in the prognostic management of patients with autoimmune diseases.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A prognostic management system for patients with autoimmune diseases, characterized in that, The architecture employs an edge-cloud collaborative approach, including sequential communication connections to form a closed-loop process. The multimodal edge data acquisition and preprocessing module is configured to simultaneously collect six types of data—clinical diagnosis and treatment, laboratory testing, wearable physiology, patient-reported outcomes, environmental exposure, and medication behavior—through various intelligent devices deployed on the patient end. It performs edge data noise reduction, time alignment, and preliminary quality control, and calculates the effective contribution coefficient of each modality of data. The edge-end real-time status monitoring module calculates the patient's real-time symptom risk index based on preprocessed high-frequency data, identifies sudden abnormal states and triggers local warnings, while encrypting and uploading standardized data to the cloud. The cloud-based progressive prognostic assessment module constructs a four-level assessment model that integrates multi-scale temporal attention mechanisms, and sequentially outputs the current disease activity correction value, the probability of recurrence risk in the next 4 weeks, the cumulative target organ damage risk index, and the acute complication risk index. The intervention intensity matching and adverse reaction early warning module calculates the overall intervention intensity by comprehensively considering the risk of recurrence, target organ damage and complication, and generates a step-by-step intervention plan. At the same time, it assesses the risk of adverse drug reactions based on the patient's genotype, liver and kidney function and medication history. The intervention effect closed-loop correction and federated learning module tracks the changes in the condition after the intervention is implemented, calculates the intervention effect score and back-corrects the model parameters and the matching rules of the protocol library. The incremental federated learning framework is used to achieve multi-center model collaborative optimization. The doctor-patient shared decision-making and data feedback module synchronously displays prognostic assessment results and intervention plans to both doctors and patients, collects patient feedback data, and sends it back to the multimodal edge data acquisition and preprocessing module.

2. The prognostic management system for patients with autoimmune diseases according to claim 1, characterized in that, The multimodal edge data acquisition and preprocessing module includes: The clinical diagnosis and treatment data collection unit automatically synchronizes patients’ SLEDAI-2K, DAS28, BASDAI disease activity scores, affected organ grades, past relapse history and medication history through the electronic medical record interface. The data is updated automatically after each outpatient visit. The laboratory test collection unit connects to the hospital's LIS system and automatically imports data on autoantibody titers, complement C3 / C4, erythrocyte sedimentation rate, C-reactive protein, liver and kidney function, and drug blood concentration. During the active phase, the collection frequency is once every 3-7 days, and during the remission phase, it is once every 1-3 months. The wearable physiological data collection unit communicates with smart bracelets and ECG patches to continuously collect data on heart rate, heart rate variability, sleep stages, daily activity levels, and skin temperature. The sampling interval is no more than 1 minute, and during nighttime sleep, the interval is increased to 10 seconds. The patient report collection unit collects fatigue scores, VAS scores for joint pain, rash area, number of oral ulcers, and gastrointestinal reactions daily via a mobile application, and collects the results of the anxiety and depression self-rating scale once a week. The environmental exposure data collection unit connects to the meteorological platform and patient location data, automatically collecting daily ultraviolet radiation duration, PM2.5 concentration, temperature, humidity, and pollen index at 1-hour intervals. The medication behavior collection unit records the time and dosage of each medication through a smart pillbox, automatically identifies missed doses, late doses, and incorrect doses, and supports patients to manually enter temporary medication adjustment information. A dynamic data acquisition priority adjustment unit is used to automatically adjust the acquisition priority of each modality of data based on the patient's current disease activity, including: S1: Obtain the patient's current disease activity correction value A, and classify it into three levels: low activity A<0.3, medium activity 0.3≤A<0.6, and high activity A≥0.6; S2: For patients with low activity levels, the sampling interval for wearable physiological data will be adjusted to 5 minutes, and the frequency of laboratory test collection will be adjusted to once every 3 months; S3: For patients with moderate activity, the sampling interval for wearable physiological data will be adjusted to 1 minute, and the frequency of laboratory test collection will be adjusted to once a month; S4: For patients with high activity levels, the sampling interval for wearable physiological data will be increased to 10 seconds, the frequency of laboratory test collection will be adjusted to once every 3 days, and a daily symptom-specific collection will be added. S5: When the real-time symptom risk index Q>0.7 is detected, the collection frequency of all modal data will be temporarily increased to the highest level, and will be restored to the default frequency of the corresponding activity level after 24 hours.

