Dynamic health management method and system based on diabetic nephropathy and storage medium

Abstract

The invention relates to the technical field of health management, in particular to a dynamic health management method and system based on diabetic nephropathy and a storage medium, and the method comprises the steps: building a multi-dimensional tag set system comprising patient, self-management, risk and scheme tag sets; generating a current personalized label combination of the patient, and matching the initial health management scheme; acquiring action data and physiological health index data of the patient executing the scheme; and inputting the data into a pre-trained dynamic adaptive updating model which is constructed based on Bi-LSTM and an attention mechanism, and directly outputting an updated personalized health management scheme through multi-objective loss function optimization. According to the invention, end-to-end self-updating of the health management scheme is realized, and the individuation degree of the scheme is improved.

Description

Technical Field

[0001] This invention relates to the field of health management technology, specifically to a dynamic health management method, system, and storage medium for diabetic nephropathy. Background Technology

[0002] Type 2 diabetic kidney disease (DKD) is one of the most common complications in patients with type 2 diabetic kidney disease and a leading cause of end-stage renal disease (ESRD), accounting for 59% of new ESRD cases. With the increasing incidence of diabetes year by year, the high prevalence of DKD has brought enormous pressure and challenges to the healthcare system and the socio-economic landscape.

[0003] Currently, most health management plans for DKD patients are subjectively designed by doctors based on clinical experience, which presents the following technical problems: ① Inconsistent standards and reliance on subjective experience: Differences in experience among doctors lead to a lack of unified standards for health management plans, making it difficult to guarantee patient health benefits; ② Delayed plans and insufficient timeliness: Plan adjustments are usually made only after patients experience discomfort or management becomes ineffective, failing to respond in real time to dynamic changes in the patient's condition; ③ Long update chain and low efficiency: Traditional monitoring-analysis-rematching plan update methods require repeated tag analysis and plan matching, resulting in a long processing chain and impacting management efficiency; ④ Lack of personalized consideration: Existing plans struggle to take into account patients' individual preferences, leading to poor patient compliance and affecting long-term management outcomes. Therefore, current health management plans for DKD patients suffer from long update chains, lack of personalization, and insufficient long-term management performance. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic health management method, system, and storage medium for diabetic nephropathy, in order to solve the technical problems of long update links, lack of personalization, and insufficient long-term management performance in existing health management solutions.

[0005] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A dynamic health management method for diabetic nephropathy includes the following steps: In the dynamic health management system, the influencing factors of the personalized health management plan for patients with type 2 diabetic nephropathy are divided into multiple tag sets, and each tag set is matched with a personalized tag that reflects the personalized characteristics of patients with type 2 diabetic nephropathy in terms of influencing factors. The influencing factors include the health information, self-management information, risk level information and expert consensus information of patients with type 2 diabetic nephropathy. The health information, self-management information and expert consensus information are all directly collected by the system, and the risk level information is indirectly calculated through a pre-established risk prediction model for the incidence of type 2 diabetic nephropathy. The dynamic health management system obtains real-time personalized tags for patients with type 2 diabetic nephropathy on various tag sets, and matches real-time personalized health management plans for patients with type 2 diabetic nephropathy based on these tags. The dynamic health management system monitors the action data of patients with type 2 diabetic nephropathy in performing the real-time personalized health management plan, and monitors the health indicators of patients with type 2 diabetic nephropathy after the implementation of the real-time personalized health management plan. The dynamic health management system dynamically adjusts the real-time personalized tags of patients with type 2 diabetic nephropathy based on the feedback of health indicators and action data to obtain new personalized tags, and uses the adjusted new personalized tags to match new personalized health management plans for patients with type 2 diabetic nephropathy. The dynamic health management system uses neural networks to establish health indicators and action data for real-time personalized health management plans, as well as the mapping relationship between real-time personalized health management plans and new personalized health management plans, to obtain a dynamic adaptive update model for personalized health management plans. After the dynamic adaptive update model is established, the dynamic health management system inputs the monitored health indicators and action data, as well as the real-time personalized health management plan, into the dynamic adaptive update model to obtain a new personalized health management plan for patients with type 2 diabetic nephropathy.

[0006] As a preferred embodiment of the present invention, the dynamic adaptive update model adopts an encoder-decoder architecture, wherein the encoder is a bidirectional long short-term memory network Bi-LSTM, used to process action data sequences and health indicator data sequences containing multiple historical time points, and the decoder is a fully connected network with residual connections, used to generate the new personalized health management scheme based on the hidden state output by the encoder. A temporal attention mechanism is introduced between the encoder and the decoder. The attention mechanism calculates attention weights based on the similarity between the hidden states of the encoder at each time step and the current decoding state, so that the model focuses on the historical data points that have the greatest impact on the scheme adjustment, while ignoring noisy data.

[0007] As a preferred embodiment of the present invention, the dynamic adaptive update model is trained using a multi-objective loss function, which includes: Loss of health improvement is measured using mean squared error to determine the gap between the updated health indicators and the expected health goals. The scheme smooths the loss and uses cosine similarity to measure the difference between the updated scheme and the historical scheme F, encouraging continuous changes in the content of the scheme; The patient preference loss is constructed based on the patient's historical implicit feedback to the plan (such as execution time and attendance rate), and the patient preference vector P is obtained through matrix factorization. The Euclidean distance between the new plan and the preference vector is used as the loss. To mitigate risk and reduce losses, the new treatment plan is input into a risk prediction model to calculate the expected change in the patient's risk level after the implementation of the new plan, and to encourage adjustments to the plan to reduce risk. By using a weighted summation to balance multiple loss terms, the model-generated solution achieves the optimal trade-off between health improvement, patient acceptance, solution continuity, and risk control.

[0008] As a preferred embodiment of the present invention, the construction of the patient preference loss includes: collecting multiple health management plans that each patient has historically executed, along with their corresponding execution durations and check-in completion rates, and constructing a patient-plan implicit feedback matrix; The weighted alternating least squares (WALS) algorithm is used to perform matrix decomposition to obtain the patient's implicit preference vector and the implicit attribute vector of the scheme; the patient preference loss is the negative cosine similarity between the implicit attribute vector of the new scheme and the patient preference vector.

[0009] As a preferred embodiment of the present invention, the method for constructing a risk prediction model for type 2 diabetic nephropathy includes: Using a multivariate logistic regression model, several characteristic factors that significantly influence the incidence of type 2 diabetic nephropathy were screened from historical cases. Principal component analysis was performed on each characteristic factor to obtain the contribution of each characteristic factor; Based on the aforementioned contribution, game theory is introduced to determine the optimal feature weight for each feature factor; Each feature factor is multiplied by its corresponding optimal feature weight and used as input features to train the risk prediction model, which is then used to output the patient's risk score. The risk prediction model is an attention-guided dual-stream temporal static fusion network. Its input consists of a static feature vector weighted by the optimal feature weight set and feature sequences from multiple follow-up time points of the patient. The output is a risk score for the onset of diabetic nephropathy, which is used to dynamically assess the patient's risk trend. The attention-guided dual-stream temporal static fusion network introduces a static feature encoder and a temporal feature encoder, which can simultaneously process baseline static data and follow-up dynamic data.

[0010] As a preferred embodiment of the present invention, the optimal feature weights are re-executed using newly added case data every preset period to update the weights of each feature factor, so that the risk prediction model can adapt to changes in the patient population.

[0011] As a preferred embodiment of the present invention, the risk level in the risk label set is determined by comparing the incidence risk score output by the risk prediction model with a preset dynamic threshold curve, and dividing the risk into high-risk, medium-risk or low-risk intervals according to the relative position of the score curve and the threshold curve. The dynamic threshold curve is adaptively adjusted based on the changes in the quantiles of historical population risk scores over time.

