Neural network prediction model-based kidney disease patient health management method and system

Abstract

The invention relates to a nephrotic patient health management method and system based on a neural network prediction model, and the method comprises the steps: obtaining the multi-dimensional health data of a target nephrotic patient at each stage, constructing a time sequence feature matrix to extract the change trend of the whole course of disease, and dividing the course of disease into a disease alleviation stage and a disease deterioration stage; analyzing the health data of the two stages to obtain an improvement factor and a deterioration factor; outputting a renal function state prediction result by using a neural network prediction model fusing an attention mechanism and a knowledge graph, and generating an initial health management scheme by combining the result, a preset health management rule base and patient preferences; and optimizing the initial scheme based on the improvement factor, the deterioration factor and a reinforcement learning algorithm to obtain an optimized health management scheme. Complete-cycle closed-loop management of data acquisition, trend analysis, factor identification, prediction and early warning, scheme generation and dynamic optimization of nephropathy patients is realized, and the accuracy and individuation level of nephropathy health management are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of patient health management, and in particular relates to a health management method and system for kidney disease patients based on a neural network prediction model. Background Art

[0002] Chronic kidney disease is characterized by a long course, insidious progression, and wide individual variability. Its renal impairment is often progressive and, without precise management, can easily progress to end-stage renal disease, placing a heavy burden on patients' quality of life and the healthcare system. Current kidney health management primarily relies on empirical medical guidance, which presents the following core issues: First, health data is underutilized. Existing management methods often rely on clinical indicators at a single point in time, failing to integrate data related to patients' long-term lifestyle behaviors (diet, exercise) and medical interventions (medication, dialysis), making it difficult to capture the dynamic patterns of disease progression. Furthermore, data preprocessing lacks specificity, using a uniform approach for data at different stages of the disease, which obscures key features. Second, the lack of correlation between disease stage classification and factors. Traditional disease course assessments are often based on static thresholds for a single indicator, making it impossible to distinguish natural fluctuations in the disease from true remission or worsening trends. Furthermore, a lack of a stage-factor correlation mechanism makes it difficult to clearly identify key improvement factors during remission (such as dietary control and regular medication) and risk factors during worsening, resulting in a lack of precise targeting of interventions. Third, the prediction model lacks integration with clinical logic. Existing renal function prediction models often rely on a single data type, failing to integrate medical knowledge and individual differences. This results in poorly interpretable predictions and struggles to support personalized intervention decisions. Furthermore, the model lacks a dynamic update mechanism, making it unable to adapt to long-term changes in a patient's condition. Fourth, health management plans lack dynamic optimization capabilities. Current plans are often generated based on fixed clinical pathways, failing to fully consider patient preferences and feasibility, leading to insufficient compliance. Plan adjustments also rely on manual evaluation, failing to dynamically optimize based on real-time health data and predictions. To address the above issues, the present invention intends to construct a method that integrates multidimensional data, fuses medical knowledge and artificial intelligence technology, and realizes dynamic and precise management. By quantifying the disease course, identifying key factors, intelligent prediction and closed-loop optimization, the accuracy and effectiveness of kidney disease management will be improved. Summary of the Invention

[0003] The purpose of the present invention is to provide a health management method and system for kidney disease patients based on a neural network prediction model, so as to improve the accuracy and effectiveness of kidney disease management through quantification of disease stages, identification of key factors, intelligent prediction and closed-loop optimization.

[0004] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows: In a first aspect, a method for health management of kidney disease patients based on a neural network prediction model is provided, comprising the following steps: S1: Obtain multidimensional health data of target kidney disease patients at various stages, preprocess the data, and obtain the disease condition change trend of the target kidney disease patients throughout the entire disease course based on the preprocessed multidimensional health data; S2: The entire course of the disease is divided into different stages based on the trend of disease changes. Different stages are divided into two categories, including remission stage and exacerbation stage; S3: Obtain corresponding multidimensional health data based on the remission stage and the exacerbation stage, obtain an improvement factor based on the analysis of the multidimensional health data in the remission stage, and obtain a deterioration factor based on the analysis of the multidimensional health data in the exacerbation stage; S4: Build a neural network prediction model that integrates the attention mechanism and knowledge graph, input preprocessed multi-dimensional health data after training, and output the predicted renal function status of the target kidney disease patient during the specified period; S5: Based on the predicted results of renal function status, the preset health management rule library is called and combined with the preferences of the target patient to generate an initial health management plan; S6: Based on the improvement factors and deterioration factors, the initial health management plan is dynamically optimized through the reinforcement learning algorithm to generate an optimized health management plan.

[0005] The optimization also includes a dynamic feedback process, as follows: S7: Execute the optimized health management plan for the target kidney disease patient, and collect multi-dimensional health data of the target kidney disease patient within a specified time period starting from the execution time node; S8: obtaining the condition change trend of the target kidney disease patient within the specified time period based on the multi-dimensional health data within the specified time period; S9: Based on the trend of disease changes in the specified time period, determine whether it belongs to the remission stage or the deterioration stage. If it is the remission stage, continue to execute the optimized health management plan. If it is the deterioration stage, re-analyze the corresponding multi-dimensional health data based on the disease change trend of the specified stage to obtain the feedback deterioration factor. Based on the feedback deterioration factor, further optimize the optimized health management plan and execute it on the target kidney disease patient. After execution, iteratively execute steps S7-S9 to achieve iterative optimization of the optimized health management plan.

