Systemic lupus erythematosus interaction management method based on patient sign data modeling

Abstract

The invention discloses a systemic lupus erythematosus interaction management method based on patient sign data modeling, and particularly relates to the field of lupus erythematosus interaction management, which comprises the following steps: collecting basic data and sign data of a patient, generating a standardized basic data set and an integrity sign data set, feature combination, feature correlation detection and factor extraction are executed, a potential feature space representing patient features is constructed, and static health behavior feature modeling is achieved; the method comprises the following steps: acquiring nurse-patient interaction behavior data, generating a preliminary interaction data set, executing interaction feature extraction and activeness judgment, and generating a standardized interaction feature matrix; a multi-dimensional feature vector and dynamic health behavior link is constructed based on patient sign data and interactive behavior features, and time sequence prediction and response feedback closed-loop optimization are combined, so that dynamic management and individualized intervention of active health behaviors of systemic lupus erythematosus patients are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of systemic lupus erythematosus interactive management, and more particularly, to a systemic lupus erythematosus interactive management method based on patient vital sign data modeling. Background Art

[0002] In the current medical management system, patients with systemic lupus erythematosus (SLE) require long-term use of hormones and immunosuppressive drugs, which often lead to decreased muscle mass and decreased strength, and can easily induce complications such as sarcopenia. Although current routine health management relies heavily on health education, brochure distribution, and telephone follow-up, these one-way, passive information delivery methods generally suffer from low patient engagement, short-lived intervention effects, and poor compliance. Furthermore, it is difficult to achieve precise management based on patient characteristics. This is particularly true for patients with SLE and sarcopenia, who lack effective dynamic behavioral regulation and individualized health interventions. This results in suboptimal clinical disease control and functional recovery, severely hindering the improvement of chronic disease management. Therefore, existing technologies urgently need a systemic lupus erythematosus interactive management method that can be based on patient vital sign data modeling, integrate dynamic interaction mechanisms, and accurately promote the patient's active health behavior transformation, so as to achieve continuous optimization of health outcomes and improve the efficiency of medical resource utilization. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an interactive management method for systemic lupus erythematosus based on patient vital sign data modeling. By constructing a multidimensional feature vector and a dynamic health behavior link based on patient vital sign data and interactive behavior characteristics, and combining time series prediction and response feedback closed-loop optimization, dynamic management and individualized intervention of the active health behavior of systemic lupus erythematosus patients are achieved to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an interactive management method for systemic lupus erythematosus based on patient vital sign data modeling, comprising: S1. Collect the patient's basic data and vital sign data, generate a standardized basic data set and a complete vital sign data set, perform feature combination, feature correlation detection and factor extraction, construct a potential feature space that represents the patient's characteristics, and realize static health behavior feature modeling; S2. Collect nurse-patient interaction behavior data, generate a preliminary interaction data set, perform interaction feature extraction and activity judgment, generate a standardized interaction feature matrix, generate a patient-interaction feature matching matrix based on feature matching, build a health behavior link, generate a health behavior conversion link diagram based on the link structure, and realize personalized interaction behavior modeling; S3. Collect vital sign follow-up data and interactive follow-up data, integrate the static feature information represented by the health behavior conversion link graph, generate a joint time series feature set, perform sequence prediction on the joint time series feature set, and generate an individualized health risk sequence; S4. Generate intervention strategies based on individualized health risk sequences, adjust health management pathways based on patient response behavior data, and feed back updated results to the health behavior chain diagram.

[0005] In a preferred embodiment, S1 further includes: collecting basic data of the patient, including demographic characteristics, social support characteristics, and disease history characteristics, and normalizing each characteristic to a standardized basic data set through a standardization method; collecting vital sign data of the patient, including bioelectrical impedance characteristics, serum biomarker characteristics, and exercise capacity characteristics, eliminating outliers and filling missing values in the vital sign data to form a complete vital sign data set; wherein the standardization method includes but is not limited to the Z-score standardization method, subtracting the mean from the original value and dividing it by the standard deviation to normalize it to a standardized basic data set; the vital sign data includes identifying and removing outliers through the IQR (interquartile range) method, and filling missing values using mean interpolation or regression interpolation methods to ensure data integrity; Perform feature combination on the standardized basic data set and the completeness vital sign data set, construct a multidimensional feature vector based on the feature correspondence, apply a feature correlation detection method to the multidimensional feature vector, calculate the Pearson correlation coefficient between pairs of multidimensional feature vectors, and determine whether there is a feature pair whose absolute value of the Pearson correlation coefficient exceeds a preset correlation coefficient threshold. If so, perform feature dimensionality reduction processing; if not, retain the original multidimensional feature vector; The retained multidimensional feature vectors are standardized to generate a set of standardized multidimensional feature vectors with unified dimensions. The standardized multidimensional feature vector set is used as input to perform covariance matrix calculation to obtain a feature covariance matrix, which is used to characterize the coordinated change relationship between the various features. Then, based on the feature covariance matrix, a factor extraction method is applied to perform principal factor analysis to extract intrinsic motivation factors, cognitive evaluation factors, and emotional response factors, generate a set of latent variable vectors, and determine whether the cumulative explained variance reaches a preset proportion threshold. If not, the number of principal factors is increased and re-extraction is performed. If so, the number of factors is fixed to determine the latent feature space. It should be noted that demographic characteristics include age (continuous value, in years), sex (binary variable, male / female), height (continuous value, in centimeters), weight (continuous value, in kilograms), and marital status (categorical variable, married / unmarried / divorced / widowed); social support characteristics include years of education (continuous value, in years), occupational category (multi-categorical variable, worker / staff / self-employed / retired, etc.), monthly household income level (continuous value, in RMB), and living status (categorical variable, living alone / living with family); disease history characteristics include age of onset of systemic lupus erythematosus (SLE) (continuous value, in years), disease duration (continuous value, in years), previous main treatment regimen (categorical variable, hormone therapy / immunosuppressants / biological agents, etc.), and comorbidities (binary variable, yes / no). In actual applications, demographic characteristics and social support characteristics are collected through electronic questionnaires, and disease history characteristics are extracted from the hospital electronic medical record system and verified through patient interviews. In the interactive management of systemic lupus erythematosus, bioelectrical impedance characteristics include skeletal muscle index (continuous value, unit: kg / m 2 ) and phase angle (continuous value, unit: degree), which are measured non-invasively by multi-frequency bioelectrical impedance analyzer (BIA); serum biomarker characteristics include C-reactive protein (CRP, continuous value, unit: mg / L), interleukin-6 (IL-6, continuous value, unit: pg / mL), tumor necrosis factor α (TNF-α, continuous value, unit: pg / mL) and serum myostatin (myostatin, continuous value, unit: ng / mL), which are obtained by venous blood collection and enzyme-linked immunosorbent assay (ELISA) detection; exercise capacity characteristics include grip strength (continuous value, unit: kg) and 6-minute walk distance (continuous value, unit: meter), which are measured by handgrip dynamometer and standardized 6-minute walk test, respectively; all physical sign data are manually reviewed and outlier checked after collection to support dynamic assessment and modeling of individual physical condition and inflammation level in interactive management of lupus erythematosus; Among them, the intrinsic motivation factor refers to the potential driving force for patients to maintain healthy behaviors based on their own will. When extracting, it is done by weighted aggregation of the eigenvectors related to autonomous participation and goal setting in the feature covariance matrix; the cognitive evaluation factor refers to the patient's degree of disease cognition and understanding of health information. When extracting, it is done by screening the eigenvectors related to disease knowledge and risk perception and performing local principal component decomposition; the emotional response factor refers to the degree of influence of the patient's emotional state on the execution of health behaviors. When extracting, it is done by identifying the eigenvectors related to emotional fluctuations and social support and performing sparse factor rotation; the factor extraction overall takes the feature covariance matrix as input, and through weighted local eigendecomposition and sparse factor rotation joint modeling, a set of independent principal factors representing the potential behavioral mechanism is extracted.