[0012] As a preferred embodiment of the present invention, a real-time personalized health management plan for patients with type 2 diabetic nephropathy is obtained by matching real-time personalized labels of patients with type 2 diabetic nephropathy on various label sets. This is achieved using deep metric learning: each plan in the plan library is pre-mapped to the same embedding space as the label combination through a Siamese network. By calculating the cosine similarity between the embedding of the label combination and the embedding of each plan, the plan with the highest similarity is selected as the matching result. The Siamese network is trained using triple loss, where the triple consists of anchor label combination, positive example plan (effective plan in history), and negative example plan (ineffective or inefficient plan in history), so that the embedding of effective plan is closer to the label combination, and the embedding of ineffective plan is further away.

[0013] As a preferred embodiment of the present invention, the present invention provides a dynamic health management system based on diabetic nephropathy, applied to a dynamic health management method based on diabetic nephropathy, the system comprising: A tag system construction unit is used to establish and store the multi-dimensional tag set system. The initial plan matching unit is used to generate the patient's current multidimensional personalized tag combination and match it with the initial health management plan; The data monitoring unit is used to monitor and acquire patients' action data and health indicator data in real time; The adaptive update unit has the dynamic adaptive update model built in, which is used to receive feedback data and directly output the updated health management plan. Preference modeling unit, used to learn and store patients' personalized preference information and participate in loss calculation; The risk dynamic assessment unit is used to dynamically classify risk levels based on the risk score curve output by the risk prediction model.

[0014] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a dynamic health management method for diabetic nephropathy.

[0015] Compared with the prior art, the present invention has the following advantages: This invention identifies a set of tags corresponding to the influencing factors of personalized health management plans for patients with type 2 diabetic nephropathy. By using the personalized tags of patients with type 2 diabetic nephropathy in each tag set, a personalized health management plan is accurately matched to the patients. Based on the combined personalized tags, a unified standard for matching health management plans is achieved. Based on the health indicators and the execution data, the personalized health management plan for patients with type 2 diabetic nephropathy is dynamically adjusted, enabling timely updates to the health management plan as the individual patient's condition changes.

[0016] This invention employs an attention-guided dual-stream temporal static fusion network to simultaneously process static baseline data and dynamic follow-up data. It also combines game theory-based combinatorial weighting to determine optimal feature weights, significantly improving the accuracy of risk prediction. The introduction of a dynamic weight update module and a dynamic threshold curve enables the risk prediction model to adapt to changes in the patient population.

[0017] This invention learns patients' personalized preferences through a preference modeling unit and incorporates the preference information into the plan update process. The generated plan is more in line with patients' habits, thereby improving patients' willingness to implement the plan and their long-term compliance, and ultimately improving the effectiveness of health management. Attached Figure Description

[0018] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart of a dynamic health management method for diabetic nephropathy provided in an embodiment of the present invention; Figure 2 A selected diagram of data collected from cases provided in an embodiment of the present invention; Figure 3The ROC curve for predicting the incidence risk based on a dual-stream temporal static fusion network provided in this embodiment of the invention; Figure 4 ROC curve for disease incidence prediction based on CNN provided in this embodiment of the invention; Figure 5 Calibration curve for disease incidence prediction based on a dual-stream temporal static fusion network provided in this embodiment of the invention; Figure 6 A comparison chart of management effects provided for embodiments of the present invention; Figure 7 This is a block diagram of a dynamic health management system for diabetic nephropathy provided in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] like Figure 1 As shown, this invention provides a dynamic health management method for diabetic nephropathy, comprising the following steps: In the dynamic health management system, the influencing factors of the personalized health management plan for patients with type 2 diabetic nephropathy are divided into multiple tag sets, and each tag set is matched with a personalized tag that reflects the personalized characteristics of the patients with type 2 diabetic nephropathy in terms of influencing factors. The influencing factors include the patients' health information, self-management information, risk level information, and expert consensus information. The health information, self-management information, and expert consensus information are all directly collected by the system, and the risk level information is indirectly calculated through a pre-established risk prediction model for the incidence of type 2 diabetic nephropathy.

[0022] Specifically, the tag set system includes at least four core dimensions: a patient tag set consisting of the patient's health information elements, a self-management tag set consisting of the patient's self-management behaviors, a risk tag set consisting of the risk level of type 2 diabetic nephropathy, and a protocol tag set consisting of the core management points extracted from clinical guidelines and expert consensus.

[0023] Patient tag set: Used to characterize the static and dynamic physiological characteristics of patients. In one specific embodiment, the personalized tags in this tag set consist of multiple health information elements with clear representativeness and discriminative power, including but not limited to: gender, age, race, marital status, education level, present medical history, past medical history, family history, medication history, smoking habits, drinking habits, exercise habits, sleep patterns, height, weight, blood pressure, blood glucose, cholesterol, glycated hemoglobin, estimated glomerular filtration rate, urine biochemistry, etc. These data are collected through hospital information systems, wearable devices, or patient self-reporting, etc. Figure 2 As shown, Figure 2 This is a selected chart of data collected from case studies. It should be noted that all personal health information collected complies with relevant laws and regulations and, with the patient's explicit authorization and consent, is used solely for disease and health management purposes.

[0024] Self-Management Tag Set: Based on the Fogg Behavioral Model, this set characterizes patients' self-management behaviors from two dimensions: motivation and ability. In one specific embodiment, the personalized tags in this tag set consist of various self-management behaviors along the ability and motivation dimensions of the Fogg Health Behavioral Model, such as 24 combined tags like "diet - high ability, low motivation," "exercise - high ability, high motivation," and "medication - low ability, low motivation." These tags reflect the patient's intrinsic motivation and execution ability to implement the health management plan and are key characteristics for predicting adherence.

[0025] Risk label set: Used to quantify the risk of a patient experiencing adverse events in the future. The personalized labels in this set consist of risk levels predicted by a type 2 diabetic nephropathy risk prediction model, including three levels: "high risk," "medium risk," and "low risk." The risk level classification is based on a comparison between the risk score output by the risk prediction model and a dynamic threshold.

[0026] The protocol tag set represents currently recognized effective management measures in the medical community. The personalized tags in this set consist of core elements of type 2 diabetic nephropathy management extracted from domestic and international literature, guidelines, and expert consensus. Examples include "monitoring blood glucose," "low-GI diet," "preventing complications," "quitting smoking and limiting alcohol consumption," "emotional management," "taking medication on time," "monitoring blood pressure," "low-fat and low-salt diet," "regular exercise," "monitoring kidney function," "adequate protein intake," and "regular lifestyle." Each protocol tag corresponds to a specific health management action or recommendation.

[0027] The dynamic health management system obtains real-time personalized tags for patients with type 2 diabetic nephropathy on various tag sets, and matches real-time personalized health management plans for patients with type 2 diabetic nephropathy based on these tags.

[0028] In this embodiment, the scheme matching is not a simple rule matching, but rather implemented using deep metric learning. Each scheme in the scheme library is pre-mapped to the same embedding space as its label combination using a Siamese network. The Siamese network consists of two fully connected neural networks with identical structures and shared weights, each with a 128-dimensional output. During matching, the first multi-dimensional personalized label combination of the current patient is input into a branch of the Siamese network to obtain the embedding vector of the label combination. Simultaneously, the vectors of all solutions in the solution library are input into another branch of the Siamese network to obtain the embedding vectors of each solution. Through calculation With each The cosine similarity is used to select the scheme with the highest similarity as the matching result.

[0029] To train this Siamese network, a triplet loss function is used. Each triplet consists of an anchor (the current patient's label combination), positive examples (solutions effective for patients with similar label combinations in historical data), and negative examples (solutions ineffective or inefficient for patients with similar label combinations in historical data). The triplet loss function is defined as follows: Where d represents the Euclidean distance, and the margin is set to 0.5. As anchor point, As a positive sample, These are negative examples. By minimizing this loss function, the embeddings of effective solutions are made closer to the label combinations, while the embeddings of ineffective solutions are made further away. This allows the network to learn the intrinsic relationship between solutions and patient status, achieving more accurate initial solution recommendations.

