A chronic kidney disease intelligent follow-up management system and method based on multi-source data

Abstract

The present application relates to the field of medical information technology, and discloses a chronic kidney disease intelligent follow-up management system and method based on multi-source data. The method forms a multi-source heterogeneous data set by collecting the patient's electronic health record, real-time wearable device monitoring data and patient active feedback information. Using a deep learning model, combined with time series data analysis and cross-modal feature alignment module, multi-source data collaborative analysis is realized to accurately assess the patient's kidney function status. According to the evaluation results and individual differences of patients, personalized follow-up strategies are dynamically generated, including follow-up cycle, mandatory items and risk intervention priority. The system automatically performs follow-up tasks, real-time tracks the execution status and feeds back to the evaluation model, forming a closed-loop iterative optimization, which significantly improves the efficiency of chronic kidney disease management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, in particular to a chronic kidney disease intelligent follow-up management system and method based on multi-source data. BACKGROUND

[0002] Traditional chronic kidney disease management relies on regular follow-up by artificial means, and there are problems of fragmented data utilization and lagging risk assessment. Electronic health records, wearable devices and patient complaint data belong to different systems, resulting in a serious information island phenomenon. Existing technologies mostly use single-modal analysis methods, which are difficult to capture dynamic evolution characteristics of kidney function, such as only evaluating through static blood test indicators while ignoring real-time physiological fluctuations and patient behavior data. At the same time, fixed follow-up cycles and uniform examination items cannot adapt to individual differences in disease progression, easily causing over-medical treatment of low-risk patients or insufficient intervention of high-risk patients. In addition, the lack of a closed-loop feedback mechanism makes it difficult to dynamically optimize management strategies according to the execution effect, making it difficult to meet the long-term and heterogeneous management needs of chronic kidney disease. SUMMARY

[0003] In view of the above technical deficiencies, the purpose of the present application is to provide a chronic kidney disease intelligent follow-up management system and method based on multi-source data, which solves the problem of lack of dynamic individualized adjustment and multi-source data collaborative analysis in the prior art.

[0004] To solve the above technical problems, the present application adopts the following technical solutions:

[0005] In a first aspect, the present application provides a chronic kidney disease intelligent follow-up management method based on multi-source data, which comprises:

[0006] Collecting electronic health records, real-time wearable device monitoring data and patient active feedback information of patients to form a multi-source heterogeneous data set;

[0007] The patient active feedback information includes symptom description and medication record;

[0008] Based on the multi-source heterogeneous data set, a deep learning model is used to evaluate the current kidney function status of the patient, and the deep learning model realizes multi-source data collaborative analysis through a time series data analysis module and a cross-modal feature alignment module;

[0009] According to the evaluation results and individual differences of the patient, a personalized follow-up strategy is dynamically generated, which includes follow-up cycle, mandatory examination item and risk intervention priority;

[0010] The personalized follow-up strategy is converted into executable task instructions, and follow-up reminders, examination appointments and patient education content push are automatically triggered, while the task execution status is tracked in real time and fed back to the evaluation model to form a closed-loop iterative optimization.

[0011] Preferably, in a possible implementation form of the first aspect, the step of implementing time series data analysis by the time series data analysis module specifically comprises:

[0012] a hybrid architecture composed of bidirectional long short-term memory network and temporal convolution network is constructed to extract time series features in electronic health records;

[0013] real-time wearable device monitoring data stream is segmented by a sliding window mechanism;

[0014] a self-attention mechanism is introduced to weight and aggregate time series features to generate a state vector representing dynamic evolution of kidney function;

[0015] the state vector is used to quantify the rate of kidney function decline and the probability of acute episode risk.

[0016] Preferably, in a possible implementation form of the first aspect, the step of implementing cross-modal feature alignment by the cross-modal feature alignment module specifically comprises:

[0017] a feature interaction layer based on cross-attention mechanism is adopted to align the state vector output by the time series data analysis module with patient active feedback information;

[0018] a learnable weight distribution matrix is used to calculate the semantic correlation degree of the three modalities of electronic health records, real-time wearable device monitoring data and patient active feedback information;

[0019] a feature mapping layer is adopted to project heterogeneous data to a unified hidden space to eliminate the scale difference of multi-source data.