[0006] In a preferred embodiment, S2 further includes: collecting nurse-patient interaction behavior data, the nurse-patient interaction behavior data including health information interaction records, emotional support interaction records, decision-making participation interaction records, and guidance interaction records, and categorizing them by interaction type to form a preliminary interaction data set; Apply the interaction feature extraction method to the preliminary interaction data set, calculate the information interaction frequency, support feedback density, decision-making participation rate and professional guidance coverage rate respectively, and generate a hierarchical interaction feature vector; judge whether each feature in the hierarchical interaction feature vector meets the preset behavioral activity standard, which includes the information interaction frequency exceeding the baseline frequency, the support feedback density exceeding the expected density, the decision-making participation rate exceeding the group median, and the professional guidance coverage rate exceeding the preset ratio. If any of the behavioral activity standards is not met, it is judged that the interaction activity is insufficient and a supplementary interaction strategy is returned; if it is met, continue to perform normalization processing on the hierarchical interaction feature vector to unify the feature scale to make the numerical values of different features comparable, and generate a standardized interaction feature matrix; Perform feature matching on the standardized interaction feature matrix and the standardized multidimensional feature vector set, calculate cosine similarity through feature space mapping, and generate a patient-interaction feature matching matrix; determine whether each matching degree in the patient-interaction feature matching matrix is higher than a preset matching threshold; if it is lower than the preset matching threshold, adjust the interaction content or frequency and re-execute feature extraction and matching; if it is higher than the preset matching threshold, solidify the current feature matching relationship; A health behavior link is constructed based on the patient-interaction feature matching matrix. The link nodes of the health behavior link represent the feature vector state, and the link edge weight represents the interaction intensity, ultimately forming a health behavior conversion link graph.

