Remote intelligent management method and system for rheumatism and immunity patients

Abstract

The present application relates to the technical field of rheumatism and immunity disease management, and particularly relates to a remote intelligent management method and system for rheumatism and immunity patients. Clinical indexes, environmental monitoring and life behavior data are acquired to construct a multi-element characteristic vector, which is input into a deep learning network to obtain a disease activity prediction result; the prediction result and the characteristic vector are input into a recurrence risk assessment network containing a collaborative constraint layer, and a consistency correction is performed based on a medical prior constraint to obtain a risk assessment result; a counterfactual replacement is performed on the characteristic vector through a causal inference algorithm, the risk assessment network is input to obtain a counterfactual prediction value, a difference value is calculated as a causal contribution degree, risk driving factors are screened, and quantitative intervention suggestions are generated; and operable early warning information is generated based on the driving factors and the intervention suggestions. The present application improves the accuracy of recurrence risk prediction and the pertinence of intervention suggestions.

Description

Technical Field

[0001] This invention relates to the field of rheumatic and immune disease management technology, and in particular to a remote intelligent management method and system for rheumatic and immune patients. Background Technology

[0002] Rheumatic and immunological diseases, as a class of chronic, relapsing autoimmune diseases, rely heavily on regular patient follow-up visits and physicians' experiential judgment for disease management. Current practices typically involve physicians assessing disease activity based on clinical indicators at the time of patient visit (such as erythrocyte sedimentation rate, C-reactive protein, and number of swollen joints), and then combining this with the patient's subjective symptom descriptions to formulate treatment plans. While some remote management systems have attempted to incorporate electronic health records or mobile questionnaires to collect patient self-reported data, these methods still primarily rely on single measurements or simple statistics, lacking integrated analysis of multidimensional dynamic data. These methods often depend on fixed thresholds or linear regression models to classify disease severity, making it difficult to capture the complex nonlinear relationship between disease activity and relapse.

[0003] A significant shortcoming of conventional approaches lies in the lack of individualization and dynamism in predicting relapse risk. Existing systems often rely solely on simple risk stratification based on current clinical indicators, neglecting the synergistic effects of environmental factors (such as UV intensity and seasonal changes) and patient lifestyle behaviors (such as medication adherence and sleep quality) on disease progression. This results in predictions lagging behind actual disease progression. Furthermore, physicians often rely on empirical rules when developing intervention recommendations, making it difficult to quantify the specific contribution of different risk factors to the relapse probability. Interventions are often characterized by "one-size-fits-all" recommendations (such as "rest and take medication on time"), lacking precise identification and prioritization of modifiable patient behaviors, leading to wasted medical resources and insufficient patient adherence. Summary of the Invention

[0004] The present invention provides a remote intelligent management method and system for rheumatic and immunological patients, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a remote intelligent management method for rheumatoid and immunological patients, comprising:

[0006] Acquire clinical indicator data, environmental monitoring data, and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain disease activity prediction results;

[0007] The disease activity prediction results and the multivariate feature vectors are input into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships.

[0008] The target feature components in the multivariate feature vector are replaced by counterfactual substitution using a causal inference algorithm. The replaced counterfactual feature vector is then input into the relapse risk assessment network to obtain a counterfactual risk prediction value. The difference between the relapse risk assessment result and the counterfactual risk prediction value is calculated as the causal contribution. Risk driving factors are screened based on the causal contribution distribution, and intervention suggestions that quantify the expected risk reduction are generated.

[0009] Actionable early warning information is generated based on the aforementioned risk drivers and intervention recommendations.

[0010] Acquire clinical indicator data, environmental monitoring data, and lifestyle behavior data of the target patient; construct a multivariate feature vector; input the multivariate feature vector into a deep learning network to obtain disease activity prediction results, including:

[0011] The clinical indicator data, environmental monitoring data, and lifestyle behavior data are time-aligned and feature-encoded to construct a multivariate feature vector;

[0012] The multivariate feature vectors are input into the deep learning network, which is trained using a combined loss function. This combined loss function includes a disease activity prediction loss term, a temporal smoothing constraint term, and a high-activity sensitivity enhancement term. The disease activity prediction loss term measures the deviation between the predicted disease activity and the labeled true disease activity. The temporal smoothing constraint term calculates the magnitude of change in predicted disease activity for the same patient within adjacent time windows and penalizes changes exceeding physiologically reasonable thresholds. The high-activity sensitivity enhancement term assigns a higher loss weight to the prediction error of samples labeled as high-activity states compared to regular samples.

[0013] Output the disease activity prediction results generated by the deep learning network after training is completed.

[0014] The recurrence risk assessment network includes a collaborative constraint layer, which performs consistency correction on the recurrence risk assessment results based on medical prior constraints, including:

[0015] The disease activity prediction results and the multivariate feature vectors are input into the relapse risk assessment network to obtain the initial relapse risk assessment results;

[0016] The collaborative constraint layer constructs hierarchical medical prior constraint relationships, including strong constraint rules and weak constraint rules. The strong constraint rules set unbreakable logical boundary conditions between disease activity state and recurrence risk state, while the weak constraint rules set statistical correlation boundary conditions between disease activity state and recurrence risk state.

[0017] Calculate the first confidence level of the disease activity prediction result and the second confidence level of the initial recurrence risk assessment result; identify the constraint rules that violate the hierarchical medical prior constraint relationship; determine the correction priority according to the type of the violated constraint rule and prioritize satisfying the strong constraint rule; calculate the reciprocal of the first confidence level as the disease activity correction weight and the reciprocal of the second confidence level as the recurrence risk correction weight; construct a constraint optimization objective function based on the disease activity correction weight and the recurrence risk correction weight; minimize the weighted sum of the disease activity correction amount and the recurrence risk correction amount under the boundary conditions of satisfying the violated constraint rules; solve the constraint optimization objective function to obtain the corrected recurrence risk assessment result.

[0018] The training of the relapse risk assessment network includes:

[0019] Basic prediction training phase: The relapse risk assessment network is trained using fully labeled training samples, which include disease activity labels, multivariate feature vector labels, and ground truth labels of relapse risk confirmed by long-term follow-up.

[0020] Constraint-guided pseudo-label generation stage: For incomplete labeled samples containing only short-term follow-up data, the network output trained on the basic prediction is constrained and corrected through the collaborative constraint layer to generate constraint-guided pseudo-label relapse risk values;

[0021] Constraint-aware joint training phase: Construct a joint training sample set, which includes fully labeled training samples and incompletely labeled samples carrying constraint-guided pseudo-label recurrence risk values. Freeze the parameters of the feature extraction layer of the recurrence risk assessment network. Train the risk assessment layer and the collaborative constraint layer. Calculate the loss for fully labeled training samples using the ground truth recurrence risk annotation, and calculate the loss for incompletely labeled samples using constraint-guided pseudo-label recurrence risk values.

[0022] The causal inference algorithm is used to perform counterfactual replacement on the target feature components in the multivariate feature vector and calculate the causal contribution, including:

[0023] The feature components in the multivariate feature vector are divided into modifiable features and non-modifiable features. The modifiable features include lifestyle-related features, and the non-modifiable features include age- and genetic-related features.

[0024] The modifiable features are subjected to single-feature counterfactual replacement and multi-feature combined counterfactual replacement. The single-feature counterfactual replacement replaces a single modifiable feature with a target improvement value, and the multi-feature combined counterfactual replacement jointly replaces multiple modifiable features with a combined target value, thereby obtaining single-feature counterfactual risk prediction values ​​and combined counterfactual risk prediction values, respectively.

[0025] The difference between the recurrence risk assessment result and the single-feature counterfactual risk prediction value is calculated as the single-feature causal contribution; the difference between the recurrence risk assessment result and the combined counterfactual risk prediction value is calculated, and the sum of the causal contributions of each single feature participating in the combination is subtracted from this difference to obtain the feature interaction causal contribution.