[0020] Preferably, in a possible implementation form of the first aspect, the cross-modal feature alignment module deploys a multi-modal fusion strategy and adopts a gating mechanism to dynamically adjust the contribution weight of each modality feature to risk assessment;

[0021] combined with a prototype contrast learning algorithm, a prototype cluster of patient state is constructed in the hidden space, and the consistency of cross-modal representation is enhanced by minimizing the multi-modal feature distance of patients of the same type and maximizing the feature distance of patients of different types.

[0022] Preferably, in a possible implementation form of the first aspect, the deep learning model is specifically implemented as:

[0023] the time series data analysis module is used as a feature extraction engine, the cross-modal feature alignment module implements feature collaborative optimization through a multi-modal fusion strategy, and an end-to-end graph neural network architecture is constructed;

[0024] multi-source heterogeneous data of patients is embedded through graph nodes, and historical cases are connected through graph edges, and kidney function state assessment based on graph reasoning is realized.

[0025] The model is optimized by combining self-supervised pre-training and a dynamic weighted loss function, wherein the loss function includes a kidney function prediction error, a complication classification cross-entropy, and a cross-modal alignment regularization term.

[0026] Preferably, in a possible implementation form of the first aspect, the kidney function status evaluation result includes a dynamic indicator quantifying the rate of kidney function decline, an acute episode risk probability value, a complication classification risk level, and a kidney function status score.

[0027] Preferably, in a possible implementation form of the first aspect, the personalized follow-up strategy generation process includes:

[0028] A dynamic optimization model based on a multi-factor decision tree is constructed, inputting the quantitative indicators in the kidney function status evaluation result and patient individual difference features, and outputting follow-up strategy parameters;

[0029] The follow-up strategy parameters include a follow-up cycle, a mandatory inspection item, and a risk intervention priority;

[0030] The follow-up cycle is adaptively calculated by integrating a learning algorithm with the rate of decline and the risk probability, the mandatory inspection item is dynamically selected by a risk classifier based on the complication probability, and the risk intervention priority is generated by an entropy weight method weighting score model;

[0031] The patient individual difference includes age, medical history, and drug compliance.

[0032] Preferably, in a possible implementation form of the first aspect, the adaptive calculation of the follow-up cycle specifically includes:

[0033] A stacked generalization model is constructed to fuse a random forest and a gradient boosting decision tree, wherein the random forest processes high-dimensional features, and the gradient boosting decision tree optimizes prediction based on patient individual difference features;

[0034] The stacked generalization model is trained by K-fold cross-validation, and outputs a dynamic follow-up interval.

[0035] Preferably, in a possible implementation form of the first aspect, the entropy weight method weighting score model is specifically implemented as:

[0036] The rate of kidney function decline, the acute episode risk probability value, and the complication risk level in the kidney function status evaluation result are extracted, and patient individual difference features are combined as input variables;

[0037] A weight distribution mechanism based on information entropy is used to calculate variable weight coefficients to reflect the different influences of the variability of the variables on the risk contribution;

[0038] The variables are normalized to eliminate the dimension and weighted fusion to generate a comprehensive risk score;

[0039] The score is dynamically mapped to a risk intervention priority sequence through an adaptive sorting algorithm, and the weight coefficient is updated in combination with real-time wearable device data.

[0040] In a second aspect, the present application provides a chronic kidney disease intelligent follow-up management system based on multi-source data, which is used to realize the method of the first aspect.

[0041] The data acquisition module acquires the patient's electronic health record, real-time wearable device monitoring data and patient active feedback information to form a multi-source heterogeneous data set.

[0042] The state evaluation module evaluates the patient's current kidney function state based on the multi-source heterogeneous data set using a deep learning model, and the deep learning model realizes multi-source data collaborative analysis through a time series data analysis module and a cross-modal feature alignment module.