[0007] In a preferred embodiment, S3 further includes: collecting physical sign follow-up data, including bioelectrical impedance follow-up indicators, serum inflammatory factor follow-up indicators, and exercise capacity follow-up indicators, to form a physical sign time series data set; collecting interaction follow-up data, including learning frequency follow-up sequence, message interaction follow-up sequence, and offline activity participation follow-up sequence, to form an interaction behavior time series data set; The vital sign time series data set and the interactive behavior time series data set are synchronized and aligned based on timestamps, and the static feature information represented by the health behavior conversion link diagram is integrated to generate a joint time series feature set; it is determined whether there are missing time segments in the synchronized joint time series feature set. If so, time series interpolation processing is performed, including linear interpolation or Lagrange interpolation. If not, the original sequence structure is retained; The joint time series feature set is input into a sequence prediction model, which includes a multi-objective regression model and a long short-term memory network (LSTM) structure to perform multi-objective prediction of disease activity, muscle mass, inflammation level, and patient satisfaction, respectively. Calculate the error indicators between the prediction results of the sequence prediction model and the actual recorded follow-up results. Error indicators include mean square error (MSE) and mean absolute error (MAE). Determine whether the error indicators are lower than the preset error tolerance threshold. If so, output the prediction results. If higher, adjust the model parameters and retrain. Generate a health risk score curve based on the prediction results. The health risk score includes a disease aggravation risk score, a malnutrition risk score, and a compliance decline risk score. Based on different score weights, a personalized health risk sequence is formed. Determine whether there are abnormal fluctuation points in the health risk score curve. The criteria for determining abnormal fluctuation points include whether the rate of change exceeds the preset change rate threshold and the risk level increases to the preset warning level. If an abnormal fluctuation point exists, the risk warning strategy is generated. If not, the individualized health risk sequence continues to be output; It should be noted that in the interactive management of systemic lupus erythematosus, the bioelectrical impedance follow-up indicators include skeletal muscle index (continuous value, unit: kg / m 2) and phase angle (continuous value, unit: degree), which were measured regularly by multi-frequency bioelectrical impedance analyzer; serum inflammatory factor follow-up indicators included C-reactive protein (CRP), interleukin-6 (IL-6), and tumor necrosis factor α (TNF-α), all of which were continuous values (unit: mg / L or pg / mL), obtained by venous blood collection and enzyme-linked immunosorbent assay (ELISA); exercise capacity follow-up indicators included grip strength (continuous value, unit: kg) and 6-minute walk distance (continuous value, unit: meter), which were measured by handgrip dynamometer test and standardized 6-minute walk test; interactive follow-up data were collected, and the learning frequency follow-up sequence refers to the number of times patients learned health education content (continuous value, unit: times / week), which was automatically recorded in the interactive platform log; the message interaction follow-up sequence refers to the number of messages patients posted on the interactive platform (continuous value, unit: messages / week), which was extracted from the platform message module log; the offline activity participation follow-up sequence refers to the number of times patients signed in to participate in offline health management activities (continuous value, unit: times / month), which was collected through the sign-in records of the activity management system; In addition, in the interactive management of systemic lupus erythematosus, static feature collection refers to the basic data and vital sign data obtained from the patient at a single time, which is used to construct the feature vector space and support initial feature modeling; dynamic follow-up collection is based on the same indicators, and forms a time series data set through periodic collection to capture the changing trends of patients' vital signs and the evolution trajectory of interactive behaviors, supporting health outcome prediction and risk assessment. The two types of data are consistent in the collection objects, but have essential differences in data attributes and application purposes. Static features are used for static modeling, and dynamic sequences are used for dynamic analysis and predictive modeling. In the calculation of the multi-objective regression model, the joint time series feature set is first grouped into feature blocks. Four sets of feature subsets are constructed based on the dynamic mutual information distribution of time series between samples according to the historical change trends of disease activity, muscle mass, inflammation level and patient satisfaction. Principal component dimensionality reduction is performed on each set of feature subsets, and a joint low-dimensional feature representation is generated through cross-feature coupling mapping. The multi-objective regression model uses the low-dimensional feature representation as input and adopts a hierarchical task weight mechanism. The initial weight is calculated based on the dynamic mutual information matrix between tasks, reflecting the strength of the historical correlation between each prediction task. Task groups with strong correlation are given higher coupling weights. The regression loss function is constructed as a target weighted square loss. A sparse penalty for the inter-feature structure is introduced in the loss term. The L1 regularization constraint is used to prevent overfitting of low-contribution features, thereby improving the coupling adaptability and generalization performance of the multi-objective prediction of disease activity, muscle mass, inflammation level and patient satisfaction. In LSTM structure calculations, the combined time series feature set is input into the long short-term memory network (LSTM), and a local weighted attention module is embedded before the LSTM input gate. The local weight coefficient is dynamically adjusted based on the kernel density estimation result of the local density of the feature time, improving the responsiveness to features in high-density variation intervals. During the hidden state update process, a task separation gating mechanism is introduced to bind the prediction paths of disease activity, muscle mass, inflammation level, and patient satisfaction to their own dedicated hidden unit subspaces. Each subspace is updated independently to avoid conflicts in multi-task hidden layers. A time-step adaptive forgetting factor is embedded in the cell state transition. The forgetting factor is dynamically corrected based on the fluctuation amplitude of the previous time window to enhance the model's time sensitivity to non-stationary feature sequences. Finally, each prediction task extracts the final state features from the corresponding hidden subspace, enters the task-specific fully connected layer, and outputs the prediction results of disease activity, muscle mass, inflammation level, and patient satisfaction. For each prediction task, the square of the residual between the prediction result and the actual follow-up result is calculated separately to form a task distribution residual set; based on the residual set, the task information entropy matrix is constructed, and the information entropy reflects the degree of uncertainty in the error distribution of each task; the MSE indicator uses the inverse of the information entropy as the weight, averages the square of the residual of each task, and improves the contribution of high-confidence tasks in the overall error index; the MAE indicator introduces density weighting in the time dimension, and weights the time step according to the kernel density estimate of the follow-up time point distribution. At the same time, the inverse weight of the fluctuation amplitude is introduced in the task dimension, and the standard deviation of the change rate of the fluctuation amplitude is calculated through a sliding window; finally, the time-task joint weighted absolute error is obtained by gradually accumulating the absolute values of the density-weighted and fluctuation-weighted residuals, ensuring that the error index is dynamically adaptive under the dual scales of time and task.

[0008] In a preferred embodiment, S4 further includes: executing intervention strategy generation based on the individualized health risk sequence, the intervention strategy generation is based on a preset rule engine, matching dietary intervention suggestions, exercise intervention suggestions, and medication compliance suggestions to generate a preliminary intervention suggestion set; The initial set of intervention suggestions is pushed to the patient interaction platform, which records the patient's response behavior data in real time. The patient response behavior data includes the number of times the suggestion is read, the feedback on the implementation of the behavior, and the number of self-consultations. The response effect evaluation is performed based on the patient's response behavior data to generate a set of behavioral response effect indicators. The evaluation indicators of the response effect evaluation include the behavior implementation rate, the suggestion adoption rate, and the degree of self-consultation enthusiasm. Determine whether the behavioral response effect indicator set reaches the preset evaluation threshold. If not, analyze the behavioral blocking point based on the response behavior data, adjust the intervention suggestion set, and push it again. If it reaches the threshold, retain the current intervention strategy. Based on the changing trend of the behavioral response effect indicators, the interaction trigger frequency and intervention content update cycle are adjusted to form a health management path based on behavioral changes. The health management path is then fed back to the health behavior link graph, and the link node status and edge weights are updated in the health behavior link graph. Determine the changing trend of the health behavior link graph after the patient characteristics are updated. If a downward trend is detected, re-execute feature modeling and link construction based on the updated feature data. If an upward trend or unchanged trend is detected, maintain the existing link structure.

[0009] In a preferred embodiment, in S1, the Pearson correlation coefficient is calculated by performing zero-mean centering on the multidimensional feature vectors, calculating the local covariance of each pair of multidimensional feature vectors by setting a local weighted kernel function, and adjusting the covariance weight according to the inverse proportion of the feature distribution density. Then, the weighted square root of the local variance of each feature vector is used as a normalization factor, and the correlation coefficient value is obtained by taking the ratio of the local weighted covariance to the normalization factor; The covariance matrix is calculated by first performing zero-mean centering on any two sets of multidimensional eigenvectors in the standardized multidimensional eigenvector set in a local density weighted manner, and then calculating the weighted covariance. The covariance value is corrected by the adjustment factor of the local sample density. The weighted covariance results of all paired multidimensional eigenvectors are combined to form a characteristic covariance matrix. The elements of the characteristic covariance matrix reflect the degree of coordinated variation of the eigenvectors under the local structure.