[0030] The dynamic health management system monitors the action data of patients with type 2 diabetic nephropathy in performing the real-time personalized health management plan, and monitors the health indicators of patients with type 2 diabetic nephropathy after the implementation of the real-time personalized health management plan.

[0031] Action data reflects the patient's adherence to the treatment plan, such as daily medication frequency (integer type), exercise duration (minutes), dietary records (Boolean type), and blood glucose monitoring frequency (times / day). This data is automatically collected through smartphone applications, smart pillboxes, wearable devices, etc., forming time-series data. Physiological health indicator data objectively reflects the effectiveness of the treatment plan, such as fasting blood glucose (mmol / L), postprandial blood glucose (mmol / L), blood pressure (mmHg), weight (kg), and urinary microalbumin / creatinine ratio (mg / g). This data can be automatically uploaded through home medical devices or entered into the system during regular hospital follow-ups, and is also stored in time-series format.

[0032] The dynamic health management system dynamically adjusts the real-time personalized tags of patients with type 2 diabetic nephropathy based on the feedback of health indicators and action data to obtain new personalized tags, and uses the adjusted new personalized tags to match new personalized health management plans for patients with type 2 diabetic nephropathy.

[0033] The dynamic health management system uses neural networks to establish health indicators and action data for real-time personalized health management plans, as well as the mapping relationship between real-time personalized health management plans and new personalized health management plans, to obtain a dynamic adaptive update model for personalized health management plans. After the dynamic adaptive update model is established, the dynamic health management system inputs the monitored health indicators and action data, as well as the real-time personalized health management plan, into the dynamic adaptive update model to obtain a new personalized health management plan for patients with type 2 diabetic nephropathy.

[0034] A dynamic adaptive update model, employing an encoder-decoder architecture, is used to process time series data and generate new schemes. The specific structure is as follows: The encoder employs a bidirectional long short-term memory network (Bi-LSTM). The input data consists of sequences of the patient's actions over the past T time steps (e.g., the past 30 days). and physiological health indicator data sequence ,in and These are the feature dimensions for behavioral data and physiological data, respectively. In the management of diabetic nephropathy, behavioral data can be specifically categorized as "number of low-GI diet check-ins," "medication adherence rate for hypoglycemic drugs," and "daily steps," etc.; physiological data can be specifically categorized as "fasting blood glucose (mmol / L)," "systolic blood pressure (mmHg)," and "urine microalbumin / creatinine ratio (mg / g)," etc. The two sequences are first mapped to the same latent space dimension d=128 through independent embedding layers, and then concatenated to form a sequence with dimension d=128. The synthesized sequence is input into a Bi-LSTM. The Bi-LSTM consists of two layers, each with a hidden state dimension of 128. The outputs of the forward and backward LSTMs are concatenated at each time step to form the hidden state at that time step. The final encoder outputs the hidden state sequence H = [ for all time steps]. ] .

[0035] An attention mechanism is introduced between the encoder and decoder. This mechanism aims to automatically identify the most critical "event points" for adjusting treatment plans during the long course of diabetic nephropathy patients, such as a drastic fluctuation in blood glucose or a missed medication dose. The attention mechanism calculates attention weights based on the similarity between the hidden states of the encoder at each time step and the current state of the decoder. Specifically, for the j-th step of the decoder, its state... Hidden states of the encoder at each time step Attention score , where v, , These are trainable parameters. Attention weights are obtained by normalization using the softmax function. Then calculate the context vector. The context vector With decoder state The concatenated data is used to generate the output of step j. Through the attention mechanism, the model can focus on the historical data points that have the greatest impact on the scheme adjustment, while effectively suppressing the interference of noisy data.

[0036] The decoder employs a fully connected network with residual connections. The decoder's input includes the current scheme vector. (Dimension 128) and context vectors generated by the attention mechanism (Dimension 256). These inputs are first concatenated to form a vector of dimension 384, then passed through three fully connected layers, each with 256, 256, and 128 neurons, respectively, using ReLU activation. A residual connection is added after each fully connected layer, meaning the output of that layer is F(x) + x, where F(x) is the transformation of the fully connected layer. The final output of the decoder is the second personalized health management plan. The output dimension is consistent with the dimension of the scheme embedding vector (128 dimensions), and then the closest specific scheme is found from the scheme library through similarity matching.

[0037] During the training of the dynamically adaptive update model, the model is trained using a supervised learning approach. The training dataset contains a large number of historical patient quadruples: (historical scheme F, historical behavior sequence) Historical physiological sequence A practical and effective new solution During the training process, F, , Input the model, and the model outputs the predicted values. Calculated through loss function With real labels The error between them is calculated, and the network parameters are updated using the backpropagation algorithm.

[0038] Data on patients with type 2 diabetic nephropathy who have completed a management cycle will be extracted from the hospital information system and wearable device database. Each sample must include: the management plan (F) for the previous cycle, daily behavioral and physiological data during the cycle, and the new plan for the next cycle developed by the physician for the patient. .

[0039] Missing values ​​were imputed and normalized for behavioral and physiological data. Management plans and new plans developed by doctors were mapped to fixed-dimensional embedding vectors.

[0040] The He initialization method is used to initialize all weight matrices and bias terms in the network. The initial hidden states and cell states of the Bi-LSTM are set to zero vectors.

[0041] Input a batch of sample data (Batch Size = 64) into the model and perform calculations according to the hierarchical connections described above to obtain the predicted new scheme embedding vector.

[0042] The error between the predicted value and the true value is calculated using a multi-objective loss function.

[0043] The Adam optimizer is used for gradient updates, with an initial learning rate of 0.001. A learning rate decay strategy is implemented, reducing the learning rate to 0.9 every 10 epochs. The number of training epochs is set to 100, and early stopping is employed: training stops when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting.

[0044] The dynamic adaptive update model is trained using a multi-objective loss function, which consists of four parts: Health improvement loss Mean squared error is used to measure the updated health indicators. With expected health goals The difference between the two; this loss aims to maximize the health benefits of the new approach. Mean squared error is used to measure the updated health indicators. (i.e., the health indicators predicted by the model that patients can achieve after the new treatment plan is implemented) and the expected health goals The gap between them. These can be target values ​​set based on clinical guidelines, or they can be the patient's personalized ideal health status. Specifically, From an auxiliary prediction network And predicted from the current patient status Smoothing loss scheme The updated scheme is measured using cosine similarity. Compared with historical plans The difference between the two encourages continuous changes in the content of the protocol; this loss is intended to encourage new protocols to maintain a certain degree of continuity with historical protocols in terms of content, avoiding drastic changes that may cause patients to find it difficult to adapt. Cosine similarity is used to measure the updated protocol. The difference between the new scheme and the historical scheme F. By minimizing this term, the new scheme can achieve a "smooth transition" in content from the old scheme.

[0045] Loss of patient preference A preference matrix is ​​constructed based on the implicit feedback of patients to the treatment plan in their history (such as execution time and attendance rate), and the patient preference vector is obtained through matrix factorization. Embedded with a new solution The Euclidean distance to the preference vector is used as the loss; this loss aims to make the new plan more aligned with the patient's personalized preferences, thereby improving long-term adherence. Patient preferences are modeled through their implicit feedback to historical plans. Specifically, this is achieved by collecting data on multiple health management plans historically implemented by each patient, along with their corresponding execution duration (minutes / day) and completion rate (%), constructing a patient-plan implicit feedback matrix R, where... Let represent the overall feedback score of patient i for plan j, calculated as a weighted sum of the normalized value of execution time and the check-in completion rate. The weighted alternating least squares algorithm is used to decompose this matrix to obtain the patient's implicit preference vector. (Dimension 64) and the implicit attribute vector of the scheme (Dimension 64). The patient preference loss is the implicit attribute vector of the new treatment plan. or With patient preference vector The Euclidean distance between them.