[0043] The strategy generation module dynamically generates a personalized follow-up strategy according to the evaluation results and individual differences of the patient, and the personalized follow-up strategy includes a follow-up cycle, a required item and a risk intervention priority.

[0044] The task execution module converts the personalized follow-up strategy into executable task instructions, automatically triggers follow-up reminders, examination appointments and patient education content pushing, and simultaneously tracks the task execution state in real time and feeds back to the evaluation model to form a closed loop iteration optimization.

[0045] The present application has the advantages that: through multi-source heterogeneous data fusion and deep learning model, the dynamic evolution characteristics of kidney function are accurately captured. The time series data analysis module combines bidirectional LSTM and time convolution network to effectively quantify the kidney function decline rate and acute attack risk; the cross-modal feature alignment module eliminates the data scale difference through the cross-attention mechanism to enhance the consistency of multi-modal representation.

[0046] The system can dynamically generate a personalized strategy including a follow-up cycle, a required item and a risk priority, and the closed loop iteration mechanism continuously optimizes the model through task execution state feedback, so that the complication prediction accuracy is improved. This scheme significantly improves the efficiency of chronic kidney disease management and reduces the incidence of acute exacerbation. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0048] Figure 1 A flow chart of a chronic kidney disease intelligent follow-up management method based on multi-source data is provided for the present application.

[0049] Figure 2 A system structure diagram of a chronic kidney disease intelligent follow-up management system based on multi-source data is provided for the present application.

[0050] Legend: 1-data acquisition module, 2-state evaluation module, 3-strategy generation module, 4-task execution module. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] Embodiment one: as shown, the present application provides a chronic kidney disease intelligent follow-up management method based on multi-source data, comprising: Figure 1

[0053] The electronic health records, real-time wearable device monitoring data and patient active feedback information of the patient are collected to form a multi-source heterogeneous data set.

[0054] In the present embodiment, the electronic health records are obtained through a hospital information system interface, including patient history test reports, image data, medication records and outpatient medical history texts, and the data format covers structured tables and unstructured texts. The real-time wearable device monitoring data is accessed to the system through Bluetooth and Internet of Things protocols, and physiological indicators such as blood pressure, heart rate, blood oxygen saturation and daily steps of the patient are dynamically collected, and the sampling frequency is set according to clinical needs. The patient active feedback information is collected by a mobile terminal application program, and the patient's subjective symptom description and medication record are obtained through a pre-set questionnaire form.

[0055] ​To address the heterogeneity of multi-source data, the system first performs data preprocessing. For unstructured text in electronic health records, named entity recognition techniques are used to extract key medical terms and convert them into standardized codes. Wearable device data streams are segmented into time slices using a sliding window mechanism, and abnormal values are calibrated using Kalman filtering. Patient-initiated feedback information is extracted using a natural language processing model to convert it into a structured feature vector.

[0056] The processed multi-source data is stored in a database after timestamp alignment, forming a multi-source heterogeneous dataset with patient ID as the primary key.

[0057] Based on the multi-source heterogeneous dataset, a deep learning model is used to evaluate the current renal function status of the patient. The deep learning model uses a time series data analysis module and a cross-modal feature alignment module to perform collaborative analysis of multi-source data.

[0058] In this embodiment, the time series data analysis module constructs a hybrid architecture composed of a bidirectional long short-term memory network and a temporal convolutional network as the feature extraction engine. Historical test indicators, medication records, and disease course text in electronic health records are input into this hybrid architecture. The bidirectional long short-term memory network captures long-term dependencies, while the temporal convolutional network extracts local temporal patterns using dilated convolution kernels.

[0059] For continuous monitoring data streams generated by real-time wearable devices, a sliding window mechanism is used to segment them into fixed-length segments. Each segment is standardized and then fed into the hybrid architecture for processing.