[0026] The risk attribution degree is calculated by counterfactual substitution of the aforementioned non-interventionist features;

[0027] Risk drivers are screened based on single-feature causal contribution, feature interaction causal contribution, and risk attribution, and then the interventionable causal targets and non-interventionable risk attribution factors are distinguished.

[0028] Generate intervention recommendations that quantify the expected reduction in risk, including:

[0029] For each operable causal target, multiple candidate intervention intensities are set. The expected risk reduction and intervention target value corresponding to different intervention intensities are calculated through counterfactual inference. The minimum intervention intensity that reduces the expected risk to the target threshold is determined as the personalized intervention amount for that causal target.

[0030] A multi-objective optimization framework is constructed, wherein the optimization objectives of the multi-objective optimization framework include maximizing the reduction of total expected risk, minimizing the cost of intervention implementation, and maximizing patient compliance feasibility;

[0031] The multi-objective optimization framework is used to solve both single-causal target intervention schemes and multi-causal target combination intervention schemes to obtain the Pareto-optimal set of intervention schemes.

[0032] Output the Pareto optimal intervention set, where each intervention set includes the personalized intervention amount, expected risk reduction, and implementation cost estimate.

[0033] Based on the aforementioned risk drivers and intervention recommendations, actionable early warning information is generated, including:

[0034] The risk level is determined based on the recurrence risk assessment results, and the risk level is divided into low risk, medium risk, and high risk;

[0035] Differentiated early warning strategies are set for different risk levels. Regular follow-up reminders are generated for low-risk patients, key monitoring early warnings are generated for medium-risk patients and the risk driving factors are marked, and emergency early warnings are generated for high-risk patients and the intervention recommendations are associated with them.

[0036] The risk drivers are translated into patient-understandable risk explanation texts that explain the reasons for the current increased risk of relapse.

[0037] The intervention recommendations are translated into specific, executable operational instructions, which include the specific execution method and frequency of the intervention measures.

[0038] An actionable warning message is generated and pushed to the patient and doctor terminals. The actionable warning message includes the risk level, the risk explanation text, and the operation instructions.

[0039] A second aspect of the present invention provides a remote intelligent management system for rheumatoid and immunological patients, comprising:

[0040] The prediction and evaluation unit is used to acquire clinical indicator data, environmental monitoring data and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain the disease activity prediction result.

[0041] A risk correction unit is used to input the disease activity prediction result and the multivariate feature vector into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships.

[0042] The causal analysis unit is used to perform counterfactual replacement on the target feature components in the multivariate feature vector using a causal inference algorithm, input the replaced counterfactual feature vector into the relapse risk assessment network to obtain a counterfactual risk prediction value, calculate the difference between the relapse risk assessment result and the counterfactual risk prediction value as the causal contribution degree, screen risk driving factors according to the causal contribution degree distribution, and generate intervention suggestions that quantify the expected risk reduction.

[0043] The early warning output unit is used to generate actionable early warning information based on the risk driving factors and the intervention recommendations.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] This method significantly improves the accuracy and timeliness of disease management for patients with rheumatology and immunology. By integrating clinical indicators, environmental and lifestyle data to construct a multivariate feature vector, and leveraging deep learning networks to predict disease activity in real time, it can capture disease fluctuation signals in advance, avoiding the lag caused by relying on subjective patient reports or regular follow-up visits. The relapse risk assessment network introduces a collaborative constraint layer, using prior medical knowledge to perform consistency correction on the prediction results, effectively suppressing misjudgments caused by data noise or model bias, thus greatly improving the accuracy of relapse early warning and reducing unnecessary medical interventions and patient anxiety.

[0050] The introduction of causal inference algorithms enables in-depth analysis of risk factors. By calculating the causal contribution of each feature component through counterfactual substitution techniques, the true risk drivers can be extracted from confounding variables, rather than relying solely on correlation analysis. This quantitative assessment mechanism can clearly distinguish between controllable factors (such as medication adherence) and uncontrollable factors (such as seasonal climate), thereby generating targeted intervention recommendations and providing patients with actionable, data-supported guidelines.

[0051] Intervention recommendations generated based on causal contribution distribution possess dynamic adaptive capabilities. The system continuously tracks the evolution of risk drivers, and when the contribution of a certain factor changes significantly, the warning information is updated synchronously. This closed-loop management strategy avoids the limitations of static solutions, enabling patients to flexibly adjust their behavior according to their own conditions, while providing doctors with quantitative decision-making basis, ultimately achieving a shift from passive treatment to proactive prevention, significantly reducing recurrence rates and medical costs. Attached Figure Description

[0052] Figure 1 A flowchart illustrating the remote intelligent management method for rheumatoid and immunological patients. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes will not be repeated in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the remote intelligent management method for rheumatic and immunological patients according to an embodiment of the present invention.

[0056] Remote intelligent management methods for rheumatoid and immunological patients include:

[0057] Acquire clinical indicator data, environmental monitoring data, and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain disease activity prediction results;

[0058] The disease activity prediction results and the multivariate feature vectors are input into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships.

[0059] The target feature components in the multivariate feature vector are replaced by counterfactual substitution using a causal inference algorithm. The replaced counterfactual feature vector is then input into the relapse risk assessment network to obtain a counterfactual risk prediction value. The difference between the relapse risk assessment result and the counterfactual risk prediction value is calculated as the causal contribution. Risk driving factors are screened based on the causal contribution distribution, and intervention suggestions that quantify the expected risk reduction are generated.

[0060] Actionable early warning information is generated based on the aforementioned risk drivers and intervention recommendations.

[0061] In one optional implementation, clinical indicator data, environmental monitoring data, and lifestyle behavior data of the target patient are acquired to construct a multivariate feature vector. This multivariate feature vector is then input into a deep learning network to obtain disease activity prediction results, including:

[0062] The clinical indicator data, environmental monitoring data, and lifestyle behavior data are time-aligned and feature-encoded to construct a multivariate feature vector;

[0063] The multivariate feature vectors are input into the deep learning network, which is trained using a combined loss function. This combined loss function includes a disease activity prediction loss term, a temporal smoothing constraint term, and a high-activity sensitivity enhancement term. The disease activity prediction loss term measures the deviation between the predicted disease activity and the labeled true disease activity. The temporal smoothing constraint term calculates the magnitude of change in predicted disease activity for the same patient within adjacent time windows and penalizes changes exceeding physiologically reasonable thresholds. The high-activity sensitivity enhancement term assigns a higher loss weight to the prediction error of samples labeled as high-activity states compared to regular samples.

[0064] Output the disease activity prediction results generated by the deep learning network after training is completed.

[0065] For example, clinical indicators typically come from hospital information systems or patient-reported test results, including erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), anti-cyclic citrullinated peptide antibody (anti-CCP), number of swollen joints, number of tender joints, and Patient Overall Health Assessment (PGA) score. The collection frequency of these indicators varies; some indicators are only recorded every few weeks, while patient-reported scores are collected daily. Environmental monitoring data comes from meteorological sensors linked to the patient or publicly available meteorological databases associated with their geographic location, including ambient temperature, relative humidity, atmospheric pressure, and air quality index (AQI), collected typically hourly or daily. Lifestyle data comes from wearable devices or patient mobile applications, including daily steps, sleep duration, sleep quality score, medication adherence records, and dietary logs, also collected daily.

[0066] Because the collection timestamps of the three types of data are inconsistent, all data need to be aligned using a unified time window (e.g., with the calendar day as the smallest granularity). For data with multiple records within a certain time window, the mean or the most recent valid value within that window is used. For data missing within a certain time window, linear interpolation or forward imputation strategies based on the same patient's historical sequence are used to complete the data, and corresponding missing marker bits are generated and incorporated into the vector as additional feature components to preserve data integrity information. After time alignment, each feature component is normalized. Continuous numerical features are standardized using z-score based on the training set distribution, making the mean of each feature 0 and the standard deviation 1. Categorical features (such as medication regimen type) are converted into numerical vectors using one-hot encoding. All aligned, completed, and encoded feature components are concatenated in a fixed order to form the feature vector for each time step. Then, the feature vectors of several consecutive time steps are stacked using a sliding time window method to construct a multivariate feature vector sequence containing temporal context, which serves as the input to the deep learning network.