[0006] Preferably, the multidimensional health data includes designated clinical indicator data, medical intervention record data, diet data, associated impact data, and exercise data. The specific process of preprocessing the multidimensional health data in step S1 is as follows: S11: dividing the multidimensional health data into continuous indicator data, classification data, and time series data according to the first division dimension, and dividing the multidimensional health data into initial diagnosis data, treatment period data, and follow-up period data according to the second division dimension; S12: Z-score normalization is used for continuous indicator data, one-hot encoding is performed on categorical data, and time series data is uniformly converted into time series with weekly units; S13: A stage-by-stage adaptive preprocessing strategy was used to process missing values ​​for data at different disease stages. S14: Through the 3σ principle combined with clinical rationality verification, true abnormal values ​​are distinguished from false abnormal values. For true abnormal values, the original value is retained and the cause of the abnormality is marked. For false abnormal values, the moving average method is used to correct them.

[0007] Preferably, the specific process of obtaining the disease condition change trend of the target kidney disease patient throughout the course of the disease based on the pre-processed multi-dimensional health data in step S1 is as follows: S15: Constructing a time series feature matrix: Arrange the preprocessed data along the time axis to generate a feature matrix with time-index as the dimension; S16: Extract trend features: Calculate the slope of change of each specified indicator within a continuous time window, and generate a smooth trend line by combining spline curve fitting; determine the indicator's rising and falling trend through the first-order derivative of the trend line, and determine the trend change rate through the second-order derivative; S17: Perform multi-indicator fusion: Perform dimensionality reduction processing on the multi-indicator trends, extract the top three principal components with the highest contribution rate, and construct a comprehensive disease trend index.

[0008] Preferably, the specific process of step S2 is as follows: S21: Determine the stage division indicators: use the comprehensive disease trend index as the basic indicator, combine the changing trends of designated key clinical indicators as auxiliary verification indicators, and set the weights of each indicator; S22: Setting stage division thresholds: including thresholds for determining the stage of disease remission, thresholds for determining the stage of disease exacerbation, and transition stage processing; S23: Perform dynamic window division to obtain the remission stage and the exacerbation stage, and classify the short-term indicator fluctuations that do not meet the duration requirements into adjacent major stages; S24: Verify and revise the stage boundaries: Verify the preliminary stage boundaries based on clinical event records and introduce stage consistency verification rules for revision; S25: Generate a disease stage division report, including the time interval of each stage, the core indicator change curve, the basis for stage division and the annotation of key clinical events.

[0009] Preferably, the specific process of step S3 is as follows: S31: extracting time interval data corresponding to the remission stage and the exacerbation stage from the multidimensional health data of the entire disease course, forming two independent data sets including a remission stage data set and a exacerbation stage data set; S32: For the dataset of the remission stage, the temporal relationship between intervention measures and indicator changes is retained, and a time alignment algorithm is used to match the execution time of intervention measures with the indicator change time; S33: Perform improvement factor analysis to screen potential related factors from the data of the remission stage and form a candidate improvement factor pool; use bivariate association analysis combined with multivariate regression model to perform improvement factor correlation analysis and improvement factor classification: S34: Perform exacerbation factor analysis and candidate exacerbation factor extraction. Screen potential risk factors from the disease exacerbation stage data to form a candidate exacerbation factor pool; use causal inference combined with risk accumulation model to perform exacerbation factor correlation analysis and exacerbation factor classification; S35: Generate a factor analysis report, including a list of improvement factors, a list of deterioration factors, and a factor correlation graph.

[0010] Preferably, the specific process of step S4 is as follows: S41: Constructing a knowledge graph for kidney disease: Construct a knowledge graph containing core entities and associated entities, define entity relationships, convert entities and relationships in the knowledge graph into low-dimensional vectors, and generate a knowledge graph embedding vector; S42: The input layer of the neural network prediction model receives the preprocessed multi-dimensional health data and loads the knowledge graph embedding vector as auxiliary input; S43: performing feature extraction to obtain a feature vector; S44: The feature vector output by the feature extraction layer is integrated with the knowledge graph embedding vector through a gated fusion mechanism. The fusion ratio of the two types of features is dynamically adjusted through the gated unit. After fusion, a data-knowledge joint feature vector is generated. S45: An improved bidirectional long short-term memory network is used to capture the long-term temporal dependencies of the joint feature vectors. A temporal attention mechanism is introduced to assign higher temporal weights to recent data, and the prediction results of renal function status are output separately: including continuous value prediction and classification prediction; Preferably, the specific process of step S5 is as follows: S51: Rule base matching and candidate solution generation: Matching basic intervention framework: Call corresponding rules based on the predicted results of renal function status to generate a basic intervention framework; Expand the candidate solution set: For each intervention item in the basic framework, retrieve 3-5 alternative solutions from the rule base to obtain candidate solutions; S52: Patient preference adaptation and program screening: Preference matching calculation: For each candidate plan, calculate the matching score with the patient's preferences. Screening the best candidate plan: select the candidate plan with a matching degree ≥80 points. If there are multiple plans that meet the standards, select a comprehensive plan based on the expected intervention effect.