[0060] In the feature aggregation stage, a self-attention mechanism is used to dynamically weight the extracted time series features, generating a state vector that represents the dynamic evolution of renal function. This state vector contains two key quantitative indicators: the first is the renal function decline rate based on the first derivative of the state vector, which represents the linear or nonlinear change trend of renal function; the second is the acute episode risk probability value calculated by the fully connected layer and the Sigmoid activation function, which reflects the likelihood of short-term deterioration of renal function.

[0061] The cross-modal feature alignment module is responsible for fusing the state vector output by the time series module and the patient-initiated feedback information. First, a feature interaction layer based on cross-attention mechanism is used to align multi-modal features: the time series state vector is used as the query vector, and the patient-initiated feedback information is used as the key-value vector. Through a learnable weight distribution matrix, the semantic correlation between the two is calculated to generate a feedback feature vector aligned with the time feature.

[0062] Meanwhile, the cross-modal feature alignment module calculates the global semantic correlation matrix of the three modalities of electronic health records, real-time monitoring data, and patient active feedback information through another parallel branch, quantifying the consistency of the contribution of different modalities to the same clinical representation. Subsequently, the feature mapping layer projects heterogeneous data into a unified hidden space, and uses a multilayer perceptron to eliminate the scale difference of different modalities of data.

[0063] To further optimize the multi-modal fusion effect, a gating mechanism is used to dynamically adjust the feature weights of each modality: a differentiable gating unit is used to calculate the contribution coefficients of electronic health records, monitoring data, and feedback information in real time according to the context information of the current patient state, and to suppress the influence of noise modalities. In addition, a prototype contrast learning algorithm is used to enhance representation consistency: K patient state prototype clusters (stable, slow progression, and high risk in this embodiment) are constructed in the hidden space, and the model is driven to learn the essential state representation independent of the modalities by minimizing the multi-modal feature distance (Euclidean distance in this embodiment) of the same patients and maximizing the feature distance of different patients.

[0064] The overall architecture of the deep learning model uses an end-to-end graph neural network to implement kidney function state evaluation. The time series data analysis module is used as a feature extraction engine, and the cross-modal feature alignment module uses a multi-modal fusion strategy to optimize feature collaboration. The patient's multi-source heterogeneous data is embedded as a graph node, and each node represents the comprehensive state of a patient. Historical cases are connected using graph edges: the feature similarity between the current patient node and the historical case node is calculated using cosine similarity, and the Top-K similar nodes are retained to construct a dynamic subgraph.

[0065] The graph reasoning process uses a graph attention network to aggregate the state information of neighboring nodes (historical complication development trajectories and intervention effects in this embodiment) to generate an enhanced kidney function state evaluation result. This result includes four dimensions of quantitative output: first, the dynamic indicator of kidney function decline rate, outputting a specific numerical value; second, the acute attack risk probability value, ranging from 0 to 1; third, the complication classification risk level, which outputs the probability distribution of diabetic nephropathy, cardiovascular events, and other complications through a Softmax layer and maps it to low, medium, and high levels; fourth, the comprehensive kidney function state score, which is generated by a fully connected layer integrating the first three indicators into a continuous value ranging from 0 to 100.

[0066] The model training strategy adopts a two-stage optimization method. In the first stage, self-supervised pre-training is performed: a mask recovery task is designed using unlabeled historical data to learn general representations through reconstruction loss. In the second stage, a dynamic weighted loss function is introduced for supervised fine-tuning. The loss function includes three terms: the Huber loss function is used for the prediction error of renal function, providing robust constraints for the regression tasks of decline rate and status score; the cross-entropy loss of complications classification optimizes the multi-label classification performance; the cross-modal alignment regularization term constrains the consistency of features in the hidden space of different modalities by maximizing mutual information. During training, an adaptive weight scheduling algorithm is used to dynamically adjust the loss term weights based on the performance of each task on the validation set, ensuring balanced convergence in multi-objective optimization.

[0067] In this embodiment, when the patient adds new electronic medical records or wearable device data, the time series data analysis module starts incremental calculation to update the state vector; after the patient submits feedback information, the cross-modal feature alignment module responds and refreshes the fused features. The graph neural network automatically reconstructs the patient subgraph every 24 hours, incorporating the latest similar case information to optimize the evaluation results.