[0067] Deep learning networks employ sequence modeling structures capable of capturing temporal dependencies, such as architectures based on Transformer encoders or Long Short-Term Memory (LSTM) networks. Their input is the aforementioned multivariate feature vector sequence, and their output is the disease activity prediction result for the current time step. The disease activity prediction result can be a continuous disease activity score (such as a predicted DAS28 score) or a probability distribution of the corresponding discrete activity levels (remission, low activity, moderate activity, high activity), depending on the task setting.

[0068] The network training process employs a combined loss function, which consists of a weighted sum of three loss terms: a disease activity prediction loss term, a temporal smoothing constraint term, and a high activity sensitivity enhancement term.

[0069] The disease activity prediction loss term measures the deviation between the disease activity prediction output by the network and the labeled true disease activity values. For regression tasks, mean squared error (MSE) or mean absolute error (MAE) is used as this loss term; for classification tasks, cross-entropy loss is used. Let the... The predicted value for each sample is The corresponding labeled truth value is Taking regression tasks as an example, predicting the loss term for: ;in This represents the total number of samples used in the calculation.

[0070] The temporal smoothing constraint is designed based on the physiological laws of rheumatic and immunological diseases: disease activity should not undergo drastic changes in the short term; if the difference between predicted values ​​within adjacent time windows exceeds the physiologically reasonable range, it indicates instability in network prediction and requires penalty. Let the same patient... The predicted value at each time step is In the The predicted value at each time step is The threshold for reasonable physiological changes is Time-series smoothing constraint term for:

[0071] ;in For the total number of patients, Indicates the first The patient in the The predicted value at each time step, The preset threshold for reasonable physiological variation can be set based on clinical experience. For example, in the DAS28 scoring system, a change exceeding 1.2 points within two consecutive weeks is considered to be outside the reasonable physiological range. This constraint only penalizes changes exceeding the threshold and does not restrict normal fluctuations within the threshold range, thus achieving a balance between smoothness and flexibility.

[0072] The high-activity sensitivity enhancement term is designed to improve the network's prediction accuracy for high-activity state samples. In the management of rheumatic and immunological diseases, a high-activity state signifies a higher risk of joint damage and functional loss. The cost of underreporting high-activity states far outweighs the prediction error of ordinary samples; therefore, a higher loss weight needs to be assigned to high-activity samples. Let the sample... The weighting coefficient corresponding to the activity level labeling is For samples labeled as high activity, For other samples, assign a boost weight greater than 1 (e.g., 3.0). If we set it to 1.0, then the high activity sensitivity enhancement term... for: .

[0073] Final combined loss function The weighted sum of the three loss terms: ;in , , The weights of the three loss terms are hyperparameters, which can be determined by grid search or Bayesian optimization on the validation set. The three loss terms work synergistically: the prediction loss term ensures overall prediction accuracy, the temporal smoothing constraint term suppresses non-physiological jitter in the predicted sequence, and the high activity sensitivity enhancement term ensures that the network has a higher recall rate under clinical critical conditions. Thus, the trained deep learning network balances accuracy, stability, and clinical safety in the task of predicting disease activity in rheumatoid and immunological patients.

[0074] After training, the multivariate feature vector sequence of the target patient is input into the network to obtain the disease activity prediction result.

[0075] In one optional implementation, the relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment results based on medical prior constraints, including:

[0076] The disease activity prediction results and the multivariate feature vectors are input into the relapse risk assessment network to obtain the initial relapse risk assessment results;

[0077] The collaborative constraint layer constructs hierarchical medical prior constraint relationships, including strong constraint rules and weak constraint rules. The strong constraint rules set unbreakable logical boundary conditions between disease activity state and recurrence risk state, while the weak constraint rules set statistical correlation boundary conditions between disease activity state and recurrence risk state.

[0078] Calculate the first confidence level of the disease activity prediction result and the second confidence level of the initial recurrence risk assessment result; identify the constraint rules that violate the hierarchical medical prior constraint relationship; determine the correction priority according to the type of the violated constraint rule and prioritize satisfying the strong constraint rule; calculate the reciprocal of the first confidence level as the disease activity correction weight and the reciprocal of the second confidence level as the recurrence risk correction weight; construct a constraint optimization objective function based on the disease activity correction weight and the recurrence risk correction weight; minimize the weighted sum of the disease activity correction amount and the recurrence risk correction amount under the boundary conditions of satisfying the violated constraint rules; solve the constraint optimization objective function to obtain the corrected recurrence risk assessment result.

[0079] For example, after simultaneously inputting the disease activity prediction result and the multivariate feature vector into the relapse risk assessment network, the network first passes through a feature fusion layer and a risk regression layer, outputting an initial relapse risk assessment result. This initial result is based solely on data-driven statistical learning and has not yet incorporated prior knowledge constraints from the medical field. Therefore, in boundary cases, predictive outputs may appear that contradict clinical cognition, such as giving an extremely low relapse risk score in a highly active disease state, or predicting an abnormally high risk value in patients in remission. The design of the collaborative constraint layer is precisely to identify and correct these inconsistencies at the output stage.

[0080] The collaborative constraint layer constructs hierarchical medical prior constraint relationships during the network inference phase, dividing constraint rules into two levels: strong constraint rules and weak constraint rules. Strong constraint rules correspond to boundary conditions with deterministic logical support in clinical practice, reflecting the inviolable logical relationship between disease activity status and relapse risk status. Taking rheumatoid arthritis as an example, when the disease activity score (DAS28) exceeds the clinically recognized high activity threshold, the relapse risk level cannot be judged as "extremely low risk"; when the patient is in complete remission and all biomarkers are within the normal range, the relapse risk cannot be judged as "extremely high risk". These rules have clear logical necessity in medicine, and predictions that violate these rules have no reasonable explanation in clinical practice, therefore they are given the highest correction priority. Weak constraint rules correspond to boundary conditions at the statistical correlation level, derived from probabilistic patterns revealed by large-scale rheumatological and immunological disease cohort studies. For example, when disease activity is moderately elevated, the relapse risk score usually falls within a certain statistical confidence interval; there is a significant positive correlation between a sustained increase in specific inflammatory markers (such as CRP and ESR) and the probability distribution of relapse risk. Weak constraint rules allow for exceptions within a certain probability range, but when the prediction results exceed the statistical boundary, a certain degree of correction is required.

[0081] When identifying constraint violations, the set of strong constraints and the set of weak constraints are traversed sequentially to determine whether the combination of the current disease activity prediction result and the initial relapse risk assessment result falls into the prohibited area or statistical out-of-bounds area defined by each rule. For each violated rule, its level (strong or weak) and the degree of violation (i.e., the deviation of the current predicted value from the boundary condition) are recorded. The determination of correction priority follows the hierarchical priority principle: the correction needs of all violated strong constraints are met before the correction needs of weak constraints; within the same level, rules with a greater degree of violation receive a higher correction priority.

[0082] Confidence calculation is a crucial preliminary step in constructing the constrained optimization objective function. The first confidence level of the disease activity prediction results. The degree of confidence a deep learning network has in predicting the current activity level can be quantified by the probability distribution entropy of the network's output layer, the prediction variance of Monte Carlo Dropout sampling, or the maximum probability value of Softmax after temperature scaling. A higher value indicates greater confidence in the prediction by the network. Second confidence level of the initial recurrence risk assessment result. Calculated in a similar manner, this reflects the risk assessment network's level of confidence in the current risk score. .

[0083] Disease activity correction weight Defined as the reciprocal of the first confidence level, i.e. Relapse risk adjusted weights Defined as the reciprocal of the second confidence level, i.e. The physical implication of this design is that when the network has a low confidence level in a prediction, the reliability of the prediction itself is questionable, and a larger correction should be allowed; conversely, when the confidence level is high, the original prediction should be preserved as much as possible, with only the minimum necessary corrections. By converting the confidence level into correction weights through a reciprocal relationship, an adaptive correction mechanism is achieved where "low confidence level corresponds to high correction tolerance, and high confidence level corresponds to low correction tolerance."