[0011] S53: Integrate the screened candidate solutions into an initial health management plan.

[0012] S54: Check the rationality of the initial health management plan.

[0013] Preferably, the specific process of step S6 is as follows: S61: Define the core elements of reinforcement learning and map the state space to the action space: S62: Perform initial adjustments directed towards improvement factors; S63: Make risk aversion adjustments guided by deterioration factors; S64: Iteratively optimize the initial solution based on reinforcement learning; S65: Check the optimization action constraints: After each round of optimization, double verification is performed: clinical compliance verification and factor synergy verification; The optimization parameters output by reinforcement learning are injected into the initial health management plan to form an optimized health management plan that includes core optimization items and auxiliary optimization items.

[0014] In a second aspect, a health management system for kidney disease patients based on a neural network prediction model is provided, which is used to implement the health management method for kidney disease patients based on the neural network prediction model, including a data acquisition module, a preprocessing module, a disease course classification module, a disease condition factor analysis module, a neural network prediction model, a health management plan generation module, and a plan optimization module; A data acquisition module is used to obtain multi-dimensional health data of target kidney disease patients at various stages; Preprocessing module, used to preprocess multidimensional health data; The disease course classification module is used to obtain the disease course change trend of the target kidney disease patient throughout the disease course based on the pre-processed multi-dimensional health data. The disease course is divided into different stages based on the disease course change trend. The different stages are divided into two categories: remission stage and exacerbation stage. A disease course classification module is used to obtain corresponding multidimensional health data based on the remission stage and the exacerbation stage, obtain improvement factors based on the analysis of the multidimensional health data in the remission stage, and obtain exacerbation factors based on the analysis of the multidimensional health data in the exacerbation stage; A neural network prediction model is used to output a prediction result of the renal function status of a target kidney disease patient during a specified period based on the input of pre-processed multi-dimensional health data; A neural network prediction model is used to call a preset health management rule library based on the predicted results of renal function status and combine it with the preferences of the target patient to generate an initial health management plan; The solution optimization module is used to dynamically optimize the initial health management plan based on improvement factors and deterioration factors through reinforcement learning algorithms to generate an optimized health management plan.

[0015] The beneficial effects of the present invention include: The neural network-based predictive model-based health management method and system for kidney disease patients, provided by this invention, first utilizes a phased adaptive preprocessing strategy to differentiate data processing for different disease stages. Multidimensional health data is standardized by type and a time series feature matrix is ​​constructed, providing a high-quality data foundation for subsequent disease trend analysis and prediction, addressing the issues of data clutter and the obscuration of key features in traditional management.

[0016] Secondly, using a comprehensive disease trend index as the core, combined with the changing trends of key clinical indicators, the system divides the disease into remission and exacerbation stages. This avoids the limitations of a single static indicator and accurately captures the true changing trends of the disease. Furthermore, improvement and exacerbation factors are extracted from data at different stages, clarifying the key drivers of disease changes. This provides precise targeting for subsequent interventions and addresses the lack of targeted interventions in traditional management.

[0017] Thirdly, the neural network prediction model, which integrates attention mechanisms and knowledge graphs, not only integrates multidimensional health data but also incorporates medical knowledge and logic. This attention mechanism strengthens the influence of key features, improving prediction accuracy. Furthermore, the model outputs clear trends in renal function and risk levels, providing a scientific basis for generating health management plans and resolving the issues of traditional prediction models being disconnected from clinical logic and lacking interpretability.

[0018] Thirdly, the initial health management plan is generated by combining renal function prediction results, a health management rule base, and patient preferences. This plan not only meets clinical standards but also takes into account individual patient needs such as dietary restrictions and exercise preferences, thereby improving plan compliance. Furthermore, a reinforcement learning algorithm dynamically optimizes the plan based on improvement and deterioration factors, further improving its adaptability and addressing the lack of personalization and low compliance associated with traditional plans.

[0019] Finally, through a dynamic feedback process, health data after the plan is executed is collected in real time to determine the trend of the disease and iterate the plan for optimization. If the disease worsens, the plan can be quickly adjusted based on the feedback of the deterioration factor, achieving a full cycle of closed-loop management: data collection - trend analysis - prediction - plan generation - optimization - feedback. This solves the problems of delayed plan adjustment and lack of dynamic response in traditional management. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the health management method for kidney disease patients based on a neural network prediction model of the present invention.

[0021] Figure 2 Schematic diagram of the architecture of the neural network prediction model of the present invention.