[0068] According to the evaluation results and individual differences of patients, personalized follow-up strategies are dynamically generated, including follow-up cycle, mandatory items, and risk intervention priority.

[0069] In this embodiment, the strategy generation process is realized by constructing a multi-factor decision tree dynamic optimization model. The model takes the quantitative indicators in the renal function status evaluation results as input, including the renal function decline rate, acute episode risk probability value, and complication classification risk level, while integrating patient individual difference features, including patient age, medical history, and drug adherence. Through the hierarchical judgment logic of the multi-factor decision tree, the follow-up cycle, mandatory item list, and risk intervention priority sequence are dynamically output as three types of strategy parameters.

[0070] The adaptive calculation of the follow-up cycle is realized by a stacked generalization model. This model integrates random forest algorithm and gradient boosting decision tree algorithm to form a two-level prediction architecture. The first-level random forest model processes high-dimensional medical features, such as analyzing the non-linear relationship between renal function decline rate and acute episode risk probability. When the patient's annual decline in glomerular filtration rate exceeds 5 and the acute episode probability is greater than 0.3, the model automatically identifies it as a high-risk mode.

[0071] The second-level gradient boosting decision tree is optimized and corrected based on the individual difference characteristics of the patient, and further shortens the follow-up interval based on the output of the random forest. The stacked generalization model is trained through K-fold cross-validation, and the early stopping mechanism is used to prevent overfitting. The final generated follow-up period is negatively correlated with the risk level. The follow-up interval of high-risk patients is compressed to 7-14 days, the follow-up interval of medium-risk patients is maintained at 30 days, and the follow-up interval of low-risk patients is extended to 60-90 days. The regularity constraint is used to ensure the clinical rationality of the interval adjustment.

[0072] The dynamic selection of the mandatory inspection items is completed by the risk classifier module. The module generates a list of mandatory inspection items based on the complication classification risk level. For example, when the model detects that the risk level of diabetic nephropathy has increased to medium, the urine protein creatinine ratio detection and fundus examination are added; if the risk probability of cardiovascular events exceeds the threshold, electrocardiogram and myocardial enzyme spectrum detection are included. The risk classifier combines real-time wearable device data streams for incremental updates. For example, if a patient has abnormal blood pressure monitoring for three consecutive days, even if the original complication risk level has not increased, the patient will be temporarily added to the 24-hour ambulatory blood pressure monitoring project.

[0073] The risk intervention priority generation adopts an entropy weight method weighted scoring model. First, the rate of kidney function decline, acute attack risk probability value, and complication risk level are extracted as three core indicators, which together with the patient's individual difference characteristics form the input variable matrix. The information entropy weight distribution mechanism is used to calculate the weight coefficients of each variable: the greater the variation, the higher the weight, reflecting the difference in the contribution of the comprehensive risk.

[0074] After all variables are normalized by the maximum and minimum to eliminate dimensional differences, the weight coefficients are weighted and fused to generate a comprehensive risk score of 0 to 100 points. Finally, the adaptive sorting algorithm is used to map the score to an intervention priority sequence. Patients with a score greater than or equal to 80 points are included in the emergency intervention group, patients with a score greater than or equal to 60 but less than 80 points are included in the priority processing group, and patients with a score less than 60 points are included in the routine follow-up group. The model updates the weight coefficients every 24 hours combined with the latest wearable device data to ensure the dynamic accuracy of the priority sorting.

[0075] The personalized follow-up strategy is converted into executable task instructions, automatically triggering follow-up reminders, examination appointments, and patient education content pushing, while real-time tracking of task execution status and feedback to the evaluation model form a closed-loop iterative optimization.

[0076] In this embodiment, the personalized follow-up strategy output by the strategy generation module is converted into a structured instruction set by the task analysis engine. The follow-up period parameter triggers the timing task scheduler, which pushes a follow-up reminder message to the patient's mobile terminal at the preset time point. The message content includes the clinic location, doctor information, and precautions. The mandatory inspection item parameter drives the hospital information system interface to automatically generate examination orders, matches the optimal examination period through the appointment algorithm, and synchronizes the appointment voucher to the patient terminal and department terminal.