[0046] Risk control loss The new treatment plan is input into a risk prediction model to calculate the expected change in patient risk level after implementation. This encourages adjustments to the treatment plan to reduce risk. This loss is intended to ensure that the new plan effectively controls the risk of disease progression in patients. The new plan... The data is input into a pre-trained risk prediction model to calculate the change in the patient's risk level over a future period after the implementation of the plan. This means encouraging risk scoring corresponding to new solutions. Not higher than the current risk score This guides the plan towards reducing risk. Risk score for the current patient.

[0047] The total loss function is ,in The weighting coefficients for each loss can be used to balance multiple loss terms through a weighted summation method, so that the model-generated solution achieves the optimal trade-off between health improvement, patient acceptance, solution continuity, and risk control.

[0048] In some specific implementations, the construction of patient preference loss includes: collecting multiple health management plans that each patient has historically executed, along with their corresponding execution durations and check-in completion rates, to construct a patient-plan implicit feedback matrix; performing matrix decomposition using the weighted alternating least squares (WALS) algorithm to obtain the patient's implicit preference vector and the plan's implicit attribute vector; and the patient preference loss is the negative cosine similarity between the new plan's implicit attribute vector and the patient's preference vector.

[0049] By optimizing the aforementioned multi-objective loss function, the model can find the optimal balance between health improvement, protocol continuity, patient preference, and risk control, resulting in a protocol that is both effective and easily accepted by patients. The algorithmic features (multi-objective loss function) and technical features (health indicators, motion data, protocol content) are functionally mutually supportive and interactive, jointly solving the technical problem of balancing multiple objectives during protocol updates.

[0050] In the context of diabetic nephropathy management, the patient's historical management plan (e.g., "low-protein diet + daily blood glucose monitoring") and execution feedback (behavioral data: diet adherence rate 90%; physiological data: elevated urinary protein) jointly determine the current management effectiveness. By learning the "plan-feedback-new plan" mapping pattern from a large amount of historical data, the model can infer the optimal adjustment strategy for specific feedback information. For example, when the model detects elevated urinary protein (physiological data) and poor dietary control (behavioral data), its output new plan should focus more on specific management actions such as "strictly restricting protein intake" and "increasing nutritionist guidance." Step S5, Risk Prediction and Dynamic Assessment: The patient's status is dynamically assessed using a risk prediction model, and the results are used to update the risk label set and participate in plan updates.

[0051] In some specific implementations, the methods for constructing a risk prediction model for type 2 diabetic nephropathy include: We used a multivariate logistic regression model to analyze the influence of factors on the pathogenesis of type 2 diabetic nephropathy and screened out characteristic factors that have a significant impact on the pathogenesis of type 2 diabetic nephropathy.

[0052] In this embodiment, it is first necessary to collect all influencing factors related to the onset of type 2 diabetic nephropathy. Specifically, influencing factors include demographic characteristics, clinical information, laboratory tests, renal biopsy pathology results, physical examination, lifestyle, medication use, past medical history and family history, socioeconomic status, and emotional state. The demographic characteristics include name, ID number, gender, date of birth, and place of residence; clinical information includes the date of diagnosis of diabetes, screening results for diabetic nephropathy, and coexisting conditions such as diabetic retinopathy, diabetic peripheral neuropathy, and diabetic macrovascular complications; laboratory tests include fasting blood glucose, glycated hemoglobin, blood pressure, blood lipids, renal function indicators (such as eGFR and urine protein levels), complete blood count, and urinalysis; renal biopsy pathology results include immunofluorescence examination, light microscopy, and electron microscopy results; physical examination includes height and weight at admission; lifestyle includes dietary habits, sleep patterns, smoking, alcohol consumption, and physical activity; medication use includes the use of hypoglycemic drugs, antihypertensive drugs, and lipid-lowering drugs; past medical history and family history include a history of hypertension, lipid metabolism disorders, coronary heart disease, and cerebrovascular disease, as well as a family history of diabetes; socioeconomic status includes per capita annual family income, education level, employment status, marital status, and type of medical reimbursement; and emotional state includes symptoms of depression and anxiety.

[0053] In one specific implementation, SPSS 26.0 and RStudio 4.2.2 software were used to analyze case data of type 2 diabetic nephropathy. Quantitative data conforming to a normal distribution were expressed as mean ± standard deviation (SD). The numbers indicate that qualitative data that do not conform to a normal distribution are expressed as frequencies and percentages (%). To compare the differences in various influencing factors between patients with non-type 2 diabetic nephropathy and patients with type 2 diabetic nephropathy, the t-test was used for comparisons of normally distributed quantitative data between groups, and the χ² test was used for comparisons of qualitative data that do not conform to a normal distribution between groups. 2 Test, rank-sum test, with For statistically significant differences, logistic regression analysis was used to screen influencing factors and test the significance level. The selected influencing factors were then used as characteristic factors.

[0054] Before using a multivariate logistic regression model for initial screening, LASSO regression is employed to preliminarily select influencing factors. LASSO regression, by constructing a penalty function, can compress the coefficients of unimportant features to zero, thereby achieving feature selection. Specifically, influencing factors with non-zero coefficients in the LASSO regression are considered as candidate factors and then incorporated into the multivariate logistic regression model for stepwise regression analysis to determine the final feature factors. This dual screening mechanism effectively handles high-dimensional data and improves the stability and accuracy of feature selection.

[0055] Principal component analysis was performed on each characteristic factor to obtain its contribution.

[0056] In this embodiment, principal component analysis is performed on each of the selected feature factors to identify the contribution of each feature factor to the risk of disease. The higher the contribution, the more important the feature factor, and the more the disease risk prediction model will favor the information provided by that feature factor during the disease risk calculation process.

[0057] After performing principal component analysis on each characteristic factor, the number of principal components is determined based on the principle that the cumulative variance contribution rate is greater than 85%. The weighted average of the coefficients of each characteristic factor in the linear combination of principal components is calculated using the variance contribution rate of each principal component as the weight, and this average is taken as the contribution of the characteristic factor.

[0058] Specifically, suppose principal component analysis yields k principal components, and the variance contribution rate of each principal component is... (j=1,2,...,k), the coefficient of each feature factor in the j-th principal component is Then the contribution of the i-th feature factor It can be represented as: Considering that the sign of the characteristic load only represents the direction of correlation, and the magnitude of its absolute value can better reflect the contribution intensity of the feature in the principal component, the absolute value is used for weighted summation.

[0059] This method can more accurately reflect the true importance of features and reduce noise interference.

[0060] Based on contribution, a centralized feature weight is constructed. In the centralized feature weight, the higher the contribution of the feature factor, the greater the weight is assigned to that feature factor, so as to achieve information centralization of the feature factor. Based on contribution, a decentralized feature weight is constructed. In the decentralized feature weight, the lower the contribution of the feature factor, the greater the weight is assigned to that feature factor, so as to achieve information decentralization of the feature factor.

[0061] In this embodiment, two methods are used to assign weights to feature factors. The first method assigns a higher feature weight to a feature factor with a higher contribution, further enhancing its importance. This causes the disease risk prediction model to favor information provided by highly important feature factors during risk assessment, thus achieving the effects of relevant information extraction and redundant information removal, and centralizing feature information. The second method assigns a higher feature weight to a feature factor with a lower contribution, weakening its importance. This causes the disease risk prediction model to favor information provided by less important feature factors during risk assessment, thus achieving comprehensive information coverage and avoiding centralized feature information.

[0062] The methods for centralizing the feature weights of feature factors include: Contribution to characteristic factors After normalization, we get ,in, , These represent the contributions of the i-th and j-th feature factors, respectively. for The normalized result, where n is the total number of feature factors; Calculate the information entropy redundancy of each feature factor. The composite function of (overall entropy) and individual contribution The centered feature weights are then obtained by normalization. , Let be the feature weight of the i-th feature factor after centralization. This method combines information entropy theory with principal component contribution, considering not only the variance contribution of the feature but also the amount of information carried by the feature, making the weight allocation more information-theoretic.