[0022] Figure 3 Schematic diagram of the architecture of the bidirectional long short-term memory network of the present invention. DETAILED DESCRIPTION

[0023] The following is combined with Figures 1 to 3 The present invention is described in further detail: Example 1 See attached Figure 1 As shown, the health management method for kidney disease patients based on the neural network prediction model includes the following steps: S1: Obtain multidimensional health data of target kidney disease patients at various stages, preprocess the data, and obtain the disease condition change trend of the target kidney disease patients throughout the entire disease course based on the preprocessed multidimensional health data; S2: The entire course of the disease is divided into different stages based on the trend of disease changes. Different stages are divided into two categories, including remission stage and exacerbation stage; S3: Obtain corresponding multidimensional health data based on the remission stage and the exacerbation stage, obtain an improvement factor based on the analysis of the multidimensional health data in the remission stage, and obtain a deterioration factor based on the analysis of the multidimensional health data in the exacerbation stage; S4: Build a neural network prediction model that integrates the attention mechanism and knowledge graph, input preprocessed multi-dimensional health data after training, and output the predicted renal function status of the target kidney disease patient during the specified period; S5: Based on the predicted results of renal function status, the preset health management rule library is called and combined with the preferences of the target patient to generate an initial health management plan; S6: Based on the improvement factors and deterioration factors, the initial health management plan is dynamically optimized through the reinforcement learning algorithm to generate an optimized health management plan.

[0024] It also includes a dynamic feedback process, as follows: S7: Execute the optimized health management plan for the target kidney disease patient, and collect multi-dimensional health data of the target kidney disease patient within a specified time period starting from the execution time node; S8: obtaining the condition change trend of the target kidney disease patient within the specified time period based on the multi-dimensional health data within the specified time period; S9: Based on the trend of disease changes in the specified time period, determine whether it belongs to the remission stage or the deterioration stage. If it is the remission stage, continue to execute the optimized health management plan. If it is the deterioration stage, re-analyze the corresponding multi-dimensional health data based on the disease change trend of the specified stage to obtain the feedback deterioration factor. Based on the feedback deterioration factor, further optimize the optimized health management plan and execute it on the target kidney disease patient. After execution, iteratively execute steps S7-S9 to achieve iterative optimization of the optimized health management plan.

[0025] Example 2 Based on Example 1, the multidimensional health data includes designated clinical indicator data, medical intervention record data, diet data, associated impact data, and exercise data. The specific process of preprocessing the multidimensional health data in step S1 is as follows: S11: Divide the multidimensional health data into continuous indicator data, categorical data, and time series data according to the first division dimension, wherein the continuous indicator data includes blood creatinine, blood sugar, etc. Divide the multidimensional health data into initial diagnosis data, treatment period data, and follow-up period data according to the second division dimension; S12: Z-score normalization was used for continuous indicator data, one-hot encoding was performed on categorical data including medication types in medical intervention record data, and time series data was uniformly converted to weekly time series. S13: A staged adaptive preprocessing strategy is then used to process missing values ​​for data at different disease stages: the migration imputation method based on similar cases is used for the data at the initial diagnosis period, the adjacent time point interpolation method is used for the data during the treatment period, and the dynamic window imputation method is used for the data during the follow-up period; S14: Combined with the 3σ principle for clinical rationality verification, such as whether a sudden increase in blood creatinine is accompanied by an acute infection record, distinguish true abnormal values ​​(acute kidney injury) from false abnormal values ​​(detection errors), retain the original value of the true abnormal value and mark the cause of the abnormality, and use the moving average method to correct the false abnormal value.

[0026] The specific process of obtaining the target kidney disease patient's condition change trend throughout the course of the disease based on the pre-processed multi-dimensional health data in step S1 is as follows: S15: Constructing a time series feature matrix: Arrange the preprocessed data along the time axis to generate a feature matrix with time-indicator as the dimension, where the time granularity is adaptively adjusted according to the length of the disease course, for example, it can be days for the short term and weeks for the long term; S16: Extract trend features: Use a sliding window algorithm to calculate the slope of change of each specified indicator within a continuous time window, and generate a smooth trend line by combining spline curve fitting; use the first-order derivative of the trend line to determine the indicator's upward and downward trend, and use the second-order derivative to determine the trend change rate; S17: Perform multi-indicator fusion: Perform dimensionality reduction on multi-indicator trends based on principal component analysis, extract the top three principal components with the highest contribution rate, and construct a comprehensive disease trend index. The direction and magnitude of change in the comprehensive disease trend index are used to characterize the overall disease trend.

[0027] The specific process of step S2 is as follows: S21: Determine the indicators for stage division: use the comprehensive disease trend index as the basic indicator, combine the changing trends of designated key clinical indicators including eGFR, serum creatinine, and urine protein quantification as auxiliary verification indicators, and set the weights of each indicator; S22: Setting the stage division threshold: The thresholds for determining the remission stage are: the comprehensive disease trend index continues to rise with an increase of ≥15%; the eGFR continues to rise or stabilize, with a fluctuation of ≤5%; the serum creatinine continues to fall with a decrease of ≥10%; the urine protein quantity continues to fall or remains within the normal range, i.e., <0.15g / 24h. The above indicators must meet the basic indicators of the comprehensive disease trend index and at least two auxiliary indicators meet the standards; The threshold for determining the stage of disease worsening is: the comprehensive disease trend index continues to decline and the decline is ≥15%; the eGFR continues to decline and the decline is ≥10%; the serum creatinine continues to rise and the increase is ≥20%; the urine protein quantity continues to rise and exceeds the normal range by 20% or more. The core indicators must also meet the standards and at least two auxiliary indicators must meet the standards; Transitional stage treatment: If the indicator change does not reach any of the above thresholds, for example, the fluctuation range of the comprehensive disease trend index is less than 15%, it is determined to be in the transitional stage and is not temporarily classified as remission or worsening. Continue to monitor until the corresponding threshold conditions are met before classification; S23: Dynamic window division: A sliding time window with a window length of 4 weeks and a step length of 1 week was used to scan the time series data of the entire disease course. When the indicator changes in the window met the judgment threshold for the remission stage and the duration of this state was ≥2 consecutive windows, the time interval corresponding to the window was marked as the remission stage; when the indicator changes in the window met the judgment threshold for the worsening stage and the duration was ≥2 consecutive windows, it was marked as the worsening stage.