[0010] Technical effects and advantages of the present invention: 1. By dynamically building a health behavior link based on physical sign data and interactive feature modeling, we can accurately identify the individual characteristics of patients with systemic lupus erythematosus and sarcopenia and optimize dynamic health management paths, thereby improving intervention targeting and compliance; 2. By integrating standardized basic data with complete vital sign data, we perform weighted feature dimensionality reduction and factor extraction to improve the information density and representation capabilities of multidimensional feature vectors and enhance the accuracy of extracting potential features of patients' health behaviors; 3. Dynamically update the evolution trajectory of patients' interactive behaviors through local density-weighted interaction feature space matching, achieve fine-grained monitoring of interactive behavior status and personalized recommendation of interaction strategies, and improve the adaptability of interaction management; 4. By combining time series features into a multi-objective regression model and a long-short-term memory network, disease activity, muscle mass, inflammation levels, and patient satisfaction can be dynamically predicted, improving the accuracy of identifying health risk trends and the timeliness of early warnings. 5. Through the error evaluation mechanism of information entropy inverse weight and time density weighting, the multi-task prediction error distribution is dynamically optimized to improve the stability and robustness of the sequence prediction model under different time series densities and task complexities. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0013] Refer to the instruction manual Figure 1 An interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to an embodiment of the present invention includes: S1. Collect the patient's basic data and vital sign data, generate a standardized basic data set and a complete vital sign data set, perform feature combination, feature correlation detection and factor extraction, construct a potential feature space that represents the patient's characteristics, and realize static health behavior feature modeling; S2. Collect nurse-patient interaction behavior data, generate a preliminary interaction data set, perform interaction feature extraction and activity judgment, generate a standardized interaction feature matrix, generate a patient-interaction feature matching matrix based on feature matching, build a health behavior link, generate a health behavior conversion link diagram based on the link structure, and realize personalized interaction behavior modeling; S3. Collect vital sign follow-up data and interactive follow-up data, integrate the static feature information represented by the health behavior conversion link graph, generate a joint time series feature set, perform sequence prediction on the joint time series feature set, and generate an individualized health risk sequence; S4. Generate intervention strategies based on individualized health risk sequences, adjust health management pathways based on patient response behavior data, and feed back updated results to the health behavior chain diagram.

[0014] S1 also includes: collecting basic data of patients, including demographic characteristics, social support characteristics and disease history characteristics, and normalizing each characteristic to a standardized basic data set through standardization methods; collecting vital sign data of patients, including bioelectrical impedance characteristics, serum biomarker characteristics and exercise capacity characteristics, eliminating outliers and filling missing values in the vital sign data to form a complete vital sign data set; the standardization method includes but is not limited to the Z-score standardization method, which subtracts the mean from the original value and divides it by the standard deviation to normalize it to a standardized basic data set; the vital sign data includes identifying and eliminating outliers through the IQR (interquartile range) method, and filling missing values with mean interpolation or regression interpolation to ensure data integrity; Perform feature combination on the standardized basic data set and the completeness vital sign data set, construct a multidimensional feature vector based on the feature correspondence, apply a feature correlation detection method to the multidimensional feature vector, calculate the Pearson correlation coefficient between pairs of multidimensional feature vectors, and determine whether there is a feature pair whose absolute value of the Pearson correlation coefficient exceeds a preset correlation coefficient threshold. If so, perform feature dimensionality reduction processing; if not, retain the original multidimensional feature vector; The retained multidimensional feature vectors are standardized to generate a set of standardized multidimensional feature vectors with unified dimensions. The standardized multidimensional feature vector set is used as input to perform covariance matrix calculation to obtain a feature covariance matrix, which is used to characterize the coordinated change relationship between the various features. Then, based on the feature covariance matrix, a factor extraction method is applied to perform principal factor analysis to extract intrinsic motivation factors, cognitive evaluation factors, and emotional response factors, generate a set of latent variable vectors, and determine whether the cumulative explained variance reaches a preset proportion threshold. If not, the number of principal factors is increased and re-extraction is performed. If so, the number of factors is fixed to determine the latent feature space. It should be noted that demographic characteristics include age (continuous value, in years), sex (binary variable, male / female), height (continuous value, in centimeters), weight (continuous value, in kilograms), and marital status (categorical variable, married / unmarried / divorced / widowed); social support characteristics include years of education (continuous value, in years), occupational category (multi-categorical variable, worker / staff / self-employed / retired, etc.), monthly household income level (continuous value, in RMB), and living status (categorical variable, living alone / living with family); disease history characteristics include age of onset of systemic lupus erythematosus (SLE) (continuous value, in years), disease duration (continuous value, in years), previous main treatment regimen (categorical variable, hormone therapy / immunosuppressants / biological agents, etc.), and comorbidities (binary variable, yes / no). In actual applications, demographic characteristics and social support characteristics are collected through electronic questionnaires, and disease history characteristics are extracted from the hospital electronic medical record system and verified through patient interviews. In the interactive management of systemic lupus erythematosus, bioelectrical impedance characteristics include skeletal muscle index (continuous value, unit: kg / m 2) and phase angle (continuous value, unit: degree), which are measured non-invasively by multi-frequency bioelectrical impedance analyzer (BIA); serum biomarker characteristics include C-reactive protein (CRP, continuous value, unit: mg / L), interleukin-6 (IL-6, continuous value, unit: pg / mL), tumor necrosis factor α (TNF-α, continuous value, unit: pg / mL) and serum myostatin (myostatin, continuous value, unit: ng / mL), which are obtained by venous blood collection and enzyme-linked immunosorbent assay (ELISA) detection; exercise capacity characteristics include grip strength (continuous value, unit: kg) and 6-minute walk distance (continuous value, unit: meter), which are measured by handgrip dynamometer and standardized 6-minute walk test, respectively; all physical sign data are manually reviewed and outlier checked after collection to support dynamic assessment and modeling of individual physical condition and inflammation level in interactive management of lupus erythematosus; Among them, the intrinsic motivation factor refers to the potential driving force for patients to maintain healthy behaviors based on their own will. When extracting, it is done by weighted aggregation of the eigenvectors related to autonomous participation and goal setting in the feature covariance matrix; the cognitive evaluation factor refers to the patient's degree of disease cognition and understanding of health information. When extracting, it is done by screening the eigenvectors related to disease knowledge and risk perception and performing local principal component decomposition; the emotional response factor refers to the degree of influence of the patient's emotional state on the execution of health behaviors. When extracting, it is done by identifying the eigenvectors related to emotional fluctuations and social support and performing sparse factor rotation; the factor extraction overall takes the feature covariance matrix as input, and through weighted local eigendecomposition and sparse factor rotation joint modeling, a set of independent principal factors representing the potential behavioral mechanism is extracted.