[0063] Decentralized methods for handling feature weights of feature factors include: Contribution of characteristic factors Convert to probability distribution ,in, , These represent the contributions of the i-th and j-th feature factors, respectively. for The normalized result, where n is the total number of feature factors; Calculate the complementary entropy weights for each feature. The decentralized feature weights are then obtained by normalization. , Let be the feature weight of the i-th feature factor after decentralized processing. High-contribution features usually correspond to low entropy (strong information certainty), while low-contribution features correspond to high entropy (high information uncertainty). Complementary entropy weights automatically assign higher weights to low-contribution features, achieving information compensation and having a clear information theory interpretation.

[0064] By introducing game theory concepts, we treat the centralized and decentralized feature weights as the two sides of a game, and solve for an optimal feature weight that balances information centralization and information decentralization.

[0065] In this embodiment, both methods of assigning weights to the above feature factors have their own advantages and complement each other. Therefore, by using a game theory approach to make optimal decisions among the feature weights obtained from the two methods of assigning weights to the above feature factors, the optimal feature weights are obtained. This allows the optimal feature weights to seek a compromise between their respective advantages and the advantages of the other, thereby achieving a balance between the two advantages.

[0066] Let the feature weights after centering be a vector. The decentralized feature weights are vectors. n is the total number of feature factors; construct a combined weight set consisting of a linear combination of W and V. ,in and The coefficients are used for combination; the objective function is established with the goal of minimizing the deviation between the combined weight set U and W and V. , This is the minimization operator; Solve the objective function to obtain the optimal combination coefficients. and ; Calculate the optimal feature weights based on the optimal combination coefficients. Then, G is normalized to obtain the final optimal feature weights. .

[0067] The essence of this optimization goal is to find a compromise that can retain the "focusing" advantage of centralized weights while also taking into account the "coverage" advantage of decentralized weights, so that the integrated weight set is neither neglected nor redundant in terms of information utilization.

[0068] Furthermore, the objective function is solved using the Shapley value method from cooperative game theory. The centralized and decentralized feature weights are considered as participants in the game, and the expected marginal contribution of each participant in all possible alliances is calculated as the optimal combination coefficient. and The alternative values ​​are used to obtain the optimal feature weights. The Shapley value can allocate weights more fairly, taking into account the marginal contribution of each weight set under different combinations, making the fusion result more reasonable and robust.

[0069] By using optimal feature weights to weight feature factors and using the weighted feature factors as input, a risk prediction model for type 2 diabetic nephropathy is constructed to output the individual's risk of developing the disease.

[0070] In this embodiment, the optimal feature weights are used to weight the feature factors to train a risk prediction model for type 2 diabetic nephropathy, thereby enabling the assessment of the risk of type 2 diabetic nephropathy in patients based on their feature factors.

[0071] A risk prediction model for type 2 diabetic nephropathy employs an attention-guided dual-stream temporal static fusion network deeply integrated with optimal feature weights. This attention-guided dual-stream temporal static fusion network comprises a static feature encoding module, a temporal feature encoding module, a cross-attention interaction module, and a prediction output module. The network design fully considers the interaction between algorithmic features (attention mechanisms, game-theoretic weights) and technical features (clinical static indicators, dynamic follow-up data). By using prior weights as input to the static encoder, the weight information guides the feature representation; through the cross-attention interaction module, static features and dynamic temporal features mutually reinforce each other within the model, jointly improving prediction accuracy.

[0072] The static feature encoding module is used to further abstract the feature factors, which have been weighted a priori using optimal feature weights, into a latent space representation adapted to subsequent interactions, while preserving the differences in feature importance assigned by the optimal weight set. The specific structure includes: The input layer takes the static feature vector as a weighted average of the optimal feature weights. ,in Let be the optimal weight for the i-th feature factor. This represents the original value of the i-th feature factor; Fully connected layer 1, containing 128 neurons, followed by a batch normalization layer and a ReLU activation function; Fully connected layer 2, containing 64 neurons, followed by a batch normalization layer and a ReLU activation function; The fully connected layer 3 contains 32 neurons, followed by a batch normalization layer and a ReLU activation function, and outputs a static higher-order representation.

[0073] This module maps the original static features to a more expressive latent space through multi-layer nonlinear transformations, while preserving the differences in feature importance assigned by the optimal weight set. The optimal feature weights are fixed weights calculated offline using game theory fusion before model training and remain unchanged during training and inference, serving as prior knowledge to guide model learning and thus avoiding the computational complexity of online game theory inference. This design embodies a close integration of algorithmic features (game theory weights) and technical features (static feature encoding): game theory weights guide the static feature encoder to focus on important features, while the static feature encoder transforms the weighted features into a more abstract representation, enabling weight information to be transmitted and utilized in deep networks.

[0074] This optimal weight set, serving as prior knowledge, is not only used for feature weighting but also, in the subsequent static feature encoding process, continuously guides the model to focus on high-value information through nonlinear transformations of multi-layer neural networks. Furthermore, it deeply interacts with temporal features in the cross-attention module, jointly optimizing the feature representation in the latent space. This design enables the game-theoretic fusion weights and the feature extraction capabilities of the deep learning model to mutually support and enhance each other, thereby improving the accuracy of the final risk score.

[0075] The temporal feature encoding module is used to capture dynamic changes in patient follow-up data. Its specific structure includes: A bidirectional LSTM layer is used to process the follow-up time-series data matrix. Input a bidirectional LSTM network, and obtain the forward hidden state at each time step t. and backward hidden state After piecing them together, the complete hidden state is obtained. d represents the number of hidden units in the LSTM, which allows for the simultaneous consideration of past and future contextual information, thus more accurately depicting the evolution trend of the disease. The temporal attention layer is used to automatically assign weights to the hidden states at each time step through an attention mechanism. The specific calculation is as follows: ; ; in, These are the learnable attention parameter vector, attention weight matrix, and bias term, respectively. Let be the attention weight at time step t. Let this be the hidden state at the j-th time step. This is a dynamic representation of the entire time series.

[0076] This design enables the model to focus on information at key time points (such as periods of rapid blood sugar spikes or blood pressure fluctuations), improving prediction accuracy. Here, the attention mechanism, as an algorithmic feature, works in conjunction with the bidirectional LSTM technique to jointly solve the problem of identifying key time points in time-series data.

[0077] The cross-attention interaction module is used to capture and The cross-modal interaction module calculates the mutual attention weights between static high-order representations and temporal dynamic representations, enabling static features to dynamically select associated temporal information and, conversely, allowing temporal features to calibrate the expression of static features, thereby achieving deep fusion and enhancement of the two modalities. Static features and dynamic temporal features are not isolated; complex interactions exist between them. For example, a patient's age (static) may influence the contribution of blood glucose changes (dynamic) to the risk of kidney disease. To fully capture this cross-modal interaction, this module introduces a cross-attention mechanism. The specific structure includes: Static-temporal attention sublayer, used for static higher-order representation As a query, the time-series dynamic representation Using the key and value as keys, calculate the attention weights of static features on temporal features: ; in, These are the weight matrices for query, key, and value in the static-temporal attention sublayer, respectively. Scaling factor This is the normalized exponential function. The output... It is a statically guided temporal enhancement representation that dynamically adjusts the expression of temporal features based on the importance of static features.

[0078] Temporal-static attention sublayer, used for temporal dynamic representation As a query, static higher-order representation As keys and values, calculate the attention weights of temporal features on static features: ; in, These are the weight matrices for the query, key, and value in the temporal-static attention sublayer, respectively. This output... It is a time-guided static enhancement representation that recalibrates the importance of static features based on time-varying patterns.

[0079] The fusion layer is used to combine... and The features are concatenated to obtain the fused feature vector. The splicing operation preserves the independence of the two types of interactive information, facilitating decision-making in subsequent fully connected layers.