[0028] For short-term indicator fluctuations, if the indicator rebounds or decreases for 1-2 weeks and does not meet the duration requirement, it will not be divided into a separate stage and will be included in the adjacent major stage. For example, if there is a small increase in eGFR for 1 week during the overall worsening trend, that week will still be included in the worsening stage and marked as a short-term fluctuation within the worsening stage.

[0029] S24: Stage boundary verification and correction: The preliminary stage boundaries were verified in combination with clinical event records. If the stage division results were inconsistent with known clinical events, such as when the patient was determined to be in the remission stage but there was a record of renal function damage caused by acute infection during this period, the stage boundaries were adjusted based on the time point of the clinical event.

[0030] At the same time, a stage consistency verification rule was introduced: the boundary between two adjacent stages must have a clear indicator turning point. For example, the first-order derivative of the comprehensive disease trend index changes from negative to positive to mark the beginning of the remission stage, and from positive to negative to mark the beginning of the exacerbation stage. If there is no clear turning point at the boundary, the window range is expanded and recalculated to ensure that the stage division conforms to the actual turning points of the disease progression.

[0031] S25: Generate a disease stage report, including the time intervals for each stage, the core indicator change curve, the basis for stage division, and key clinical event annotations. The report must clearly distinguish between the primary stage (lasting ≥12 weeks) and the secondary stage (lasting 8-11 weeks). The secondary stage must be marked with "requires continuous monitoring" to indicate that its stability needs to be paid special attention to during subsequent health management.

[0032] Example 3 Based on Example 1 or Example 2, the specific process of step S3 is as follows: S31: extracting time interval data corresponding to the remission stage and the exacerbation stage from the multidimensional health data of the entire disease course, forming two independent data sets including a remission stage data set and a exacerbation stage data set; S32: For the dataset of the remission stage, the temporal relationship between intervention measures and indicator changes is retained, including the temporal correspondence between blood pressure changes after adjusting antihypertensive drugs, and a time alignment algorithm is used to match the execution time of intervention measures with the indicator change time; The dose-effect relationship annotation of risk factors in the dataset of the disease exacerbation stage was strengthened, including the correspondence between the number of days with excessive salt intake and the increase in blood creatinine. Through the time series segmentation method, the persistent risk factors were accurately matched with the period of indicator exacerbation. S33: Perform improvement factor analysis to screen potential related factors from the data of the remission stage and form a candidate improvement factor pool, including: Treatment-related factors: including medication adjustment plans for increasing the dose of ACEI drugs and optimization of dialysis parameters to improve dialysis adequacy; Behavioral-related factors: daily protein intake controlled within 0.8g / kg, daily exercise duration ≥30 minutes, and salt intake ≤5g / day; Clinical management factors: including timely control of infection with decreased inflammatory indicators after antibiotic use, and achievement of comorbidity targets including HbA1c <7%; Constitution-related factors: Changes in tongue appearance corresponding to improvement of Qi deficiency in TCM syndrome differentiation and increased abundance of probiotics in intestinal flora; Bivariate correlation analysis combined with multivariate regression model was used to analyze the correlation between improvement factors: Temporal correlation analysis: The cross-correlation function (CCF) was used to calculate the time-lagged correlation between candidate factors and core indicators (including eGFR). Factors that were significantly associated with improvement in indicators within 2-4 weeks after intervention were screened, such as factors associated with an increase in eGFR in the third week after dietary adjustment. Dose-effect analysis: For continuous factors including exercise duration, segmented regression was used to calculate the quantitative relationship between different dose intervals and the magnitude of improvement in indicators. For example, the improvement in eGFR when exercising ≥5 times per week was 1.8 times that of exercising <3 times per week. Multivariate regression validation: A multivariate linear regression model was constructed with the eGFR change rate as the dependent variable and the candidate improvement factors as the independent variables, and significant correlation factors with P < 0.05 were screened; Verification of effectiveness of improvement factors: Intra-group validation: Divide the remission phase data into a training set and a validation set. Reproduce the effects of factors selected in the training set in the validation set, and retain factors with a correlation of ≥0.7 in the validation set. Clinical rationality verification: Based on the guidelines for the diagnosis and treatment of kidney disease, factors with statistical significance but no clear clinical mechanism were eliminated, including irrelevant behaviors with accidental associations. Factors consistent with the intervention-improvement pathological mechanism were retained, including the protective effect of blood pressure reduction on renal function; Improvement factor classification: According to the contribution ranking, the improvement factors are divided into core factors (contribution rate ≥ 30%), important factors (contribution rate 15%-30%) and auxiliary factors (contribution rate <15%): Core improvement factors: low-protein diet implementation when eGFR is less than 60ml / min, and blood pressure controlled within 130 / 80mmHg after standardized use of antihypertensive drugs; Important improvement factors: exercise ≥5 times a week and intestinal probiotic supplementation; Auxiliary improvement factors: sleep duration ≥ 7 hours and decreased psychological stress score.