[0084] Constrained optimization objective function Disease activity correction With recurrence risk correction The weighted sum is the minimization objective, expressed as: ,in This is the adjustment amount for the predicted disease activity value relative to the original predicted value. This represents the adjustment amount of the recurrence risk assessment result relative to the initial assessment result. The objective function is in squared form to penalize excessive adjustments, ensuring that the corrected result is as close as possible to the original prediction while satisfying the constraints. The constraints consist of all boundary conditions of the rules identified as being violated: for each violated strong constraint, the corresponding inequality or equality constraint is added to the optimization problem as a hard constraint; for each violated weak constraint, the corresponding statistical boundary condition is added as a soft constraint, which can be handled by introducing slack variables and adding a penalty term to the objective function.

[0085] When solving the above constrained optimization problem, since the objective function is quadratic and the constraints are mostly linear inequalities, the overall problem constitutes a quadratic programming (QP) problem, which can be solved efficiently using the interior-point method or the effective set method. Under the hierarchical principle of prioritizing strong constraint rules, the solution process is divided into two stages: the first stage activates only the strong constraints to obtain the minimum correction amount that satisfies all strong constraints; the second stage, based on the solution from the first stage, further activates the weak constraints, satisfying the weak constraint boundaries as much as possible without violating the strong constraints. The final solution obtained... and The original disease activity prediction value and the initial relapse risk assessment result are respectively superimposed to obtain the corrected relapse risk assessment result.

[0086] The corrected relapse risk assessment results simultaneously satisfy the logical and statistical boundaries of medical prior constraints, eliminating inconsistencies with clinical knowledge while retaining the advantages of data-driven prediction. For predictions with low confidence, more significant corrections are allowed to align with prior constraints; for predictions with high confidence, only minimal adjustments are made when constraints are violated, thus achieving a dynamic balance between data-driven and knowledge-driven approaches. This correction mechanism enhances the interpretability and credibility of relapse risk assessment results in clinical practice, providing a reliable risk assessment basis for subsequent causal inference and intervention recommendations.

[0087] In one alternative implementation, the training of the relapse risk assessment network includes:

[0088] Basic prediction training phase: The relapse risk assessment network is trained using fully labeled training samples, which include disease activity labels, multivariate feature vector labels, and ground truth labels of relapse risk confirmed by long-term follow-up.

[0089] Constraint-guided pseudo-label generation stage: For incomplete labeled samples containing only short-term follow-up data, the network output trained on the basic prediction is constrained and corrected through the collaborative constraint layer to generate constraint-guided pseudo-label relapse risk values;

[0090] Constraint-aware joint training phase: Construct a joint training sample set, which includes fully labeled training samples and incompletely labeled samples carrying constraint-guided pseudo-label recurrence risk values. Freeze the parameters of the feature extraction layer of the recurrence risk assessment network. Train the risk assessment layer and the collaborative constraint layer. Calculate the loss for fully labeled training samples using the ground truth recurrence risk annotation, and calculate the loss for incompletely labeled samples using constraint-guided pseudo-label recurrence risk values.

[0091] For example, the training process of the relapse risk assessment network is divided into three progressive stages. Each stage adopts a differentiated training strategy for different types of labeled data, thereby making full use of the incomplete labeling problem that is common in clinical follow-up data and improving the network's generalization ability in real medical scenarios.

[0092] Basic prediction training phase:

[0093] In the clinical follow-up system for rheumatic and immunological diseases, patient samples with complete long-term follow-up records are the most valuable source of supervisory signals. Completely annotated training samples must simultaneously possess three types of annotation information: disease activity annotations assessed by specialists according to international standards such as ACR / EULAR; multivariate feature vector annotations covering clinical indicators, environmental monitoring, and lifestyle behaviors; and ground truth annotations of relapse risk confirmed through at least 12 to 24 months of long-term follow-up. Relapse risk ground truth annotations are typically recorded in a binary form (relapse / no relapse) or a continuous score, corresponding to whether the patient experienced clinical events such as a significant increase in disease activity, progression of joint damage, or the need for escalation of treatment during the follow-up period.

[0094] The basic prediction training phase uses the aforementioned complete labeled samples to train all parameters of the recurrence risk assessment network in an end-to-end manner, including the feature extraction layer, risk assessment layer, and collaborative constraint layer. The feature extraction layer is responsible for mapping multivariate feature vectors to a high-dimensional semantic representation space, capturing the nonlinear interaction relationships between different feature dimensions; the risk assessment layer outputs a recurrence risk score based on the semantic representation; and the collaborative constraint layer applies medical prior constraints to the risk score during forward propagation, ensuring that the output results satisfy common-sense medical relationships such as the monotonicity constraint between disease activity and recurrence risk, and the risk range constraint corresponding to different combinations of biomarkers.

[0095] In terms of loss function design, the difference between the ground truth label of relapse risk and the network output is directly supervised during the basic prediction training phase. Simultaneously, auxiliary supervision is introduced for the disease activity prediction branch, enabling the feature extraction layer to form an effective representation of the disease state from the early stages of training. The training process employs a mini-batch stochastic gradient descent strategy, balancing positive and negative samples (relapsed / non-relapsed) within each batch according to a certain ratio to alleviate the class imbalance problem commonly found in clinical data. After sufficient basic prediction training, the network possesses a stable prediction ability for fully labeled samples, laying a reliable initialization foundation for the subsequent pseudo-label generation stage.

[0096] Constraint-guided pseudo-label generation stage:

[0097] In real-world clinical databases, many patients only have short-term follow-up records (e.g., less than 6 months), making it impossible to obtain ground truth labels for relapse risk confirmed by long-term follow-up. However, these samples contain rich feature information and short-term disease activity patterns, and discarding them directly would result in significant data waste. The goal of the constraint-guided pseudo-label generation stage is to generate reliable relapse risk pseudo-labels for these incomplete labeled samples, thereby expanding the scale of effective training data.

[0098] For each incompletely labeled sample, its multivariate feature vector and disease activity prediction result are input into the relapse risk assessment network trained on the base prediction to obtain the initial network output relapse risk score. Since this sample lacks supervision from long-term follow-up ground truth values, the initial output has some uncertainty. The collaborative constraint layer plays a crucial role here: it corrects the initial output based on medical prior constraints, specifically including the joint application of the following types of constraints.

[0099] First, the monotonicity constraint of disease activity-relapse risk: if the disease activity prediction result of the current sample is in the high activity range (e.g., DAS28 score exceeds 5.1), then the corresponding relapse risk score should not be lower than the preset minimum threshold. The collaborative constraint layer will adjust the output above the threshold. Secondly, biomarker consistency constraint: if key biomarkers such as anti-CCP antibody and RF factor are all positive and at high titers in the sample, the relapse risk score should fall into the corresponding high-risk range. The collaborative constraint layer performs boundary truncation and softening correction on outputs exceeding this range. Thirdly, there is the temporal trend consistency constraint: if the short-term follow-up records of a sample show a continuous upward trend in disease activity, the relapse risk score should be correspondingly higher than that of samples with a stable or declining trend. The collaborative constraint layer corrects scores that violate trend consistency by comparing the outputs of the same patient at different time steps.

[0100] The output value after the above multiple constraint corrections is used as the constraint-guided pseudo-label recurrence risk value for the incomplete labeled sample. This is used in the subsequent joint training phase. To control the quality of pseudo-labels, a constraint satisfaction score is calculated for each pseudo-label. Only constraints that are satisfied to a degree higher than the threshold will be considered. The pseudo-labels are included in the joint training sample set. For samples with low constraint satisfaction (usually corresponding to missing or abnormal feature vectors), pseudo-labels are not generated temporarily to avoid low-quality pseudo-labels from negatively interfering with subsequent training.

[0101] Constraint-aware joint training phase:

[0102] The joint training sample set consists of two parts: fully labeled samples with ground truth relapse risk annotations, and relapse risk values ​​with constraint-guided pseudo-labels. Incomplete labeled samples. Different loss calculation strategies are used for the two types of samples during joint training.