[0080] The cross-attention module is key to the deep integration of algorithmic features and technical features: through the attention mechanism, static features and dynamic temporal features can be mutually calibrated and enhanced, so that the final feature representation contains both static background information and dynamic evolution information, thereby more comprehensively reflecting disease risk.

[0081] The prediction output module takes the fused feature vector F as input to a single-layer fully connected network and outputs the final disease risk score. ; in, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, Using the Sigmoid activation function, the output is mapped to the range of 0-1, representing the probability that a patient will develop type 2 diabetic nephropathy within a certain future timeframe. This output directly corresponds to the "risk of disease" in the technical problem, achieving end-to-end prediction from input features to risk score.

[0082] In training an attention-guided dual-stream temporal static fusion network, a multi-dimensional cross-validation method is used to optimize the network's hyperparameters. Specifically, K-fold cross-validation is employed, randomly dividing the training data into K parts. K-1 parts are used as the training set, and 1 part as the validation set, for K training and validation iterations. The optimal combination of hyperparameters, including the number of network layers, number of neurons, learning rate, and batch size, is then selected. The loss function used during training is binary cross-entropy loss. ; in, For true labels (0 or 1). Let N be the probability predicted by the model and N be the number of samples. The Adam optimizer is used for parameter updates, with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. An early stopping strategy is employed to prevent overfitting. During hyperparameter tuning, Bayesian optimization is used to search for the optimal combination of hyperparameters, including learning rate, batch size, number of LSTM hidden units, and number of attention heads, to maximize the AUC value on the validation set.

[0083] In the performance evaluation of the risk prediction model for type 2 diabetic nephropathy, the model's calibration degree was assessed using a calibration curve, and the area under the ROC curve (AUC) was used to evaluate the model's discriminative ability. Calibration degree reflects the consistency between the model's predicted probability and the actual observed probability, while discriminative ability reflects the model's ability to distinguish between individuals with and without the disease. These evaluation metrics directly validate the model's effectiveness in addressing the technical problem (improving prediction accuracy).

[0084] In one specific implementation, a training set and a dataset are divided based on type 2 diabetes patients diagnosed at a hospital between 2016 and 2018 for model construction and validation.

[0085] First, LASSO regression was used to preliminarily screen influencing factors. The 25 influencing factors with non-zero coefficients in the LASSO regression were selected as candidate factors and then included in a multivariate logistic regression model for stepwise regression analysis. Finally, 15 characteristic factors with significant independent influence on the incidence of type 2 diabetic nephropathy were identified, including: age, duration of diabetes, glycated hemoglobin, systolic blood pressure, diastolic blood pressure, urinary protein quantification, eGFR, BMI, smoking history, family history of diabetes, high-fat diet, lack of exercise, retinopathy, peripheral neuropathy, and history of cardiovascular disease.

[0086] Secondly, principal component analysis was performed on the 15 selected feature factors. Based on the principle that the cumulative variance contribution rate is greater than 85%, the number of principal components was determined to be 6. The weighted average of the coefficients of each feature factor in the linear combination of principal components was calculated using the variance contribution rate of each principal component as the weight, which is taken as the contribution of the feature factor.

[0087] Then, based on contribution, centralized and decentralized feature weights are constructed. The centralized feature weights are constructed using a composite function of information entropy redundancy and individual contribution, while the decentralized feature weights are constructed using complementary entropy weights. The optimal feature weights are obtained by using the Shapley value method in cooperative game theory.

[0088] Next, an attention-guided dual-stream temporal static fusion network was constructed. The static feature encoding module adopted a three-layer fully connected network (128→64→32), the temporal feature encoding module adopted a bidirectional LSTM (hidden unit number d=32) combined with an attention mechanism, the cross-attention interaction module adopted a multi-head attention mechanism (number of heads=4), and the prediction output module adopted a single-layer fully connected network. The network training used the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. An early stopping strategy was adopted to prevent overfitting, and the loss function was binary cross-entropy loss. Bayesian optimization was used to tune the hyperparameters. The search space included the learning rate (0.0001~0.01), batch size (16~64), number of LSTM hidden units (16~64), and number of attention heads (2~8). With the goal of maximizing the validation set AUC, the optimal hyperparameter combination was finally determined to be: learning rate 0.0005, batch size 32, number of LSTM hidden units 32, and number of attention heads 4.

[0089] This invention constructs a comparative model based on a conventional convolutional neural network (CNN). This comparative model first uses the same optimal feature weights as this invention to weight the 15 selected feature factors, and then uses the weighted feature vector as input to a conventional convolutional neural network. Its specific network structure and training parameters are as follows: Input layer: Receives a 15-dimensional weighted feature vector.

[0090] Reshaping layer: Reshapes the input to The two-dimensional format is adapted to accommodate two-dimensional convolution operations.

[0091] First convolutional layer: Two-dimensional convolution, kernel size is [size missing]. The number of filters is 32, the activation function is ReLU, and the padding method is "same".

[0092] First pooling layer: Two-dimensional max pooling, pooling window size is... The step size is 2.

[0093] The second convolutional layer: a two-dimensional convolution with a kernel size of [size missing]. The number of filters is 64, the activation function is ReLU, and the padding method is "same".

[0094] Global average pooling layer: Performs global average pooling on the feature map, flattening the output into a one-dimensional vector.

[0095] Fully connected layer: 64 neurons, ReLU activation function.

[0096] Output layer: a single neuron with the sigmoid activation function, used to output binary classification probabilities (i.e., disease risk scores).

[0097] The model training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. It also employs the same early stopping strategy as the model in this invention (stopping training when the validation set loss no longer decreases for 10 consecutive epochs) to prevent overfitting.

[0098] Finally, the model's performance is evaluated. For example... Figure 3 , Figure 4 , Figure 5 As shown in Table 1, Figure 3 The ROC curve is used to evaluate the predictive performance of type 2 diabetic nephropathy formed by combining the optimal weights with an attention-guided dual-stream temporal static fusion network in this invention. Figure 4 This is the ROC curve used in the performance evaluation of type 2 diabetic nephropathy prediction, formed by combining the optimal weights with a conventional convolutional neural network (CNN) in this invention. Figure 5 This is the calibration curve used in the performance evaluation of predicting type 2 diabetic nephropathy, which is formed by combining the optimal weights with an attention-guided dual-stream temporal static fusion network in this invention.

[0099] As can be seen, the AUC value of this model is 0.872 on the validation set, and the calibration curve shows that the model's predicted probability is highly consistent with the actual observed probability, indicating that the model has good calibration. Furthermore, compared with the prediction performance of type 2 diabetic nephropathy based on existing CNN networks, this invention demonstrates that the network structure of the attention-guided dual-stream temporal static fusion network and the deep fusion of optimal feature weights can significantly improve the prediction performance of type 2 diabetic nephropathy risk without increasing online inference complexity.

[0100] These experimental data fully demonstrate that the technical solution of the present invention can effectively solve the technical problem of insufficient prediction accuracy in the prior art, and achieve beneficial technical effects.

[0101] Table 1 Performance Comparison of Different Models As can be seen from the above, the embodiments of this application first use a multivariate logistic regression model to screen out characteristic factors that have a significant impact on the incidence of type 2 diabetic nephropathy; secondly, principal component analysis is performed on the characteristic factors to obtain their contribution; then, based on the contribution, centralized feature weights (with information centralization) and decentralized feature weights (with information decentralization) are constructed, and the optimal feature weights are obtained through game theory fusion; finally, the characteristic factors are weighted using the optimal feature weights to construct an incidence risk prediction model based on an attention-guided dual-stream temporal static fusion network. In this way, the accuracy of prediction in the incidence risk prediction process of type 2 diabetic nephropathy is improved, thereby enhancing the practical value of clinical applications.

[0102] The disease risk score output by the risk prediction model is compared with a preset dynamic threshold curve, and high-risk, medium-risk, or low-risk intervals are divided according to the relative position of the score curve and the threshold curve. The dynamic threshold curve is adaptively adjusted based on the changes in the quantiles of historical population risk scores over time.