3. The prognostic management system for patients with autoimmune diseases according to claim 2, characterized in that, The multimodal edge data acquisition and preprocessing module performs the following operations: Step X1: Time alignment processing, align all modal data to a 1-day time step, calculate the daily mean, maximum and coefficient of variation for high-frequency wearable data, and use cubic spline interpolation to complete the time series for low-frequency test data; Step X2: Outlier identification. Establish an indicator fluctuation range model related to disease activity. When an indicator exceeds the range but is consistent with the trend of disease activity change, it is determined to be a valid outlier. Otherwise, it is determined to be noise and replaced by the median of the three valid values ​​before and after. Step X3: Standardization processing: Minimum-maximum scaling is used to map clinical scores and test indicators to the [0,1] interval; Z-score standardization is used for wearable physiological data; and logarithmic standardization is used for environmental data. Step X4: Calculate the effective contribution coefficient of each modal data. The formula is: ； in, For the first The effective contribution of modal data to prognostic assessment, with a value ranging from 0 to 1; For the first The amount of effective data for a modality within the statistical period; For the first The total amount of data that should be collected for a modality within the statistical period; For the first The number of days since the last valid modal data collection; This is the time decay coefficient, with a value ranging from 0.15 to 0.25; This is the decay rate coefficient, with a value ranging from 0.08 to 0.12; For the first The basic weights of the modalities are predefined by clinical experts and satisfy the following conditions: Clinical diagnosis and treatment data Laboratory test data Wearable physiological data Patient Reported Data Environmental exposure data Medication behavior data When any mode At that time, a reminder to supplement the data collection is sent to the patient's terminal; The multimodal edge data acquisition and preprocessing module also includes a data quality grading and storage optimization unit, used to grade and store data according to effective contribution coefficients, including: T1: Effective contribution coefficient The data is marked as high-quality data and permanently stored on cloud servers; T2: Effective contribution coefficient The data is labeled as medium quality data and will be stored for 5 years. T3: Effective contribution coefficient The data is marked as low-quality data and stored for one year, after which it will be automatically deleted. T4: When the effective contribution coefficient of low-quality data increases to above 0.65 after supplementary collection, it will be automatically upgraded to medium-quality data and the storage period will be adjusted.

4. The prognostic management system for patients with autoimmune diseases according to claim 3, characterized in that, The edge-end real-time status monitoring module performs the following operations: Step Y1: Extract the symptom data and physiological data for the day, and calculate the real-time symptom risk index. The formula is: ； in, This is the patient's real-time symptom risk index for the day, with a value ranging from 0 to 1. For the first The severity of each symptom is scored on a scale of 0-3. Risk weights are assigned to corresponding symptoms, including joint pain. ,fever ,rash fatigue Gastrointestinal reactions ; This represents the average heart rate for the day. The patient's baseline heart rate during remission; Step Y2: When When this occurs, a local audible and visual alarm is triggered, and the data for the day is immediately uploaded to the cloud. Step Y3: When At that time, the standardized data for the day is encrypted and uploaded to the cloud every day at midnight; The edge-end real-time status monitoring module also includes a local anomaly data tracing and verification unit, used to trace and verify the anomaly data that triggers warnings, avoiding false warnings, including: U1: Extract the abnormal data segment that triggered the warning and check the operating status of the corresponding acquisition device and the data transmission link; U2: Compare the changing trends of other modal data of the patient during the same period to determine whether the abnormal data is correlated with other modal data; U3: Sends an abnormal data confirmation request to the patient's terminal, allowing the patient to manually confirm whether there are corresponding symptoms or equipment malfunctions; U4: If the problem is confirmed to be a device malfunction or data transmission error, cancel the alert and mark the data as invalid; if the problem is confirmed to be a genuine change in the patient's condition, retain the alert and immediately upload it to the cloud.