[0015] S2 also includes: collecting nurse-patient interaction behavior data, which includes health information interaction records, emotional support interaction records, decision-making participation interaction records, and guidance interaction records, and categorizing them by interaction type to form a preliminary interaction data set; Apply the interaction feature extraction method to the preliminary interaction data set, calculate the information interaction frequency, support feedback density, decision-making participation rate and professional guidance coverage rate respectively, and generate a hierarchical interaction feature vector; judge whether each feature in the hierarchical interaction feature vector meets the preset behavioral activity standard, which includes the information interaction frequency exceeding the baseline frequency, the support feedback density exceeding the expected density, the decision-making participation rate exceeding the group median, and the professional guidance coverage rate exceeding the preset ratio. If any of the behavioral activity standards is not met, it is judged that the interaction activity is insufficient and a supplementary interaction strategy is returned; if it is met, continue to perform normalization processing on the hierarchical interaction feature vector to unify the feature scale to make the numerical values of different features comparable, and generate a standardized interaction feature matrix; Perform feature matching on the standardized interaction feature matrix and the standardized multidimensional feature vector set, calculate cosine similarity through feature space mapping, and generate a patient-interaction feature matching matrix; determine whether each matching degree in the patient-interaction feature matching matrix is higher than a preset matching threshold; if it is lower than the preset matching threshold, adjust the interaction content or frequency and re-execute feature extraction and matching; if it is higher than the preset matching threshold, solidify the current feature matching relationship; A health behavior link is constructed based on the patient-interaction feature matching matrix. The link nodes of the health behavior link represent the feature vector state, and the link edge weight represents the interaction intensity, ultimately forming a health behavior conversion link graph.

[0016] S3 also includes: collecting physical sign follow-up data, including bioelectrical impedance follow-up indicators, serum inflammatory factor follow-up indicators, and exercise capacity follow-up indicators, forming a physical sign time series data set; collecting interactive follow-up data, including learning frequency follow-up series, message interaction follow-up series, and offline activity participation follow-up series, forming an interactive behavior time series data set; The vital sign time series data set and the interactive behavior time series data set are synchronized and aligned based on timestamps, and the static feature information represented by the health behavior conversion link diagram is integrated to generate a joint time series feature set; it is determined whether there are missing time segments in the synchronized joint time series feature set. If so, time series interpolation processing is performed, including linear interpolation or Lagrange interpolation. If not, the original sequence structure is retained; The joint time series feature set is input into a sequence prediction model, which includes a multi-objective regression model and a long short-term memory network (LSTM) structure to perform multi-objective prediction of disease activity, muscle mass, inflammation level, and patient satisfaction, respectively. Calculate the error indicators between the prediction results of the sequence prediction model and the actual recorded follow-up results. Error indicators include mean square error (MSE) and mean absolute error (MAE). Determine whether the error indicators are lower than the preset error tolerance threshold. If so, output the prediction results. If higher, adjust the model parameters and retrain. Generate a health risk score curve based on the prediction results. The health risk score includes a disease aggravation risk score, a malnutrition risk score, and a compliance decline risk score. Based on different score weights, a personalized health risk sequence is formed. Determine whether there are abnormal fluctuation points in the health risk score curve. The criteria for determining abnormal fluctuation points include whether the rate of change exceeds the preset change rate threshold and the risk level increases to the preset warning level. If an abnormal fluctuation point exists, the risk warning strategy is generated. If not, the individualized health risk sequence continues to be output; It should be noted that in the interactive management of systemic lupus erythematosus, the bioelectrical impedance follow-up indicators include skeletal muscle index (continuous value, unit: kg / m 2 ) and phase angle (continuous value, unit: degree), which were measured regularly by multi-frequency bioelectrical impedance analyzer; serum inflammatory factor follow-up indicators included C-reactive protein (CRP), interleukin-6 (IL-6), and tumor necrosis factor α (TNF-α), all of which were continuous values (unit: mg / L or pg / mL), obtained by venous blood collection and enzyme-linked immunosorbent assay (ELISA); exercise capacity follow-up indicators included grip strength (continuous value, unit: kg) and 6-minute walk distance (continuous value, unit: meter), which were measured by handgrip dynamometer test and standardized 6-minute walk test; interactive follow-up data were collected, and the learning frequency follow-up sequence refers to the number of times patients learned health education content (continuous value, unit: times / week), which was automatically recorded in the interactive platform log; the message interaction follow-up sequence refers to the number of messages patients posted on the interactive platform (continuous value, unit: messages / week), which was extracted from the platform message module log; the offline activity participation follow-up sequence refers to the number of times patients signed in to participate in offline health management activities (continuous value, unit: times / month), which was collected through the sign-in records of the activity management system; In addition, in the interactive management of systemic lupus erythematosus, static feature collection refers to the basic data and vital sign data obtained from the patient at a single time, which is used to construct the feature vector space and support initial feature modeling; dynamic follow-up collection is based on the same indicators, and forms a time series data set through periodic collection to capture the changing trends of patients' vital signs and the evolution trajectory of interactive behaviors, supporting health outcome prediction and risk assessment. The two types of data are consistent in the collection objects, but have essential differences in data attributes and application purposes. Static features are used for static modeling, and dynamic sequences are used for dynamic analysis and predictive modeling. In the calculation of the multi-objective regression model, the joint time series feature set is first grouped into feature blocks. Four sets of feature subsets are constructed based on the dynamic mutual information distribution of time series between samples according to the historical change trends of disease activity, muscle mass, inflammation level and patient satisfaction. Principal component dimensionality reduction is performed on each set of feature subsets, and a joint low-dimensional feature representation is generated through cross-feature coupling mapping. The multi-objective regression model uses the low-dimensional feature representation as input and adopts a hierarchical task weight mechanism. The initial weight is calculated based on the dynamic mutual information matrix between tasks, reflecting the strength of the historical correlation between each prediction task. Task groups with strong correlation are given higher coupling weights. The regression loss function is constructed as a target weighted square loss. A sparse penalty for the inter-feature structure is introduced in the loss term. The L1 regularization constraint is used to prevent overfitting of low-contribution features, thereby improving the coupling adaptability and generalization performance of the multi-objective prediction of disease activity, muscle mass, inflammation level and patient satisfaction. In LSTM structure calculations, the combined time series feature set is input into the long short-term memory network (LSTM), and a local weighted attention module is embedded before the LSTM input gate. The local weight coefficient is dynamically adjusted based on the kernel density estimation result of the local density of the feature time, improving the responsiveness to features in high-density variation intervals. During the hidden state update process, a task separation gating mechanism is introduced to bind the prediction paths of disease activity, muscle mass, inflammation level, and patient satisfaction to their own dedicated hidden unit subspaces. Each subspace is updated independently to avoid conflicts in multi-task hidden layers. A time-step adaptive forgetting factor is embedded in the cell state transition. The forgetting factor is dynamically corrected based on the fluctuation amplitude of the previous time window to enhance the model's time sensitivity to non-stationary feature sequences. Finally, each prediction task extracts the final state features from the corresponding hidden subspace, enters the task-specific fully connected layer, and outputs the prediction results of disease activity, muscle mass, inflammation level, and patient satisfaction. For each prediction task, the square of the residual between the prediction result and the actual follow-up result is calculated separately to form a task distribution residual set; based on the residual set, the task information entropy matrix is constructed, and the information entropy reflects the degree of uncertainty in the error distribution of each task; the MSE indicator uses the inverse of the information entropy as the weight, averages the square of the residual of each task, and improves the contribution of high-confidence tasks in the overall error index; the MAE indicator introduces density weighting in the time dimension, and weights the time step according to the kernel density estimate of the follow-up time point distribution. At the same time, the inverse weight of the fluctuation amplitude is introduced in the task dimension, and the standard deviation of the change rate of the fluctuation amplitude is calculated through a sliding window; finally, the time-task joint weighted absolute error is obtained by gradually accumulating the absolute values of the density-weighted and fluctuation-weighted residuals, ensuring that the error index is dynamically adaptive under the dual scales of time and task.