[0103] To prevent noise from false labels from propagating throughout the network via the feature extraction layer, the parameters of the feature extraction layer of the constraint-aware joint training phase are frozen, and only the parameters of the risk assessment layer and the co-constraint layer are updated. The rationale for this freezing strategy is that the feature extraction layer has already been sufficiently trained with fully labeled samples during the basic prediction training phase, possessing stable feature representation capabilities; while the parameters of the risk assessment layer and the co-constraint layer need to be further adapted to a larger sample distribution through joint training to improve robustness to data with varying label quality.

[0104] For fully annotated training samples, supervised loss is calculated using ground truth annotation of recurrence risk. This loss directly measures the difference between the network output and the true value confirmed by long-term follow-up, has the highest supervision reliability, and is assigned a large weight coefficient in the joint loss. For incompletely labeled samples, a constraint-guided pseudo-label recurrence risk value is used. Calculate pseudo-label loss Considering the inherent uncertainty of pseudo-labels, a smaller weighting coefficient is assigned. Furthermore, this weight gradually increases with each training epoch, achieving a smooth transition from ground truth supervision to pseudo-label assistance. In addition, a constraint consistency regularization term is introduced into the joint loss. This method penalizes instances where the collaborative constraint layer outputs violate prior medical constraints during joint training, ensuring the network maintains its adherence to medical constraints while expanding its training data. The total loss during joint training is... ,in The weighting coefficients for the consistency regularization term.

[0105] During the joint training phase, a small learning rate is used for fine-tuning. At fixed training epochs, pseudo-labels for incompletely labeled samples are regenerated to update the pseudo-label quality using the latest network state, forming an iterative self-training closed-loop optimization mechanism. After constraint-aware joint training, the relapse risk assessment network maintains high-accuracy predictions for fully labeled samples while significantly improving its coverage of short-term follow-up samples, effectively alleviating the problem of scarce training data caused by incomplete clinical follow-up data labeling.

[0106] In one optional implementation, the target feature components in the multivariate feature vector are counterfactually replaced and the causal contribution is calculated using a causal inference algorithm, including:

[0107] The feature components in the multivariate feature vector are divided into modifiable features and non-modifiable features. The modifiable features include lifestyle-related features, and the non-modifiable features include age- and genetic-related features.

[0108] The modifiable features are subjected to single-feature counterfactual replacement and multi-feature combined counterfactual replacement. The single-feature counterfactual replacement replaces a single modifiable feature with a target improvement value, and the multi-feature combined counterfactual replacement jointly replaces multiple modifiable features with a combined target value, thereby obtaining single-feature counterfactual risk prediction values ​​and combined counterfactual risk prediction values, respectively.

[0109] The difference between the recurrence risk assessment result and the single-feature counterfactual risk prediction value is calculated as the single-feature causal contribution.

[0110] The difference between the recurrence risk assessment result and the combined counterfactual risk prediction value is calculated, and the sum of the causal contributions of each individual feature participating in the combination is subtracted from this difference to obtain the feature interaction causal contribution.

[0111] The risk attribution degree is calculated by counterfactual substitution of the aforementioned non-interventionist features;

[0112] Risk drivers are screened based on single-feature causal contribution, feature interaction causal contribution, and risk attribution, and then the interventionable causal targets and non-interventionable risk attribution factors are distinguished.

[0113] For example, before conducting causal analysis on multivariate feature vectors, it is necessary to first classify the modifiability of the feature components. Feature components are divided into two main categories based on clinical significance: modifiable features and non-modifiable features. Modifiable features include lifestyle-related components, such as sleep duration, physical activity intensity, dietary structure scores, medication adherence scores, and emotional stress indices; these features are practically operable within a clinical intervention framework, and patients can change them through behavioral adjustments or medical intervention. Non-modifiable features include age- and genetically related components, such as risk allele carrier status at specific gene loci and family history scores; these features cannot be changed through intervention under current medical conditions and are only used as background risk factors for attribution analysis. Clearly classifying feature components into two categories helps subsequent intervention recommendations focus on truly actionable causal targets, avoiding the delivery of meaningless advice to patients.

[0114] For modifiable features, two operations are performed: single-feature counterfactual replacement and multi-feature combination counterfactual replacement. The logic of single-feature counterfactual replacement is as follows: while keeping all other components in the multivariate feature vector unchanged, one modifiable feature component is replaced with a preset target improvement value to construct a counterfactual feature vector. This counterfactual feature vector is then input into the relapse risk assessment network to obtain the corresponding single-feature counterfactual risk prediction value. The target improvement value is set based on clinical reference ranges or behavioral intervention goals; for example, sleep duration is replaced from the current value to the recommended 7 to 8 hours, and physical activity intensity is replaced with the corresponding encoded value for moderate-intensity activity. By sequentially performing the above replacement operation on each modifiable feature component in the multivariate feature vector, a sequence of single-feature counterfactual risk prediction values ​​corresponding to each modifiable feature can be obtained.

[0115] Let the original recurrence risk assessment result be , No. The single-feature counterfactual risk prediction value corresponding to each operable feature is: Then the first Single-feature causal contribution of an operable feature Defined as: ; When the value is positive, it indicates that the risk of relapse is reduced after replacing the feature with the target improvement value, which shows that the feature has a positive driving effect on the current risk of relapse. The larger the value, the more significant the individual causal contribution of this feature to the risk of relapse.

[0116] Multi-feature combined counterfactual replacement extends the single-feature replacement by simultaneously replacing multiple modifiable feature components with their respective combined target values, constructing a combined counterfactual feature vector, and inputting it into a relapse risk assessment network to obtain the combined counterfactual risk prediction value. The selection of the combined target value can be consistent with the target improvement value when replacing a single feature, or it can be specifically set according to the joint intervention plan. Let the set of interventionable features participating in the combined replacement be... Then the feature interaction causal contribution degree Defined as: ; This reflects the additional risk reduction effect generated when multiple interventionable features are jointly improved, exceeding the sum of the individual contributions of each feature; that is, the synergistic intervention effect among features. If A significantly positive value indicates a synergistic gain from the combined intervention; simultaneous improvement of multiple features can lead to a greater reduction in relapse risk than improvement of each feature individually. A value close to zero or negative indicates that the intervention effects of each feature are relatively independent or antagonistic. In the generation of actual intervention recommendations, the causal contribution of feature interactions provides a quantitative basis for developing combined intervention programs, helping clinicians determine whether it is necessary to simultaneously promote multiple lifestyle interventions.

[0117] For non-interventional characteristics, the same counterfactual substitution operation is performed, but its purpose differs from that for interventional characteristics. Counterfactual substitution of non-interventional characteristics involves replacing them with baseline levels or low-risk reference values ​​from the reference population. After constructing a counterfactual feature vector, it is input into the relapse risk assessment network to obtain the counterfactual risk prediction value for the non-interventional characteristic. Let the... The counterfactual risk prediction value of an unmanageable feature is Then the risk attribution degree of this feature Defined as: Risk attribution This quantifies the background contribution of the non-modifiable characteristic to the current patient's relapse risk. For example, older patients or those carrying high-risk genetic loci have higher risk attribution, indicating that a significant portion of their relapse risk stems from unchangeable intrinsic factors. This information is of significant value for clinical decision-making: for patients with high non-modifiable risk attribution, even if the causal contribution of the modifiable characteristic is relatively limited, monitoring frequency and follow-up density should be increased, rather than relying solely on behavioral interventions.

[0118] In obtaining single-feature causal contribution Feature interaction causal contribution Risk attribution with non-intervention characteristics Subsequently, risk drivers were screened and stratified using a combination of three types of indicators. For modifiable features, their causal contribution was determined based on the degree of individual feature contribution. Sort by size, Features exceeding a preset screening threshold are identified as potentially actionable causal targets. Simultaneously, based on the causal contribution of feature interactions, combinations of features with synergistic intervention potential are identified and labeled as joint intervention targets. For non-interventionable features, risk attribution is used... Sort by size, Higher-risk characteristics are identified as attributable factors of non-intervention risks and are presented as background risk descriptions in early warning information, rather than as targets for intervention recommendations.