5. The prognostic management system for patients with autoimmune diseases according to claim 4, characterized in that, The cloud-based progressive prognostic assessment module includes: The multi-scale temporal feature encoding unit uses a three-layer bidirectional long short-term memory network to encode multimodal data from the past 7 days, 30 days, and 90 days, respectively, and outputs a hidden state sequence. , , ; The cross-scale attention fusion unit calculates attention weights at different time scales, and the steps are as follows: Step Z1: Calculate attention scores at each time scale ,in , For learnable weight matrix, For bias terms; Step Z2: Perform Softmax normalization on the attention scores to obtain cross-scale attention weights. ; Step Z3: Fuse to obtain the comprehensive temporal feature vector ; The Level 4 risk assessment unit calculates the following four related indicators in sequence: (1) Current disease activity correction value The formula is: ； in, This represents the corrected current disease activity level for the patient, ranging from 0 to 1. A higher value indicates higher disease activity. For the first The mapping function of modal features to disease activity was obtained by training a random forest model containing data from more than 10,000 patients. (2) Risk of recurrence in the next 4 weeks The formula is: ； in, This represents the probability of disease recurrence within the next 4 weeks, with a value ranging from 0 to 1. Use the Sigmoid activation function; , These are the learnable parameters for the recurrence risk prediction model; To ensure patient adherence to medication over a 30-day rolling period; This is an environmental risk correction factor, with a value ranging from 0.5 to 1.

5. (3) Target organ cumulative injury risk index The formula is: ； in, This is the cumulative risk index for damage to various target organs of the patient, with a value ranging from 0 to 1. The higher the value, the higher the risk of damage to the target organ. The patient's disease duration (in years); For the patient in the course of the disease The average recurrence risk probability per year; For the first Damage weighting coefficients for each affected organ, including the kidneys. ,heart ,lung ,joint ,skin ; (4) Risk index of acute complications The formula is: ； in, It is a risk index for patients to develop acute complications, with a value ranging from 0 to 1, which combines the risk of long-term recurrence and the risk of short-term symptoms; The cloud-based progressive prognostic assessment module also includes a cross-patient prognostic similarity matching unit, used to match historical patients with similar prognostic characteristics to the current patient, assisting in prognostic assessment, including: V1: Extract the comprehensive temporal feature vector of the current patient. Information on disease type, course of disease, and affected organs; V2: In the cloud-based historical patient database, cosine similarity is used to calculate the feature similarity with the current patient; V3: Select the top 10 historical patients with the highest similarity and extract their actual prognostic results and the effectiveness of the intervention plan; V4: Uses the prognostic statistics of similar patients as a reference and supplements them in the doctor's prognostic assessment report.

6. The prognostic management system for patients with autoimmune diseases according to claim 5, characterized in that, The cloud-based progressive prognostic assessment module also includes a dynamic risk grading and source tracing unit, which performs the following steps: Step AA1: Based on the patient's individual historical data and data from the same disease population, a Bayesian optimization algorithm is used to determine three risk thresholds. , , ; Step AA2: Classification of recurrence risk levels: Low risk Low to medium risk It is classified as medium to high risk. High risk; Step AA3: Calculate the cumulative recurrence risk value over the past 14 days. The formula is: ； in, This represents the patient's cumulative recurrence risk value over the past 14 days. For the first The probability of recurrence within 1 day; This is the recent risk weighting coefficient, ranging from 0.1 to 0.2, used to assign higher weight to recent data; when Furthermore, if the current risk level is level 2, it will automatically be upgraded to a level 3 warning. Step AA4: Risk factor tracing. The contribution of each feature to the recurrence risk is calculated using multi-scale SHAP values. The top 3 features with the highest absolute values ​​are selected as the main risk factors, and a risk tracing report is generated. Step AA5: When the risk level rises to level 3 or above, or Immediately send early warning information and risk tracing reports to the attending physician's terminal; The cloud-based progressive prognostic assessment module also includes a risk warning and graded response scheduling unit, which automatically schedules corresponding medical resources according to different warning levels, including: W1: Level 1 warning: Health reminders are only sent to patients, without notifying medical staff; W2: Level 2 warning: Push intervention suggestions to the patient's terminal and send a reminder to the responsible nurse's terminal at the same time; W3: Level 3 warning: Push emergency intervention suggestions to the patient's terminal, and send warning information to the terminal of the attending physician and the responsible nurse, requiring follow-up within 24 hours; W4: Level 4 Warning: Immediately activate the emergency green channel, send warning information to the emergency department, attending physician and patient's family simultaneously, and push pre-hospitalization notice.