[0017] S4 also includes: executing intervention strategy generation based on individualized health risk sequences. Intervention strategy generation is based on a preset rule engine, matching dietary intervention recommendations, exercise intervention recommendations, and medication compliance recommendations to generate a preliminary set of intervention recommendations; The initial set of intervention suggestions is pushed to the patient interaction platform, which records the patient's response behavior data in real time. The patient response behavior data includes the number of times the suggestion is read, the feedback on the implementation of the behavior, and the number of self-consultations. The response effect evaluation is performed based on the patient's response behavior data to generate a set of behavioral response effect indicators. The evaluation indicators of the response effect evaluation include the behavior implementation rate, the suggestion adoption rate, and the degree of self-consultation enthusiasm. Determine whether the behavioral response effect indicator set reaches the preset evaluation threshold. If not, analyze the behavioral blocking point based on the response behavior data, adjust the intervention suggestion set, and push it again. If it reaches the threshold, retain the current intervention strategy. Based on the changing trend of the behavioral response effect indicators, the interaction trigger frequency and intervention content update cycle are adjusted to form a health management path based on behavioral changes. The health management path is then fed back to the health behavior link graph, and the link node status and edge weights are updated in the health behavior link graph. Determine the changing trend of the health behavior link graph after the patient characteristics are updated. If a downward trend is detected, re-execute feature modeling and link construction based on the updated feature data. If an upward trend or unchanged trend is detected, maintain the existing link structure.

[0018] In S1, the Pearson correlation coefficient is calculated by performing zero-mean centering on the multidimensional feature vectors, and calculating the local covariance of each pair of multidimensional feature vectors by setting a local weighted kernel function. The covariance weight is adjusted according to the inverse proportion of the feature distribution density. Then, the weighted square root of the local variance of each feature vector is used as the normalization factor, and the correlation coefficient value is obtained by taking the ratio of the local weighted covariance to the normalization factor; The covariance matrix is calculated by first performing zero-mean centering on any two sets of multidimensional eigenvectors in the standardized multidimensional eigenvector set in a local density weighted manner, and then calculating the weighted covariance. The covariance value is corrected by the adjustment factor of the local sample density. The weighted covariance results of all paired multidimensional eigenvectors are combined to form a characteristic covariance matrix. The elements of the characteristic covariance matrix reflect the degree of coordinated variation of the eigenvectors under the local structure.