[0103] In some specific implementations, the risk score range of the top 25% of all cases in the case data of type 2 diabetic nephropathy can be classified as the high-risk threshold range, the middle 50% as the medium-risk threshold range, and the bottom 25% as the low-risk threshold range.

[0104] Correspondingly, such as Figure 6 As shown, this application also provides a risk prediction system for type 2 diabetic nephropathy, applied to the aforementioned risk prediction method for type 2 diabetic nephropathy. The system includes: The feature selection unit is used to analyze the influence of factors on the pathogenesis of type 2 diabetic nephropathy using a multivariate logistic regression model, and to screen out feature factors that have a significant impact on the pathogenesis of type 2 diabetic nephropathy. The contribution analysis unit is used to perform principal component analysis on each characteristic factor to obtain the contribution of the characteristic factor; The weight configuration and fusion unit is used to determine the feature weights for centralizing the feature factors based on their contribution; determine the feature weights for decentralizing the feature factors based on their contribution; and perform a game between the decentralized feature weights and the centralized feature weights to obtain the optimal feature weights. The model building and prediction unit is used to build a risk prediction model for type 2 diabetic nephropathy based on feature factors weighted with optimal feature weights.

[0105] like Figure 7 As shown, this invention provides a dynamic health management system based on diabetic nephropathy, applied to a dynamic health management method based on diabetic nephropathy. The system includes: Tag system building unit, used to establish and store a multi-dimensional tag set system; The initial plan matching unit is used to generate the patient's current multidimensional personalized tag combination and match it with the initial health management plan; The data monitoring unit is used to monitor and acquire patients' action data and health indicator data in real time; The adaptive update unit has a built-in dynamic adaptive update model, which is used to receive feedback data and directly output the updated health management plan. Preference modeling unit, used to learn and store patients' personalized preference information and participate in loss calculation; The risk dynamic assessment unit is used to dynamically classify risk levels based on the risk score curve output by the risk prediction model.

[0106] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a dynamic health management method for diabetic nephropathy.

[0107] In some specific implementations, the tag set includes: a patient tag set corresponding to health information, a self-management tag set corresponding to self-management information, a risk tag set corresponding to risk level information, and a protocol tag set corresponding to expert consensus information, wherein... The personalized tags in the patient tag set consist of multiple health information elements that are clearly representative and distinctive. The specific personalized tags in the patient tag set include tags such as gender, age, race, marital status, education level, present medical history, past medical history, family history, medication history, smoking habits, drinking habits, exercise habits, sleep status, height, weight, blood pressure, blood sugar, cholesterol, AIC, eGFR, urine biochemistry, etc. The self-management tag set is composed of various self-management behaviors on the ability and motivation dimensions of the Fogg Health Behavior Model. Specifically, the self-management tag set includes high ability and low motivation for diet, high ability and low motivation for exercise, high ability and low motivation for medication, high ability and low motivation for limiting alcohol, high ability and low motivation for quitting smoking, high ability and low motivation for emotions, high ability and high motivation for diet, high ability and high motivation for exercise, high ability and high motivation for medication, high ability and high motivation for limiting alcohol, high ability and high motivation for quitting smoking, high ability and high motivation for emotions, low ability and low motivation for diet, low ability and low motivation for exercise, low ability and low motivation for medication, low ability and low motivation for limiting alcohol, low ability and low motivation for quitting smoking, low ability and low motivation for emotions, low ability and low motivation for diet, low ability and high motivation for exercise, low ability and high motivation for medication, low ability and high motivation for limiting alcohol, low ability and high motivation for quitting smoking, and low ability and high motivation for emotions. The personalized labels in the risk label set consist of various levels of risk of developing type 2 diabetic nephropathy predicted by the type 2 diabetic nephropathy risk prediction model, specifically including: high risk level, medium risk level and low risk level. The personalized tags in the protocol tag set are composed of the core points of type 2 diabetic nephropathy management extracted from domestic and foreign literature, guidelines and expert consensus. The specific personalized tags include blood glucose monitoring, low-GI diet, prevention of complications, smoking cessation and alcohol limitation, emotional management, timely medication, blood pressure monitoring, low oil and low salt, regular exercise, monitoring of kidney function, adequate protein, and regular lifestyle.

[0108] In some specific embodiments, the present invention labels the various influencing factors of the personalized health management plan for patients with type 2 diabetic nephropathy, dividing them into a plan label set, a risk label set, a self-management label set, and a patient label set. By using the positioning and matching of personalized labels in the label set, the present invention achieves the effect of matching a personalized health management plan for patients with type 2 diabetic nephropathy.

[0109] The personalized tags for patients with type 2 diabetic nephropathy in this invention can display multiple aspects of the patient's individual information. Matching health management plans with these tags can achieve a high degree of fit between the plan and the individual patient. For example, if a patient does not control their diet, they will be matched with a low-motivation, high-ability, low-risk diet plan; if a patient does not take medication regularly, they will be matched with a low-motivation, low-ability, high-risk medication plan; if a patient has not quit smoking effectively, they will be matched with a high-motivation, low-ability, medium-risk smoking plan; if a patient lacks exercise, they will be matched with a low-motivation, low-ability, medium-risk exercise plan, and so on.

[0110] This invention achieves standardized matching of health management plans based on combined personalized tags, eliminating the need for subjective health management that relies on experience and ensuring the objectivity of health management plan formulation.

[0111] In some specific embodiments, this invention promotes the self-management of research subjects from three dimensions: motivation, ability, and trigger, based on the Fogg Health Behavior Model. It uses two functions: monitoring and interaction. The monitoring function links multiple monitoring devices (combining wearable devices, home monitoring devices, and hospital follow-up records) to form a cloud data ecosystem, which collects patients' health indicators (such as blood sugar, blood pressure, and weight) and interaction data (such as usage frequency and feedback information) in real time.

[0112] Interactive Functions: Research participants can view and record their health status at any time, using an AI-powered question-and-answer system to receive reminders, record entries, and check-ins. Based on monitoring and check-in results, the platform provides personalized health education content and feedback, supporting patients in gaining comprehensive support in terms of motivation, ability, and triggers in their daily lives. The patient's personalized combination tag status changes dynamically with time and interaction. After each follow-up visit, the health management plan for intervening in the patient's health is adjusted based on the updated personalized combination tag status. This allows for dynamic adjustments to the personalized health management plan for patients with type 2 diabetic nephropathy based on patient health indicators and execution data from health management check-ins and interactions, ensuring timely updates to the health management plan as the individual patient's situation changes, guaranteeing the timeliness of health management.

[0113] The above-mentioned methods enable dynamic adjustments to personalized health management plans for patients with type 2 diabetic nephropathy, allowing for timely updates to the plans as the individual patient's condition changes. However, patient tag analysis and plan matching are still necessary. To further improve the efficiency of plan updates and ensure timeliness, this invention constructs a dynamic adaptive update model for health management plans. This model predicts new personalized health management plans based on the current personalized health management plan's health indicators and execution data, thus automating the personalized health management plan model update and simplifying the processes of tag duplication analysis and plan duplication matching.

[0114] This invention identifies a set of tags corresponding to the influencing factors of personalized health management plans for patients with type 2 diabetic nephropathy. By using the personalized tags of patients with type 2 diabetic nephropathy in each tag set, a personalized health management plan is accurately matched to the patients. Based on the combined personalized tags, a unified standard for matching health management plans is achieved. Based on the health indicators and the execution data, the personalized health management plan for patients with type 2 diabetic nephropathy is dynamically adjusted, enabling timely updates to the health management plan as the individual patient's condition changes.