7. The prognostic management system for patients with autoimmune diseases according to claim 6, characterized in that, The intervention intensity matching and adverse reaction early warning module includes: Intervention intensity calculation unit, calculates the overall intervention intensity The formula is: ； in, The overall intervention intensity, ranging from 0 to 1, considers the risks of recurrence, target organ damage, and acute complications; based on... The value of is used to divide the intervention intensity into four levels: Basic level, For the enhanced level, It is an enhanced level. Level 3: Emergency; The medication adjustment recommendation unit generates tiered suggestions based on the intensity of the intervention: Basic level: Maintain the current medication dosage and have a follow-up examination every 3 months; Enhanced level: Increase the immunosuppressant dose by 10%-15%, or shorten the dosing interval of biologics by 10%; Enhanced level: Add low-dose glucocorticoids and adjust the type of immunosuppressant; Emergency level: Immediately activate the emergency green channel, send a pre-hospitalization notice, and recommend high-dose hormone pulse therapy; Adverse reaction risk assessment unit calculates the adverse drug reaction risk index. The formula is: ； in, The adverse reaction risk index for patients using the current medication regimen, with a value ranging from 0 to 1; For the first The basic adverse reaction risk of this drug is obtained from the drug's instructions and clinical data statistics; The individual risk correction factor is calculated based on the patient's liver and kidney function, genotype, age, and concomitant medications; when If the drug is deemed high-risk, an alternative treatment should be recommended. The Lifestyle and Follow-up Planning section generates personalized recommendations: Exercise duration: ,in The baseline exercise duration is set by the doctor based on the patient's physical condition. Follow-up examination interval: ; The follow-up examination items are automatically matched according to the affected organs. For example, if the kidneys are involved, a 24-hour urine protein quantification test will be added. The intervention intensity matching and adverse reaction early warning module also includes an intervention protocol personalization adaptation unit, used to adjust the intervention protocol according to the patient's lifestyle and preferences, including: X1: Collect information on the patient's daily routine, dietary habits, exercise preferences, and work schedule; X2: Adjust the medication reminder time according to the patient's daily routine to avoid conflicts with work or rest; X3: Recommend corresponding exercise methods based on the patient's exercise preferences, such as recommending swimming instead of brisk walking if the patient prefers swimming; X4: Adjust dietary recommendations according to the patient's eating habits, and avoid recommending foods that the patient dislikes or cannot obtain; X5: Generates the final personalized intervention plan, which is then implemented after patient confirmation.