[0077] Risk intervention priority instructions activate a multi-level response mechanism. High-priority tasks trigger immediate voice alarms and SMS notifications, medium-priority tasks are integrated into the patient's daily to-do list, and low-priority tasks are incorporated into periodic health education plans. Patient education content is dynamically generated based on a knowledge graph engine. According to the current risk assessment results and the patient's historical feedback preferences, video explanations, illustrated manuals, and interactive Q&A materials are automatically combined and pushed to mobile terminals in batches according to risk level.

[0078] Task execution status is tracked in real time through a distributed log system, recording events such as instruction issuance time, patient confirmation time, examination completion rate, and educational content access rate, and annotating abnormal status codes such as appointment overdue, unread reminders, or missing data. The tracked data is cleaned by a stream processing engine to generate an execution performance report, which is injected into the incremental training dataset of the status assessment module through a feedback interface. When the same patient experiences two consecutive task overdues or key physiological indicators deviate from the predicted trajectory, the system automatically triggers a recalculation process for the assessment model, dynamically updating the patient's status assessment results using the latest execution status data, completing a closed-loop optimization iteration.

[0079] Example 2: Figure 2 As shown, this invention provides an intelligent follow-up management system for chronic kidney disease based on multi-source data, comprising:

[0080] Data acquisition module 1 collects patients' electronic health records, real-time wearable device monitoring data, and patients' proactive feedback information to form a multi-source heterogeneous dataset;

[0081] Status assessment module 2, based on multi-source heterogeneous datasets, uses a deep learning model to assess the patient's current renal function status. The deep learning model achieves collaborative analysis of multi-source data through a time-series data analysis module and a cross-modal feature alignment module.

[0082] Strategy generation module 3 dynamically generates personalized follow-up strategies based on assessment results and individual patient differences. Personalized follow-up strategies include follow-up period, mandatory test items, and risk intervention priorities.

[0083] Task execution module 4 transforms personalized follow-up strategies into executable task instructions, automatically triggering follow-up reminders, examination appointments, and patient education content pushes. At the same time, it tracks the task execution status in real time and feeds it back to the evaluation model to form a closed-loop iterative optimization.