[0119] Through the above distinction, risk drivers are clearly divided into two categories of outputs: one is modifiable causal targets, which directly correspond to the generation of subsequent intervention recommendations. The expected risk reduction in the intervention recommendations comes from the quantitative results of the corresponding single-feature causal contribution or combined causal contribution. The other is non-modifiable risk attribution factors, which serve as explanatory information on the patient's individualized risk background, helping doctors understand the patient's baseline risk level and providing a basis for adjusting follow-up plans and medication intensity. This categorized output method ensures the operability of intervention recommendations while avoiding the misleading presentation of unchangeable risk factors as intervention targets, thereby enhancing the practical value and patient compliance of remote intelligent management solutions in clinical practice.

[0120] At the implementation level, both single-feature counterfactual substitution and multi-feature combination counterfactual substitution are computed during the inference phase of the recurrence risk assessment network. This eliminates the need for network retraining; only the input feature vector needs to be replaced before forward inference is performed. For cases with a large number of feature components, counterfactual substitution can be prioritized for high-priority, clinically significant interventionable feature subsets to reduce computational overhead. The selection of feature subsets for multi-feature combination substitution can be based on the ranking of single-feature causal contributions, prioritizing the selection of the features with the highest contributions to form the combination set. It balances computational efficiency with analytical depth.

[0121] In one alternative implementation, generating intervention recommendations that quantify the expected reduction in risk includes:

[0122] For each operable causal target, multiple candidate intervention intensities are set. The expected risk reduction and intervention target value corresponding to different intervention intensities are calculated through counterfactual inference. The minimum intervention intensity that reduces the expected risk to the target threshold is determined as the personalized intervention amount for that causal target.

[0123] A multi-objective optimization framework is constructed, wherein the optimization objectives of the multi-objective optimization framework include maximizing the reduction of total expected risk, minimizing the cost of intervention implementation, and maximizing patient compliance feasibility;

[0124] The multi-objective optimization framework is used to solve both single-causal target intervention schemes and multi-causal target combination intervention schemes to obtain the Pareto-optimal set of intervention schemes.

[0125] Output the Pareto optimal intervention set, where each intervention set includes the personalized intervention amount, expected risk reduction, and implementation cost estimate.

[0126] For example, after identifying the risk drivers, the causal analysis results need to be further translated into clinically actionable intervention recommendations. For each feature identified as an interventional causal target—such as medication adherence score, physical activity intensity, sleep quality index, and dietary inflammation index—a set of discrete candidate intervention intensity levels should be established. For example, for medication adherence, candidate intervention intensity levels could be set to increase the current adherence rate by 10%, 20%, 30%, 50%, etc.; for physical activity intensity, candidate intervention intensity levels could be set to increase moderate-intensity aerobic exercise by 30 minutes, 60 minutes, or 90 minutes per week, etc.

[0127] For each candidate intervention level, the corresponding counterfactual feature value is substituted into the relapse risk assessment network to obtain the predicted relapse risk value under that intervention level, and then the expected risk reduction relative to the original relapse risk assessment result is calculated. Let there be an interventionizable causal target. Corresponding to the When considering the strength of each candidate intervention, the counterfactual risk prediction value is... The original recurrence risk assessment result was (Using the aforementioned definition), the expected risk reduction corresponding to this level of intervention is: Meanwhile, the characteristic target value corresponding to this intervention intensity is the intervention target value, such as increasing the medication adherence rate from the current value to the target value, or increasing the weekly exercise time from the current value to the target value.

[0128] After obtaining the expected risk reduction for each candidate intervention level, determine the threshold at which the expected risk reduction first reaches or exceeds the preset target threshold. The minimum intervention level is used as the personalized intervention amount for this causal target. Target threshold. The target threshold can be dynamically set based on the patient's current risk level: for high-risk patients, the target threshold can be set as the minimum reduction required to lower the recurrence risk to the intermediate risk range; for intermediate-risk patients, the target threshold can be set as the minimum reduction required to lower the recurrence risk by more than 20%. Adopting the principle of minimum intervention intensity aims to minimize the intervention burden on patients and improve the long-term sustainability of the intervention plan while ensuring the intervention effect. If all candidate intervention intensities fail to reduce the expected risk to the target threshold, the maximum achievable risk reduction for that target is recorded and considered in subsequent multi-target combination optimization.

[0129] When constructing a multi-objective optimization framework, the optimization objectives encompass three dimensions. The first optimization objective is to maximize the overall expected risk reduction, that is, to maximize the reduction in relapse risk resulting from the combined effect of all intervention targets under the selected intervention regimen. The second optimization objective is to minimize the intervention implementation cost, which comprehensively considers economic costs (such as medication adjustment costs and examination costs), time costs (such as the frequency of visits and time invested in self-management), and physiological burdens (such as the physical demands of exercise on patients). The unit cost coefficient for each type of intervention is pre-calibrated by the clinical knowledge base, and the total intervention cost is the weighted sum of the intervention costs for each target. Let the i-th... Each target point at the intervention intensity level The cost of intervention is The total cost of the plan is the sum of the intervention costs for each selected target. The third optimization objective is to maximize patient compliance feasibility. The compliance feasibility score comprehensively considers the patient's historical compliance record, the complexity of the intervention, the patient's subjective willingness assessment results, and the degree to which the intervention matches the patient's lifestyle. For patients with poor historical compliance, interventions that are easy to implement and integrated into daily life are preferred; for patients with strong lifestyle regularity, interventions that require regular implementation can be appropriately introduced.

[0130] The decision variable in the multi-objective optimization framework is the selection of the intervention intensity level for each causal target, which falls under the category of mixed-integer optimization problems. A multi-objective evolutionary algorithm based on non-dominated ranking (such as NSGA-II methods) is employed for solving the problem, searching for the Pareto front in the objective space composed of three optimization objectives through population iterative evolution. During the optimization process, solutions are found for both single-causal target intervention schemes (i.e., intervening in only one target at a time) and multi-causal target combination intervention schemes (i.e., intervening in multiple targets simultaneously). The advantage of single-target schemes lies in their clear intervention logic, ease of understanding and execution by patients, making them suitable for patients with weak adherence or those with stable current conditions; multi-target combination schemes are suitable for high-risk patients or situations where single-target interventions cannot achieve the target risk reduction.

[0131] When calculating the total expected risk reduction for multi-target combined intervention programs, the interaction effects between features must be considered. All counterfactual feature values ​​corresponding to the combined intervention program should be simultaneously substituted into the relapse risk assessment network to obtain the combined counterfactual risk prediction value. This value should be used as the basis for the total expected risk reduction of the combined program, rather than simply summing the reductions of individual targets, thus avoiding overestimation of the intervention effect due to ignoring feature interactions.

[0132] Each intervention in the Pareto-optimal intervention set is a non-dominated solution, meaning that no other intervention is superior to it in all three optimization objectives. The final Pareto-optimal intervention set includes the following elements for each intervention: the personalized intervention amount for each intervention target (i.e., the specific target value, such as "increasing the daily medication frequency from once to twice" or "increasing weekly walking by 60 minutes"), the corresponding total expected risk reduction (quantified as a percentage), and an estimated intervention implementation cost (listing economic costs, time costs, and physiological burden ratings).

[0133] When presenting Pareto-optimal intervention options to patients or clinicians, a preference weighting mechanism can be further introduced. This mechanism ranks the Pareto frontier options based on the patient's individual priorities (e.g., prioritizing economic cost or adherence feasibility), recommending several options that best suit the patient's individual preferences for clinical decision-making. For patients with adherence feasibility scores below a preset threshold... Even if the proposed intervention significantly reduces risk, it will be downgraded in the recommendation list or have additional compliance risk warnings added to ensure its feasibility in real-world clinical scenarios. The entire intervention recommendation generation process forms a complete closed loop, from causal target identification, personalized intervention dosage determination, multi-objective optimization solution to protocol output, providing quantifiable and traceable precision intervention support for the remote intelligent management of rheumatology and immunology patients.