[0033] S34: Perform deterioration factor analysis, Extraction of candidate exacerbation factors: Screen potential risk factors from the data of disease exacerbation stage to form a pool of candidate exacerbation factors, including: Risk behavior factors: salt intake > 6 g / day, daily water intake > 2000 ml (renal insufficiency stage), smoking ≥ 3 times / day; Treatment-related factors: including the use of nephrotoxic drugs such as nonsteroidal anti-inflammatory drugs, unauthorized discontinuation of antihypertensive drugs, and irregular dialysis, including dialysis frequency less than 3 times per week; Clinical risk factors: acute infection including pneumonia leading to elevated inflammatory markers, blood pressure fluctuations including pneumonia leading to elevated inflammatory markers, and hyperkalemia including serum potassium >5.5mmol / L; Environmental factors: including persistent haze with PM2.5 > 100 μg / m³ and insufficient water intake in high temperature environments; Causal inference combined with risk accumulation model is used to analyze the correlation between deterioration factors: Temporal causal analysis: Granger causality test is used to determine whether the risk factor is the antecedent cause of the indicator deterioration, such as excessive salt intake occurring 2 weeks before the increase in blood creatinine, and other interfering factors are excluded; Cumulative Risk Analysis: A cumulative risk factor scoring model was constructed to calculate the probability of renal function worsening when a single factor persisted or when multiple factors were combined. Key factors with a cumulative risk greater than 50% were identified. The persistent presence of a single factor included excessive salt intake for seven consecutive days, while the combined presence of multiple factors included high salt intake, hypertension, and infection.

[0034] Dose-effect verification: Through curve fitting, the quantitative relationship between the intensity of factor exposure and the degree of deterioration was clarified, including the corresponding relationship between every 0.5mmol / L increase in blood potassium and a 2ml / min decrease in eGFR.

[0035] Verify the validity of the deterioration factor: Reverse validation: Substitute the screened exacerbation factors into the remission stage data to verify whether they are in a low exposure state during the remission stage, including whether the intake of the exacerbation factor "high salt intake" in the remission stage is significantly lower than that in the exacerbation stage; Clinical mechanism verification: Based on the pathological mechanism of kidney disease, false correlation factors were eliminated, including non-causal associations between seasonal changes and renal function fluctuations, while factors with clear pathological pathways were retained, including renal tubular damage caused by nephrotoxic drugs; Classify exacerbation factors: Based on risk intensity and intervention feasibility, exacerbation factors are divided into: High-risk controllable factors: risk intensity > 70% and can be eliminated through intervention, including the use of nephrotoxic drugs, excessive salt intake, and unauthorized discontinuation of medication. Moderate-risk controllable factors: risk intensity 30%-70%, requiring long-term intervention and improvement, including poor blood pressure control, lack of sleep, and too little exercise; Low-risk uncontrollable factors: risk intensity <30%, intervention is more difficult, including increasing age and underlying kidney disease type.

[0036] S35: Generate factor analysis report, including: A list of improvement factors, including factor name, contribution, mechanism of action, and intervention threshold; A list of exacerbation factors, including factor name, risk intensity, associated indicators, and control objectives; Factor association map showing the positive correlation between improvement factors and renal function indicators, and the negative correlation between deterioration factors and renal function indicators.

[0037] The specific process of step S4 is as follows: S41: Constructing a knowledge graph in the field of kidney disease: Construct a knowledge graph containing core entities and associated entities. The core entities include kidney function indicators, intervention measures, and complications, and the associated entities include lifestyle behaviors and environmental factors. Define entity relationships, including the causal relationship of high-salt diet → increased blood pressure → kidney function damage, and the therapeutic relationship of ACEI drugs → improved eGFR. Knowledge embedding: Use the TransE algorithm to convert the entities and relationships in the knowledge graph into low-dimensional vectors to generate a knowledge graph embedding vector. Among them, a higher embedding weight is given to kidney disease-specific relationships to ensure that the core medical logic is retained first in the vector space. Kidney disease-specific relationships include increased urine protein quantification → glomerular damage.

[0038] S42: The input layer of the neural network prediction model receives the preprocessed multi-dimensional health data and loads the knowledge graph embedding vector as auxiliary input. The neural network prediction model structure is shown in Figure 2 shown.

[0039] S43: Feature extraction: Convolutional neural networks are used to extract local features of the data, including the characteristics of changes in blood creatinine over three consecutive weeks. A multi-headed self-attention mechanism is introduced to dynamically assign weights to input features according to their clinical importance. A basic weight of 0.3-0.4 is assigned to core indicators including eGFR and blood creatinine, a weight of 0.1-0.2 is assigned to related indicators including eGFR and blood creatinine, and a weight of <0.1 is assigned to secondary features including sleep duration. An M×M attention weight matrix is ​​used to strengthen the association between key features, such as the coordinated change characteristics of blood creatinine and urine protein, and finally a feature vector is obtained.