[0115] This invention simplifies the traditional long "monitoring-analysis-rematching" chain into a short "input feedback data-direct output of new plan" chain by constructing an end-to-end dynamic adaptive update model. This significantly improves the efficiency and real-time performance of health management plan updates. Simultaneously, the initial plan matching based on a multi-dimensional label system, combined with an intelligent matching algorithm based on deep metric learning, ensures the accuracy of the initial plan. The dynamic adaptive update model, through a multi-objective loss function, balances health improvement with plan smoothness and patient preference, generating updated plans that are both effective and easily accepted by patients.

[0116] This invention employs an attention-guided dual-stream temporal static fusion network to simultaneously process static baseline data and dynamic follow-up data. It also combines game theory-based combinatorial weighting to determine optimal feature weights, significantly improving the accuracy of risk prediction. The introduction of a dynamic weight update module and a dynamic threshold curve enables the risk prediction model to adapt to changes in the patient population.

[0117] This invention learns patients' personalized preferences through a preference modeling unit and incorporates the preference information into the plan update process. The generated plan is more in line with patients' habits, thereby improving patients' willingness to implement the plan and their long-term compliance, and ultimately improving the effectiveness of health management.

[0118] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A dynamic health management method for diabetic nephropathy, characterized in that, Includes the following steps: In the dynamic health management system, the influencing factors of the personalized health management plan for patients with type 2 diabetic nephropathy are divided into multiple tag sets, and each tag set is matched with a personalized tag that reflects the personalized characteristics of patients with type 2 diabetic nephropathy in terms of influencing factors. The influencing factors include the health information, self-management information, risk level information and expert consensus information of patients with type 2 diabetic nephropathy. The health information, self-management information and expert consensus information are all directly collected by the system, and the risk level information is indirectly calculated through a pre-established risk prediction model for the incidence of type 2 diabetic nephropathy. The dynamic health management system obtains real-time personalized tags for patients with type 2 diabetic nephropathy on various tag sets, and matches real-time personalized health management plans for patients with type 2 diabetic nephropathy based on these tags. The dynamic health management system monitors the action data of patients with type 2 diabetic nephropathy in performing the real-time personalized health management plan, and monitors the health indicators of patients with type 2 diabetic nephropathy after the implementation of the real-time personalized health management plan. The dynamic health management system dynamically adjusts the real-time personalized tags of patients with type 2 diabetic nephropathy based on the feedback of health indicators and action data to obtain new personalized tags, and uses the adjusted new personalized tags to match new personalized health management plans for patients with type 2 diabetic nephropathy. The dynamic health management system uses neural networks to establish health indicators and action data for real-time personalized health management plans, as well as the mapping relationship between real-time personalized health management plans and new personalized health management plans, to obtain a dynamic adaptive update model for personalized health management plans. After the dynamic adaptive update model is established, the dynamic health management system inputs the monitored health indicators and action data, as well as the real-time personalized health management plan, into the dynamic adaptive update model to obtain a new personalized health management plan for patients with type 2 diabetic nephropathy.

2. The method according to claim 1, characterized in that, The dynamic adaptive update model adopts an encoder-decoder architecture, where the encoder is a bidirectional long short-term memory network Bi-LSTM, used to process action data sequences and health indicator data sequences containing multiple historical time points, and the decoder is a fully connected network with residual connections, used to generate the new personalized health management plan based on the hidden state output by the encoder. A temporal attention mechanism is introduced between the encoder and the decoder. The attention mechanism calculates attention weights based on the similarity between the hidden states of the encoder at each time step and the current decoding state, so that the model focuses on the historical data points that have the greatest impact on the scheme adjustment, while ignoring noisy data.

3. The method according to claim 2, characterized in that, The dynamic adaptive update model is trained using a multi-objective loss function, which includes: Loss of health improvement is measured using mean squared error (MSE) of the updated health indicators. With expected health goals The gap between ; The scheme smooths the loss and uses cosine similarity. Measure the updated solution The differences between the proposed scheme and historical scheme F encourage continuous changes in the scheme's content. ; Patient preference loss is addressed by constructing a preference matrix based on patients' historical implicit feedback to the proposed solutions. This is then used to obtain the patient preference vector P through matrix factorization, along with the implicit attribute vector of the new solution. The obtained Euclidean distance to the preference vector is used as the loss. ; To mitigate risk and prevent losses, the new personalized health management plan is input into a risk prediction model to calculate the expected change in the patient's risk level after implementation, encouraging adjustments to the plan to reduce risk. , Risk scores corresponding to the new personalized health management plan This represents the current risk score, with `max` being the maximum value operator. By using a weighted summation to balance multiple loss terms, the model-generated solution achieves an optimal trade-off between health improvement, patient acceptance, solution continuity, and risk control. ,in These are the weighting coefficients for each loss.

4. The method according to claim 3, characterized in that, The construction of the patient preference loss includes: collecting multiple health management plans that each patient has historically executed, along with their corresponding execution duration and check-in completion rate, and constructing a patient-plan implicit feedback matrix; The weighted alternating least squares (WALS) algorithm is used for matrix decomposition to obtain the patient's implicit preference vector and the implicit attribute vector of the treatment plan. The patient preference loss is the negative cosine similarity between the implicit attribute vector of the new personalized health management plan and the patient preference vector.

5. The method according to claim 1, characterized in that, Methods for constructing risk prediction models for type 2 diabetic nephropathy include: Using a multivariate logistic regression model, several characteristic factors that significantly influence the incidence of type 2 diabetic nephropathy were screened from historical cases. Principal component analysis was performed on each characteristic factor to obtain the contribution of each characteristic factor; Based on the aforementioned contribution, game theory is introduced to determine the optimal feature weight for each feature factor; Each feature factor is multiplied by its corresponding optimal feature weight and used as input features to train the risk prediction model, which is then used to output the patient's risk score. The risk prediction model is an attention-guided dual-stream temporal static fusion network. Its input consists of a static feature vector weighted by the optimal feature weight set and feature sequences from multiple follow-up time points of the patient. The output is a risk score for the onset of diabetic nephropathy, which is used to dynamically assess the patient's risk trend. The attention-guided dual-stream temporal static fusion network introduces a static feature encoder and a temporal feature encoder, which can simultaneously process baseline static data and follow-up dynamic data.

6. The method according to claim 5, characterized in that, The optimal feature weights are re-executed using newly added case data every preset period to update the weights of each feature factor, enabling the risk prediction model to adapt to changes in the patient population.

7. The method according to claim 5, characterized in that, The methods for determining the risk level in the risk label set include: The disease risk score output by the risk prediction model is compared with a preset dynamic threshold curve, and high-risk, medium-risk, or low-risk intervals are divided according to the relative position of the score curve and the threshold curve. The dynamic threshold curve is adaptively adjusted based on the changes in the quantiles of historical population risk scores over time.

8. The method according to claim 1, characterized in that, Real-time personalized health management plans for patients with type 2 diabetic nephropathy are developed based on real-time personalized label matching across various label sets, using deep metric learning. These plans include: Each solution in the solution library is pre-mapped to the same embedding space as the label combination through a Siamese network. The solution with the highest similarity is selected as the matching result by calculating the cosine similarity between the label combination embedding and the embedding of each solution. The Siamese network is trained using triplet loss, where a triplet consists of an anchor label combination, a positive example scheme, and a negative example scheme, which makes the embedding of effective schemes closer to the label combination and the embedding of invalid schemes farther away.

9. A dynamic health management system for diabetic nephropathy, characterized in that, The system, applicable to the method of any one of claims 1-8, comprises: A tag system construction unit is used to establish and store the multi-dimensional tag set system. The initial plan matching unit is used to generate the patient's current multidimensional personalized tag combination and match it with the initial health management plan; The data monitoring unit is used to monitor and acquire patients' movement data and health indicators in real time; The adaptive update unit has the dynamic adaptive update model built in, which is used to receive feedback data and directly output the updated health management plan. Preference modeling unit, used to learn and store patients' personalized preference information and participate in loss calculation; The risk dynamic assessment unit is used to dynamically classify risk levels based on the risk score curve output by the risk prediction model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images