[0019] It should be noted that this program addresses the problems of large individual differences, low compliance, and difficulty in dynamic adjustment in the interactive management of systemic lupus erythematosus, and proposes an interactive management method based on patient vital sign data modeling. The program first establishes a standardized basic data set by collecting the patient's demographic characteristics, social support characteristics, and disease history characteristics. At the same time, it collects bioelectrical impedance characteristics, serum biomarker characteristics, and exercise capacity characteristics to form a complete vital sign data set. The basic data and vital sign data are used to construct a multidimensional feature vector based on the feature correspondence relationship, and the correlation between features is detected by the Pearson correlation coefficient adjusted by the local weighted kernel function to avoid the error caused by the global linear assumption. On this basis, weighted local feature decomposition and sparse factor rotation are applied to extract intrinsic motivation factors, cognitive evaluation factors, and emotional response factors to form a potential feature space that characterizes the patient's potential health behavior characteristics and complete static feature modeling. Subsequently, nurse-patient interaction behavior data were collected and hierarchically categorized according to health information interaction records, emotional support interaction records, decision-making participation interaction records, and guidance interaction records. The interaction characteristics of each layer were calculated and dynamically screened and supplemented based on the behavioral activity standard. After normalization, the interaction characteristics were matched with the static feature vector set. The patient-interaction feature matching matrix was generated based on the cosine similarity calculation of the feature space, and the health behavior link was constructed. The link nodes represented the feature vector status, and the edge weights quantified the interaction strength. Finally, a health behavior conversion link diagram was formed, realizing the coupled expression of static characteristics and dynamic interaction behaviors. During the dynamic follow-up phase, vital sign follow-up data and interaction follow-up data are continuously collected to form a vital sign time series data set and an interaction behavior time series data set, respectively. The two types of time series are synchronously aligned based on timestamps, and the static feature information represented by the health behavior conversion link graph is integrated to generate a joint time series feature set. The sequence feature set is input into a multi-target regression model and a long-short-term memory network structure. The former groups feature subsets through a dynamic mutual information matrix, performs principal component dimensionality reduction and cross-feature coupling, and improves information sharing and task collaboration between multi-target predictions. The latter embeds a local weighted attention module in the input gate, adjusts the weighting coefficient according to the local density of the feature time, introduces a task separation gating mechanism and a time-step adaptive forgetting factor in the hidden state update process, and improves the non-stationary time series modeling capability. The two models output prediction results for disease activity, muscle mass, inflammation level, and patient satisfaction, respectively. The error between the predicted results and the actual follow-up results is calculated using the mean square error (MSE) weighted by the inverse weight of information entropy and the mean absolute error (MAE) weighted by the time-task combination, dynamically adjusting the contribution of each task and time step to the overall error. A health risk score curve is generated based on the predicted results, covering the risk of disease aggravation, malnutrition, and decreased compliance, and comprehensively constructing an individualized health risk sequence. The risk sequence is detected for abnormal fluctuations, and based on the change rate and the preset threshold of the risk level, it is determined whether to trigger an early warning strategy, continuously outputting stable individualized risk prediction results. During the intervention decision-making stage, intervention strategies are generated based on individualized health risk sequences, matching dietary, exercise, and medication compliance recommendations, and pushed to the patient interaction platform. The platform records patient response behavior data in real time, collects reading times, behavioral feedback, and autonomous consultation behaviors, and generates behavioral response effect indicators based on the response effects. If the response effect does not meet the standard, the behavioral blocking point is located, the intervention strategy is adjusted, and the strategy is pushed again, forming a feedback loop. If the response effect meets the standard, the current strategy is solidified, and based on the changing trend of the response effect indicator, the interaction trigger frequency and intervention content update cycle are dynamically adjusted. Finally, the changes in the health management path are fed back to the health behavior link diagram, and the link node status and edge weights are updated in real time, forming an adaptive optimization closed loop of interactive management. The original intention of designing this program is to address the real problems of strong individual characteristics of systemic lupus erythematosus patients, dynamic disease progression and poor compliance in interactive behaviors. This program aims to build an interactive management system based on patient vital sign data modeling, integrating static features and dynamic interactive behaviors, and based on time series prediction and response feedback closed-loop optimization. This system captures basic states through static feature modeling, monitors changing trends through dynamic time series, provides accurate warnings through sequence prediction and anomaly detection, and implements strategy adjustments and path optimization through interactive feedback, systematically improving patient health outcomes and interactive management effectiveness, thereby effectively breaking through the static, one-way, passive and extensive problems of traditional health management models and meeting the long-term, dynamic and personalized health management needs of lupus patients.

[0020] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An interactive management method for systemic lupus erythematosus based on patient vital sign data modeling, including, characterized by: S1. Collect the patient's basic data and vital sign data, generate a standardized basic data set and a complete vital sign data set, perform feature combination, feature correlation detection and factor extraction, construct a potential feature space that represents the patient's characteristics, and realize static health behavior feature modeling; S2. Collect nurse-patient interaction behavior data, generate a preliminary interaction data set, perform interaction feature extraction and activity judgment, generate a standardized interaction feature matrix, generate a patient-interaction feature matching matrix based on feature matching, build a health behavior link, generate a health behavior conversion link diagram based on the link structure, and realize personalized interaction behavior modeling; S3. Collect vital sign follow-up data and interactive follow-up data, integrate the static feature information represented by the health behavior conversion link graph, generate a joint time series feature set, perform sequence prediction on the joint time series feature set, and generate an individualized health risk sequence; S4. Generate intervention strategies based on individualized health risk sequences, adjust health management pathways based on patient response behavior data, and feed back updated results to the health behavior chain diagram.

2. The interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to claim 1, characterized in that: S1 also includes: collecting basic data of patients, including demographic characteristics, social support characteristics, and disease history characteristics, and normalizing each characteristic to a standardized basic data set through standardization methods; collecting vital sign data of patients, including bioelectrical impedance characteristics, serum biomarker characteristics, and exercise capacity characteristics, eliminating outliers and filling missing values in the vital sign data to form a complete vital sign data set; Perform feature combination on the standardized basic data set and the completeness vital sign data set, construct a multidimensional feature vector based on the feature correspondence, apply a feature correlation detection method to the multidimensional feature vector, calculate the Pearson correlation coefficient between pairs of multidimensional feature vectors, and determine whether there is a feature pair whose absolute value of the Pearson correlation coefficient exceeds a preset correlation coefficient threshold. If so, perform feature dimensionality reduction processing; if not, retain the original multidimensional feature vector; The retained multidimensional feature vectors are standardized to generate a set of standardized multidimensional feature vectors with unified dimensions. The covariance matrix is calculated with the standardized multidimensional feature vector set as input to obtain a feature covariance matrix, which is used to characterize the coordinated change relationship between the various features. Then, the factor extraction method is applied based on the feature covariance matrix to perform principal factor analysis, extract intrinsic motivation factors, cognitive evaluation factors, and emotional response factors, generate a set of latent variable vectors, and determine whether the cumulative explained variance reaches the preset proportion threshold. If not, the number of principal factors is increased and re-extracted. If so, the number of factors is fixed to determine the potential feature space.