[0084] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent follow-up management of chronic kidney disease based on multi-source data, characterized in that, The method includes: Collect patients' electronic health records, real-time wearable device monitoring data, and patients' proactive feedback information to form a multi-source heterogeneous dataset; The patient's proactive feedback information includes symptom descriptions and medication records; Based on the multi-source heterogeneous dataset, a deep learning model is used to assess the patient's current renal function status. The deep learning model achieves multi-source data collaborative analysis through a time-series data analysis module and a cross-modal feature alignment module. The time series data analysis module performs time series data analysis in the following specific steps: A hybrid architecture consisting of a bidirectional long short-term memory network and a temporal convolutional network was constructed to extract temporal features from electronic health records. Bidirectional long short-term memory networks capture long-term dependencies, while temporal convolutional networks extract local temporal patterns through dilated convolutional kernels. Real-time wearable device monitoring data streams are segmented using a sliding window mechanism; A self-attention mechanism is introduced to weighted aggregate temporal features, generating a state vector representing the dynamic evolution of kidney function. The state vector is used to quantify the rate of renal function decline and the probability of acute exacerbation. The specific steps for the cross-modal feature alignment module to achieve cross-modal feature alignment include: A feature interaction layer based on the cross-attention mechanism is adopted to align the state vector output by the time series data analysis module with the patient's active feedback information. The semantic relevance of three modalities—electronic health records, real-time wearable device monitoring data, and patient proactive feedback information—is calculated using a learnable weighting matrix. A feature mapping layer is used to project heterogeneous data onto a unified latent space, eliminating scale differences between multi-source data. The cross-modal feature alignment module deploys a multimodal fusion strategy and uses a gating mechanism to dynamically adjust the contribution weight of each modal feature to risk assessment. By combining prototype contrastive learning algorithms, a prototype cluster of patient states is constructed in the latent space. By minimizing the multimodal feature distance of patients of the same type and maximizing the feature distance of patients of different types, the consistency of cross-modal representations is enhanced. The deep learning model is specifically implemented as follows: Using the time-series data analysis module as the feature extraction engine, the cross-modal feature alignment module achieves feature collaborative optimization through a multi-modal fusion strategy, thus constructing an end-to-end graph neural network architecture; By embedding multi-source heterogeneous patient data into graph nodes and connecting historical cases using graph edges, a graph-based reasoning-based assessment of renal function status can be achieved. The results of the renal function status assessment include: dynamic indicators that quantify the rate of renal function decline, probability values ​​of acute exacerbation risk, risk levels of complication classification, and renal function status scores. The model is jointly optimized using self-supervised pre-training and a dynamic weighted loss function, where the loss function includes renal function prediction error, complication classification cross-entropy, and cross-modal alignment regularization term; Based on the assessment results and individual patient differences, a personalized follow-up strategy is dynamically generated. The personalized follow-up strategy includes the follow-up period, mandatory examination items, and risk intervention priority. The personalized follow-up strategy generation process includes: A dynamic optimization model based on a multi-factor decision tree is constructed, which takes into account the quantitative indicators in the renal function status assessment results and the individual differences of patients, and outputs the follow-up strategy parameters. Individual patient differences include age, medical history, and medication adherence; Follow-up strategy parameters include follow-up period, mandatory testing items, and risk intervention priority; Among them, the follow-up period is adaptively calculated by combining the decline rate and risk probability through an ensemble learning algorithm, the mandatory test items are dynamically selected by the risk classifier based on the probability of complications, and the risk intervention priority is generated by an entropy weighted scoring model. The adaptive calculation of the follow-up period specifically includes: A stacked generalization model was constructed that integrates random forest and gradient boosting decision tree, where random forest handles high-dimensional features and gradient boosting decision tree optimizes predictions based on individual patient differences. The stacked generalization model is trained using K-fold cross-validation and outputs a dynamic follow-up interval. The entropy weighted scoring model is specifically implemented as follows: The rate of renal function decline, the probability value of acute exacerbation risk, and the risk level of complications in the renal function status assessment results were extracted and combined with the individual differences of patients as input variables. A weight allocation mechanism based on information entropy is adopted to calculate the weight coefficients of each variable to reflect the differential impact of their variability on risk contribution; After normalizing the variables to eliminate dimensions, a weighted fusion is performed to generate a comprehensive risk score. The scores are dynamically mapped to a risk intervention priority sequence through an adaptive sorting algorithm, and the weight coefficients are updated in combination with real-time wearable device data. The personalized follow-up strategy is transformed into executable task instructions, which automatically trigger follow-up reminders, examination appointments, and push patient education content. At the same time, the task execution status is tracked in real time and fed back to the evaluation model to form a closed-loop iterative optimization.

2. A smart follow-up management system for chronic kidney disease based on multi-source data, characterized in that, The system is used to implement the intelligent follow-up management method for chronic kidney disease based on multi-source data as described in claim 1, including: The data acquisition module collects patients' electronic health records, real-time wearable device monitoring data, and patients' proactive feedback information to form a multi-source heterogeneous dataset; The status assessment module, based on the multi-source heterogeneous dataset, uses a deep learning model to assess the patient's current renal function status. The deep learning model achieves multi-source data collaborative analysis through a time-series data analysis module and a cross-modal feature alignment module. The strategy generation module dynamically generates personalized follow-up strategies based on the assessment results and individual patient differences. The personalized follow-up strategies include the follow-up period, mandatory examination items, and risk intervention priorities. The task execution module transforms the personalized follow-up strategy into executable task instructions, automatically triggering follow-up reminders, examination appointments, and patient education content pushes. At the same time, it tracks the task execution status in real time and feeds it back to the evaluation model to form a closed-loop iterative optimization.

Images