[0040] S44: Integrate the feature vector output by the feature extraction layer with the knowledge graph embedding vector through a gated fusion mechanism: Dynamically adjust the fusion ratio of the two types of features through the gating unit, including increasing the weight of the "drug-kidney function" relationship feature in the knowledge graph when the "drug intervention" feature is clear in the input data; After fusion, a "data-knowledge" joint feature vector is generated, while the core information of the original data features is retained through residual connections.

[0041] S45: Using an improved bidirectional long short-term memory network, see Figure 3 As shown, x1 and x 2 is the input sequence. Two hidden layers, a feature concatenation layer, and a fully connected layer are set to capture the long-term temporal dependencies of the joint feature vector, including the impact of feature changes over the past three months on the next month. By introducing a temporal attention mechanism, a higher temporal weight is assigned to recent data, improving short-term prediction accuracy.

[0042] Design a dual-output channel structure to output the renal function status prediction results respectively, and output the continuous value prediction results and classification prediction results through the dual output channels: Continuous value prediction: The fully connected layer outputs the predicted values ​​of renal function indicators for a specified period, including eGFR and serum creatinine, using a linear activation function. The specified period can be set to 1 month. Classification prediction: The softmax activation function is used to output the short-term health risk level (low / medium / high) and the probability of complication warning, including the risk of hyperkalemia.

[0043] Example 4 Based on Example 1, Example 2, or Example 3, the health management rule base includes a basic rule layer, a prediction adaptation layer, and a preference adaptation layer: Basic rule layer: Develop universal rules based on clinical guidelines for kidney disease, covering: Intervention rules corresponding to renal function indicators include protein intake ≤0.6 g / kg / day when eGFR <30 ml / min; The risk level response rules include weekly review for high risk and every three months for low risk; Complication prevention rules include limiting potassium intake to <2 g / day in the event of hyperkalemia. Prediction adaptation layer: Mapping rules linking predicted renal function status results with intervention intensity. For example, when the predicted eGFR decreases by >5 ml / min in the next month, an enhanced intervention rule is triggered, increasing the frequency of medication adjustments and shortening the review interval. If the predicted risk level is medium and shows a downward trend, the progressive intervention rule will be triggered to gradually reduce the intensity of intervention. Preference adaptation layer: stores the adaptation rules between patient preferences and intervention plans, including mapping vegetarian preferences to a diet based on plant protein, and mapping sedentary occupations to fragmented exercise recommendations.

[0044] Set up a dynamic update mechanism for the health management rule base: The basic rules are revised every quarter in combination with the latest clinical guidelines (KDIGO guideline updates); after accumulating the execution data of every 100 patients, the prediction adaptation layer rules are optimized through association analysis, including adjusting the matching threshold of "risk level-intervention intensity"; patient preference feedback is received in real time, and the mapping relationship of the preference adaptation layer is updated, including the addition of a recipe library corresponding to low-sugar diet preferences.

[0045] S51: Rule base matching and candidate solution generation: Matching the basic intervention framework: Based on the predicted renal function status, the corresponding rules are called to generate the basic intervention framework. If the predicted result is "eGFR 28ml / min + hyperkalemia risk 62% + high risk", the basic framework includes: protein intake 0.6g / kg / day, potassium intake <2g / day, renal function review once a week, and daily blood pressure monitoring; Expand the set of candidate solutions: For each intervention item in the basic framework, 3-5 alternative solutions are retrieved from the rule library. Taking the diet plan as an example, based on the basic rules of low protein and low potassium, candidate solutions such as Chinese vegetarian plan, Mediterranean diet modification plan, and high protein and low phosphorus alternative plan are generated; S52: Patient preference adaptation and program screening: Preference Matching Calculation: For each candidate plan, calculate its match score with the patient's preferences, with a maximum score of 100. For example, if a patient's preference for vegetarianism is 5 points, then the Chinese vegetarian plan's match score will be increased by 30 points; if a patient dislikes swimming (with a preference score of 1 point), then the match score for an exercise plan that includes swimming will be deducted by 20 points. Screening the best candidate plan: Select the candidate plan with a matching degree ≥80 points. If there are multiple plans that meet the standards, make a comprehensive selection based on the expected intervention effect.

[0046] S53: Integrate the selected candidate solutions into an initial health management plan, including: Diet module: Specific recipes planned by week, ingredient replacement lists considering preferences and taboos, and cooking suggestions that suit your preferred cooking methods; Exercise module: Matches the preferred exercise type, duration based on tolerance, and intensity of low-intensity exercise for high-risk patients in combination with predicted risk level; Medical management module: medication reminders with patient-preferred medication times, review plans, and indicator monitoring frequencies; Traditional Chinese Medicine conditioning module: If the patient has a preference for traditional Chinese medicine, a conditioning plan corresponding to the patient's constitution will be added, including the recommendation of soaking Astragalus in water for patients with Qi deficiency.