[0134] In one optional implementation, actionable early warning information is generated based on the risk drivers and the intervention recommendations, including:

[0135] The risk level is determined based on the recurrence risk assessment results, and the risk level is divided into low risk, medium risk, and high risk;

[0136] Differentiated early warning strategies are set for different risk levels. Regular follow-up reminders are generated for low-risk patients, key monitoring early warnings are generated for medium-risk patients and the risk driving factors are marked, and emergency early warnings are generated for high-risk patients and the intervention recommendations are associated with them.

[0137] The risk drivers are translated into patient-understandable risk explanation texts that explain the reasons for the current increased risk of relapse.

[0138] The intervention recommendations are translated into specific, executable operational instructions, which include the specific execution method and frequency of the intervention measures.

[0139] An actionable warning message is generated and pushed to the patient and doctor terminals. The actionable warning message includes the risk level, the risk explanation text, and the operation instructions.

[0140] For example, after obtaining the recurrence risk assessment results, they need to be transformed into actionable early warning information that has practical guiding value for both patients and doctors. This transformation process involves risk level classification, development of differentiated early warning strategies, generation of risk interpretation texts, specification of operational instructions, and multi-platform push notifications.

[0141] Risk levels are determined based on the numerical range of relapse risk assessment results. After normalizing the relapse risk score to a continuous range of 0 to 1, it is mapped to three discrete levels according to a segmentation rule: a risk score below a preset upper threshold for low risk is considered low risk; a risk score between the upper and lower thresholds for low and high risk is considered medium risk; and a risk score above the lower threshold for high risk is considered high risk. These two thresholds are jointly calibrated based on clinical expert experience and historical cohort data before model deployment, and allow for setting independent threshold parameters for different disease types (such as rheumatoid arthritis, systemic lupus erythematosus, ankylosing spondylitis, etc.) to reflect the differences in relapse risk distribution among different diseases. The risk level determination results will serve as the entry condition for all subsequent early warning strategy selections, ensuring that patients with different risk levels receive management responses that match their actual situation.

[0142] For low-risk patients, the early warning strategy primarily relies on maintenance follow-up reminders. Specifically, this involves automatically generating follow-up reminder messages according to a preset follow-up cycle (e.g., every 4 weeks). The reminders include the suggested date for the next follow-up, clinical indicators that need to be self-assessed and reported (such as the number of joint tenderness points, duration of morning stiffness, and visual analog scale score for pain), and daily life precautions. The wording of the early warning messages for low-risk patients is primarily positive and encouraging to avoid causing unnecessary anxiety. At the same time, patients are clearly informed that they should contact their doctor immediately if their symptoms suddenly worsen, preventing delays in responding to changes in their condition due to over-reliance on the low-risk assessment.

[0143] For patients at medium risk, the early warning strategy has been upgraded to focused monitoring and early warning. Building upon follow-up reminders, the follow-up cycle has been shortened (e.g., adjusted to once every two weeks), and the currently identified risk drivers are explicitly marked in the early warning information. The risk drivers are marked in a structured list format, with each entry including the factor name, a comparison of the current measured value with the normal reference range, and a ranking of the factor's causal contribution to the current risk score, enabling doctors to quickly identify indicators requiring focused attention. Early warning information for patients at medium risk is simultaneously pushed to both the patient's and doctor's end. The doctor's end also includes trend charts showing the dynamic changes of the main risk drivers within a recent time window, helping doctors determine whether the risk is showing a sustained upward trend.

[0144] For high-risk patients, the early warning strategy has been upgraded to emergency warnings. Emergency warning messages are pushed through a high-priority channel to ensure that doctors receive and confirm them within a preset timeframe. In addition to complete annotations of the risk drivers, high-risk warning information also includes intervention suggestions generated from causal inferences, displaying each suggestion in relation to its corresponding risk driver, forming a complete "problem-cause-suggestion" chain. For high-risk patients, the system also triggers an active outbound call or message confirmation mechanism. If the patient does not read the warning information within the specified time, the responsible doctor is automatically notified for manual follow-up, preventing the warning information from becoming invalid due to the patient's failure to view it in time.

[0145] The risk interpretation text is generated using a combination of templated natural language generation and dynamic data imputation. Standardized interpretation templates are pre-designed for common risk driver categories (such as abnormal inflammatory markers, decreased medication adherence, abnormal environmental exposure, and irregular lifestyle behaviors). These templates include dynamic slots for filling in the patient's actual values, baseline values, and descriptions of trends. For example, when C-reactive protein (CRP) levels are identified as a major risk driver, the interpretation text automatically generates a statement with the following structure: explaining the increase in the current CRP value compared to the patient's baseline value; indicating the association mechanism between the elevated CRP and increased recurrence risk (using concise, non-technical language, such as "elevated inflammation levels indicate worsening joint inflammation activity"); and stating the CRP's contribution ranking among all current risk drivers. The risk interpretation text differentiates between patient and physician versions in terms of language style: the patient version uses everyday language, avoiding excessive technical jargon; the physician version retains the clinical indicator names, quantitative values, and statistical confidence information to facilitate professional judgment by physicians.

[0146] The process of translating intervention recommendations into operational instructions requires mapping abstract causal intervention targets into concrete actions that patients can directly execute in their daily lives. For interventions related to medication adherence, the operational instructions clearly state the target drug name, the recommended time of administration (e.g., after breakfast daily), the single dose, and the minimum continuous period of administration, along with reminder setting suggestions (e.g., setting a daily medication reminder in the phone calendar). For interventions related to lifestyle behaviors, the operational instructions are refined to the type of exercise (e.g., low-intensity aerobic exercise), the duration of each session (e.g., 30 minutes each time), the weekly frequency (e.g., 3 times a week), and a method for self-judging exercise intensity (e.g., if one can talk normally after exercise, the intensity is appropriate). For interventions related to environmental exposure, the operational instructions provide specific avoidance suggestions, such as reducing outdoor activity time when the air quality index exceeds a certain value, or increasing joint warmth measures when the temperature drops sharply. Each operational instruction includes an explanation of the expected effect, presented in the form of a quantified expected risk reduction (this value is derived from the counterfactual risk reduction estimate calculated in the causal inference stage), enabling patients to intuitively understand the expected benefits of adhering to the intervention measures, thereby improving their adherence motivation.

[0147] The final push notifications of actionable alerts are organized according to a unified structure: the first part displays a risk level indicator (distinguished by color coding: green for low risk, orange for medium risk, and red for high risk); the middle part displays explanatory text about the risk; and the last part lists a list of operational instructions. The patient interface is adapted for mobile interaction, allowing patients to mark each instruction as "understood" or "questionable." These markings are synchronized to the doctor's end, allowing doctors to prioritize addressing patient feedback with questions. In addition to receiving alerts consistent with those on the patient's end, the doctor's end also displays a causal contribution distribution chart, a time-series trend chart of each risk driver, and a ranking of the adherence and feasibility scores of intervention recommendations, providing data support for doctors to develop personalized treatment plans. All alert push records and patient viewing status are fully saved to the patient's file for subsequent follow-up reviews and model feedback optimization.

[0148] A second aspect of the present invention provides a remote intelligent management system for rheumatoid and immunological patients, comprising:

[0149] The prediction and evaluation unit is used to acquire clinical indicator data, environmental monitoring data and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain the disease activity prediction result.

[0150] A risk correction unit is used to input the disease activity prediction result and the multivariate feature vector into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships.

[0151] The causal analysis unit is used to perform counterfactual replacement on the target feature components in the multivariate feature vector using a causal inference algorithm, input the replaced counterfactual feature vector into the relapse risk assessment network to obtain a counterfactual risk prediction value, calculate the difference between the relapse risk assessment result and the counterfactual risk prediction value as the causal contribution degree, screen risk driving factors according to the causal contribution degree distribution, and generate intervention suggestions that quantify the expected risk reduction.