3. The interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to claim 2, characterized in that: S2 also includes: collecting nurse-patient interaction behavior data, which includes health information interaction records, emotional support interaction records, decision-making participation interaction records, and guidance interaction records, and categorizing them by interaction type to form a preliminary interaction data set; Apply the interaction feature extraction method to the preliminary interaction data set, calculate the information interaction frequency, support feedback density, decision-making participation rate and professional guidance coverage rate, and generate a hierarchical interaction feature vector. Determine whether each feature in the hierarchical interaction feature vector meets the preset behavioral activity standards. The behavioral activity standards include information interaction frequency exceeding the baseline frequency, support feedback density exceeding the expected density, decision-making participation rate exceeding the group median, and professional guidance coverage exceeding the preset ratio. If any of the behavioral activity standards is not met, it is determined that the interaction activity is insufficient and a supplementary interaction strategy is returned. If it is met, the hierarchical interaction feature vector is further normalized to unify the feature scale and generate a standardized interaction feature matrix. Perform feature matching on the standardized interaction feature matrix and the standardized multidimensional feature vector set, calculate cosine similarity through feature space mapping, and generate a patient-interaction feature matching matrix; determine whether each matching degree in the patient-interaction feature matching matrix is higher than a preset matching threshold; if it is lower than the preset matching threshold, adjust the interaction content or frequency and re-execute feature extraction and matching; if it is higher than the preset matching threshold, solidify the current feature matching relationship; A health behavior link is constructed based on the patient-interaction feature matching matrix. The link nodes of the health behavior link represent the feature vector state, and the link edge weight represents the interaction intensity, ultimately forming a health behavior conversion link graph.

4. The interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to claim 3, characterized in that: S3 also includes: collecting physical sign follow-up data, including bioelectrical impedance follow-up indicators, serum inflammatory factor follow-up indicators, and exercise capacity follow-up indicators, forming a physical sign time series data set; collecting interactive follow-up data, including learning frequency follow-up series, message interaction follow-up series, and offline activity participation follow-up series, forming an interactive behavior time series data set; The vital sign time series data set and the interactive behavior time series data set are synchronized and aligned based on timestamps, and the static feature information represented by the health behavior conversion link diagram is integrated to generate a joint time series feature set; it is determined whether there are missing time segments in the synchronized joint time series feature set. If so, time series interpolation processing is performed, including linear interpolation or Lagrange interpolation. If not, the original sequence structure is retained; The joint time series feature set is input into a sequence prediction model, which includes a multi-objective regression model and a long short-term memory network structure to perform multi-objective prediction of disease activity, muscle mass, inflammation level, and patient satisfaction respectively. Calculate the error indicators between the prediction results of the sequence prediction model and the actual recorded follow-up results. The error indicators include mean square error and mean absolute error. Determine whether the error indicators are lower than the preset error tolerance threshold. If so, output the prediction results. If higher, adjust the model parameters and retrain. Generate a health risk score curve based on the prediction results. The health risk score includes a disease aggravation risk score, a malnutrition risk score, and a compliance decline risk score. Based on different score weights, a personalized health risk sequence is formed. Determine whether there are abnormal fluctuation points in the health risk score curve. The judgment criteria for abnormal fluctuation points include the change rate exceeding the preset change rate threshold and the risk level rising to the preset warning level. If there are abnormal fluctuation points, the risk warning strategy generation is triggered. If not, the individualized health risk sequence continues to be output.

5. The interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to claim 4, characterized in that: S4 also includes: executing intervention strategy generation based on individualized health risk sequences. Intervention strategy generation is based on a preset rule engine, matching dietary intervention recommendations, exercise intervention recommendations, and medication compliance recommendations to generate a preliminary set of intervention recommendations; The initial set of intervention suggestions is pushed to the patient interaction platform, which records the patient's response behavior data in real time. The patient response behavior data includes the number of times the suggestion is read, the feedback on the implementation of the behavior, and the number of self-consultations. The response effect evaluation is performed based on the patient's response behavior data to generate a set of behavioral response effect indicators. The evaluation indicators of the response effect evaluation include the behavior implementation rate, the suggestion adoption rate, and the degree of self-consultation enthusiasm. Determine whether the behavioral response effect indicator set reaches the preset evaluation threshold. If not, analyze the behavioral blocking point based on the response behavior data, adjust the intervention suggestion set, and push it again. If it reaches the threshold, retain the current intervention strategy. Based on the changing trend of the behavioral response effect indicators, the interaction trigger frequency and intervention content update cycle are adjusted to form a health management path based on behavioral changes. The health management path is then fed back to the health behavior link graph, and the link node status and edge weights are updated in the health behavior link graph. Determine the changing trend of the health behavior link graph after the patient characteristics are updated. If a downward trend is detected, re-execute feature modeling and link construction based on the updated feature data. If an upward trend or unchanged trend is detected, maintain the existing link structure.

6. The interactive management method for systemic lupus erythematosus based on patient vital sign data modeling according to claim 5, characterized in that: In S1, the Pearson correlation coefficient is calculated by performing zero-mean centering on the multidimensional feature vectors, and calculating the local covariance of each pair of multidimensional feature vectors by setting a local weighted kernel function. The covariance weight is adjusted according to the inverse proportion of the feature distribution density. Then, the weighted square root of the local variance of each feature vector is used as the normalization factor, and the correlation coefficient value is obtained by taking the ratio of the local weighted covariance to the normalization factor; The covariance matrix is calculated by first performing zero-mean centering on any two sets of multidimensional eigenvectors in the standardized multidimensional eigenvector set in a local density weighted manner, and then calculating the weighted covariance. The covariance value is corrected by the adjustment factor of the local sample density. The weighted covariance results of all paired multidimensional eigenvectors are combined to form a characteristic covariance matrix. The elements of the characteristic covariance matrix reflect the degree of coordinated variation of the eigenvectors under the local structure.

Images