8. The prognostic management system for patients with autoimmune diseases according to claim 7, characterized in that, The intervention effect closed-loop correction and federated learning module performs the following operations: Step BB1: Collect patient data on days 7, 14, and 28 after intervention implementation, and calculate intervention effectiveness scores. The formula is: ； in, The effectiveness of the intervention program is scored, with a value ranging from -1 to 1. A larger value indicates a better intervention effect. , , The disease activity correction value, relapse risk probability, and real-time symptom risk index before intervention; , , The corresponding value after intervention; Step BB2: Grading of Intervention Effects: For significant effectiveness, For it to be effective, Invalid; Step BB3: Model Correction: When When, keep the parameters unchanged; when When, the matching weight of the corresponding intervention plan is increased proportionally; when At that time, incremental training of the model is triggered; Step BB4: Incremental Federated Learning Update: Each edge node trains a local model based on the latest local data and calculates the incremental gradient. ; The incremental gradient is encrypted using a homomorphic encryption algorithm and then uploaded to the cloud. Federated average aggregation is performed in the cloud to generate a global incremental gradient. ,in The number of edge nodes participating in federated learning; The global model is updated in the cloud and encrypted and distributed to each edge node. The intervention effect closed-loop correction and federated learning module also includes a model incremental update triggering and verification unit, used to control the frequency and quality of model incremental updates and avoid model drift, including: Y1: Set the minimum trigger interval for incremental model updates to 1 month to avoid frequent updates that could lead to model instability; Y2: When incremental training of the model is triggered, use local data from the most recent 3 months for training; Y3: After training is complete, use an independent validation set to calculate the prediction accuracy of the updated model; Y4: If the accuracy of the updated model improves by ≥2%, the new model will be adopted; otherwise, the original model will be retained and the model will be waited for the next update.

9. The prognostic management system for patients with autoimmune diseases according to claim 8, characterized in that, The autoimmune disease patient prognosis management system also includes a multimodal loss robustness module, which performs the following operations: Step CC1: When a certain mode When data is missing, renormalize the effective contribution coefficients of other available modes: ； Formula explanation: For the renormalized first The effective modal contribution coefficient ensures that the sum of the contribution coefficients of all available modes is 1. Step CC2: Based on generative adversarial networks, generate virtual features for missing modalities using available modal data. ; Step CC3: Calculate the robust fusion feature vector The formula is: ； in, This is the robust fusion feature vector when modality is missing, with the same dimension as the original fusion feature vector; The feature vectors of the available modes; This is the virtual feature weight coefficient, with a value of 0.2-0.4, used to control the influence of virtual features on the fusion result; Step CC4: Use Replace the original fusion features for prognostic assessment, and send a data missing reminder to the patient's terminal; The intervention effect closed-loop correction and federated learning module also includes a missing modality data completion quality assessment unit, used to evaluate the completion quality of virtual features and dynamically adjust the weights of virtual features, including: Z1: After the missing modal data is completed, calculate the mean square error (MSE) between the virtual features and the real features; Z2: If MSE < 0.1, then the virtual feature weights will be adjusted. Adjusted to 0.4; Z3: If 0.1 ≤ MSE < 0.3, then the virtual feature weights will be adjusted. Adjusted to 0.3; Z4: If MSE ≥ 0.3, then the virtual feature weights will be adjusted. Adjust it to 0.2 and mark the completed data as low confidence.

10. A prognostic management system for patients with autoimmune diseases according to claim 9, characterized in that, The shared decision-making and data feedback module for doctors and patients includes: The visualization unit simultaneously displays disease activity curves, recurrence risk trend charts, target organ damage radar charts, and adverse reaction risk warnings to both doctors and patients. The prognostic simulation unit supports simulating prognostic changes under different intervention protocols and calculating the simulated risk of recurrence. ,in The effectiveness coefficient of the intervention program is derived from historical data in the intervention program database. To simulate the intensity of intervention; The decision recording unit automatically records the discussion process between doctors and patients, the selection of treatment options, and the patient's informed consent, generating electronic medical record archives. The data feedback unit encrypts and transmits patient symptom feedback, medication administration, and intervention effect data back to the multimodal edge data acquisition and preprocessing module, triggering the next round of prognostic assessment. The patient decision-making preference learning unit is used to learn patients' decision-making preferences and optimize subsequent intervention recommendations, including: AA1: Record the patient's acceptance or rejection of each recommended intervention, and the reason for rejection; AA2: Extract characteristics of patients who refuse treatment plans, such as high risk of drug side effects, high exercise intensity, and high frequency of follow-up examinations; AA3: When recommending interventions in the future, reduce the matching weight of protocols that include patient rejection characteristics; AA4: Update the patient's decision preference model every 3 months to ensure that the recommended plan is more in line with the patient's needs.

Images