[0152] The early warning output unit is used to generate actionable early warning information based on the risk driving factors and the intervention recommendations.

[0153] A third aspect of the present invention provides an electronic device, comprising:

[0154] processor;

[0155] Memory used to store processor-executable instructions;

[0156] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0158] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

Claims

1. A remote intelligent management method for rheumatoid and immunological patients, characterized in that, include: Acquire clinical indicator data, environmental monitoring data, and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain disease activity prediction results; The disease activity prediction results and the multivariate feature vectors are input into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships. The target feature components in the multivariate feature vector are replaced by counterfactual substitution using a causal inference algorithm. The replaced counterfactual feature vector is then input into the relapse risk assessment network to obtain a counterfactual risk prediction value. The difference between the relapse risk assessment result and the counterfactual risk prediction value is calculated as the causal contribution. Risk driving factors are screened based on the causal contribution distribution, and intervention suggestions that quantify the expected risk reduction are generated. Actionable early warning information is generated based on the aforementioned risk drivers and intervention recommendations.

2. The method according to claim 1, characterized in that, The multivariate feature vectors are input into a deep learning network to obtain disease activity prediction results, including: The deep learning network is trained using a combined loss function, which includes a disease activity prediction loss term, a temporal smoothing constraint term, and a high activity sensitivity enhancement term. The disease activity prediction loss term measures the deviation between the predicted disease activity and the labeled true disease activity. The temporal smoothing constraint term calculates the magnitude of change in predicted disease activity for the same patient within adjacent time windows and penalizes changes exceeding a physiologically reasonable threshold. The high activity sensitivity enhancement term assigns a higher loss weight to the prediction error of samples labeled as high-activity states compared to regular samples. Output the disease activity prediction results generated by the deep learning network after training is completed.

3. The method according to claim 1, characterized in that, The recurrence risk assessment network includes a collaborative constraint layer, which performs consistency correction on the recurrence risk assessment results based on medical prior constraints, including: The disease activity prediction results and the multivariate feature vectors are input into the relapse risk assessment network to obtain the initial relapse risk assessment results; The collaborative constraint layer constructs hierarchical medical prior constraint relationships, including strong constraint rules and weak constraint rules. The strong constraint rules set unbreakable logical boundary conditions between disease activity state and recurrence risk state, while the weak constraint rules set statistical correlation boundary conditions between disease activity state and recurrence risk state. Calculate the first confidence level of the disease activity prediction result and the second confidence level of the initial recurrence risk assessment result; identify the constraint rules that violate the hierarchical medical prior constraint relationship; determine the correction priority according to the type of the violated constraint rule and prioritize satisfying the strong constraint rule; calculate the reciprocal of the first confidence level as the disease activity correction weight and the reciprocal of the second confidence level as the recurrence risk correction weight; construct a constraint optimization objective function based on the disease activity correction weight and the recurrence risk correction weight; minimize the weighted sum of the disease activity correction amount and the recurrence risk correction amount under the boundary conditions of satisfying the violated constraint rules; solve the constraint optimization objective function to obtain the corrected recurrence risk assessment result.

4. The method according to claim 1, characterized in that, The training of the relapse risk assessment network includes: Basic prediction training phase: The relapse risk assessment network is trained using fully labeled training samples, which include disease activity labels, multivariate feature vector labels, and ground truth labels of relapse risk confirmed by long-term follow-up. Constraint-guided pseudo-label generation stage: For incomplete labeled samples, the network output trained on the basic prediction is constrained and corrected through the collaborative constraint layer to generate constraint-guided pseudo-label recurrence risk values; Constraint-aware joint training phase: Construct a joint training sample set, which includes fully labeled training samples and incompletely labeled samples carrying constraint-guided pseudo-label recurrence risk values. Freeze the parameters of the feature extraction layer of the recurrence risk assessment network. Train the risk assessment layer and the collaborative constraint layer. Calculate the loss for fully labeled training samples using the ground truth recurrence risk annotation, and calculate the loss for incompletely labeled samples using constraint-guided pseudo-label recurrence risk values.

5. The method according to claim 1, characterized in that, The causal inference algorithm is used to perform counterfactual replacement on the target feature components in the multivariate feature vector and calculate the causal contribution, including: The feature components in the multivariate feature vector are divided into operable features and non-operable features; The modifiable features are subjected to single-feature counterfactual replacement and multi-feature combined counterfactual replacement. The single-feature counterfactual replacement replaces a single modifiable feature with a target improvement value, and the multi-feature combined counterfactual replacement jointly replaces multiple modifiable features with a combined target value, thereby obtaining single-feature counterfactual risk prediction values ​​and combined counterfactual risk prediction values, respectively. The difference between the recurrence risk assessment result and the single-feature counterfactual risk prediction value is calculated as the single-feature causal contribution; the difference between the recurrence risk assessment result and the combined counterfactual risk prediction value is calculated, and the sum of the causal contributions of each single feature participating in the combination is subtracted from this difference to obtain the feature interaction causal contribution. The risk attribution degree is calculated by counterfactual substitution of the aforementioned non-interventionist features; Risk drivers are screened based on single-feature causal contribution, feature interaction causal contribution, and risk attribution, and then the interventionable causal targets and non-interventionable risk attribution factors are distinguished.

6. The method according to claim 5, characterized in that, Generate intervention recommendations that quantify the expected reduction in risk, including: For each operable causal target, multiple candidate intervention intensities are set. The expected risk reduction and intervention target value corresponding to different intervention intensities are calculated through counterfactual inference. The minimum intervention intensity that reduces the expected risk to the target threshold is determined as the personalized intervention amount for that causal target. A multi-objective optimization framework is constructed, wherein the optimization objectives of the multi-objective optimization framework include maximizing the reduction of total expected risk, minimizing the cost of intervention implementation, and maximizing patient compliance feasibility; The multi-objective optimization framework is used to solve both single-causal target intervention schemes and multi-causal target combination intervention schemes to obtain the Pareto-optimal set of intervention schemes. Output the Pareto optimal intervention set, where each intervention set includes the personalized intervention amount, expected risk reduction, and implementation cost estimate.

7. The method according to claim 1, characterized in that, Based on the aforementioned risk drivers and intervention recommendations, actionable early warning information is generated, including: The risk level is determined based on the recurrence risk assessment results, and the risk level is divided into low risk, medium risk, and high risk; Differentiated early warning strategies are set for different risk levels. Regular follow-up reminders are generated for low-risk patients, key monitoring early warnings are generated for medium-risk patients and the risk driving factors are marked, and emergency early warnings are generated for high-risk patients and the intervention recommendations are associated with them. The risk drivers are translated into patient-understandable risk explanation texts that explain the reasons for the current increased risk of relapse. The intervention recommendations are translated into specific, executable operational instructions, which include the specific execution method and frequency of the intervention measures. Generate actionable warning information, which includes the risk level, the risk explanation text, and the operation instructions.

8. A remote intelligent management system for rheumatoid and immunological patients, used to implement the method as described in any one of claims 1-7, characterized in that, include: The prediction and evaluation unit is used to acquire clinical indicator data, environmental monitoring data and lifestyle behavior data of the target patient, construct a multivariate feature vector, and input the multivariate feature vector into a deep learning network to obtain the disease activity prediction result. A risk correction unit is used to input the disease activity prediction result and the multivariate feature vector into a relapse risk assessment network to obtain a relapse risk assessment result. The relapse risk assessment network includes a collaborative constraint layer, which performs consistency correction on the relapse risk assessment result based on medical prior constraint relationships. The causal analysis unit is used to perform counterfactual replacement on the target feature components in the multivariate feature vector using a causal inference algorithm, input the replaced counterfactual feature vector into the relapse risk assessment network to obtain a counterfactual risk prediction value, calculate the difference between the relapse risk assessment result and the counterfactual risk prediction value as the causal contribution degree, screen risk driving factors according to the causal contribution degree distribution, and generate intervention suggestions that quantify the expected risk reduction. The early warning output unit is used to generate actionable early warning information based on the risk driving factors and the intervention recommendations.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images