[0047] S54: Check the rationality of the initial health management plan: Clinical rationality verification: Calling the nephrology expert system to verify whether the plan meets the intervention requirements corresponding to the predicted results, such as whether the high-risk plan includes intervention measures of sufficient intensity; Perform feasibility check: Evaluate the adaptability of the plan to the patient's living conditions, including whether the recommended exercise requires a special venue and whether the recipe ingredients are easily accessible. Adjust items with feasibility less than 60%, such as changing gym exercise to home yoga.

[0048] The specific process of step S6 is as follows: S61: Definition of core elements of reinforcement learning: Agent: The health management plan optimization module is responsible for adjusting plan parameters based on input data; Environment: It is composed of the health status, executive ability and external conditions of the target kidney patients, with the degree of improvement factor achievement, degree of deterioration factor control and program implementation feedback as environmental state variables; Actions: 12 basic optimization actions are defined, including replacement of key ingredients in the diet plan, adjustment of exercise duration by ±10 minutes, increase or decrease in the frequency of medication reminders, adjustment of review intervals by ±3 days, etc. Among them, replacement of key ingredients includes replacement of low-potassium ingredients.

[0049] Reward: Design a multi-level reward function. Positive rewards include: Core rewards: Improve factor activation, including a +50 reward when the low salt intake factor reaches the target after implementing the plan; Auxiliary rewards: Deterioration factor suppression, including a reward of +30 when the exposure to high salt intake factor decreases by 30%; Additional reward: Patient compliance score, based on historical execution data prediction, reward +20 when the score is ≥80 points.

[0050] Mapping between state space and action space: State space quantification: The environmental state is converted into an 18-dimensional state vector, which includes six improvement factor compliance states (0-1 variables, 1 means compliance), six deterioration factor control states (0-100 points, 100 points means complete control), three execution capacity indicators (time, energy, and economic cost), and three recent health indicator change values; Action space constraints: Set clinical safety boundaries for each action, including protein intake adjustment actions must meet the clinical range of 0.6-1.2g / kg / day, and exercise intensity adjustment actions must not exceed the patient's physical tolerance threshold.

[0051] S62: Make initial adjustments guided by improvement factors: establish a correspondence table between improvement factors and program intervention items, including the execution frequency parameters of the exercise module corresponding to the regular exercise factor. If the contribution of this factor reaches 40% in historical data, the execution frequency of the exercise program will be increased from 3 times a week to 4 times a week; the intestinal flora balance factor corresponds to the proportion of probiotic ingredients in the diet module. If this factor is the core improvement factor, the frequency of probiotic ingredients in the recipe will be increased by 20%.

[0052] Improvement potential assessment: For each improvement factor, the gap between the current state and the target state is calculated, including the current achievement rate of high-quality protein intake of 60% and the target of 80%. Factors with larger gaps have higher optimization priorities for the corresponding intervention items, and the highest priority is given to factors with a gap of >40%.

[0053] S63: Make risk aversion adjustments guided by deterioration factors: Risk level ranking: Sort by the risk intensity of the exacerbating factors, high-risk controllable factors > medium-risk controllable factors > low-risk uncontrollable factors, and give priority to addressing the solution loopholes corresponding to high-risk controllable factors. If the use of nephrotoxic drugs is a high-risk factor, immediately add nephrotoxic drug substitution recommendations to the medication module, including replacing non-steroidal anti-inflammatory drugs that may damage the kidneys with ARBs; If excessive salt intake is a high-risk factor, a new requirement for visual recording of daily salt intake will be added to the diet module, and a recommended list of low-potassium and low-sodium seasonings will be provided.

[0054] Factor control target embedding: The control threshold of the exacerbation factor is converted into program parameters, including the control target of the blood pressure fluctuation factor is systolic blood pressure <130mmHg. In the exercise module, high-intensity exercise is limited to avoid sudden increases in blood pressure, and the quantitative standard of daily sodium intake <2g is clearly defined in the diet module.

[0055] S64: Iteratively optimize the initial solution based on reinforcement learning: Conduct status assessment of the initial health management plan: Calculate the baseline score of the factor matching of the initial health management plan, with a maximum score of 100. The assessment dimensions include: Improvement factor coverage, such as whether the intervention items corresponding to all core improvement factors are included; The degree of avoidance of exacerbating factors, such as whether targeted interventions are set for high-risk factors; Implementation feasibility, which quantifies the probability of implementation of outcome predictions based on patient preferences. Plans with a baseline score of <60 points will directly enter the reinforcement learning optimization process, plans with a score of 60-80 will focus on optimizing the intervention items corresponding to the weak factors, and plans with a score of ≥80 will only undergo fine-tuning.

[0056] Perform multiple rounds of iterative optimization: Round 1-2: Factor activation and risk control.

[0057] The intelligent agent prioritizes actions related to core improvement factors and high-risk deterioration factors, including adding daily protein source labeling actions for high-quality protein intake factors; The reward value is calculated after each round of iteration. If the positive reward is greater than 80 points for two consecutive rounds, the current optimization action is retained. If a negative penalty occurs, including a decrease in the control degree of the deterioration factor after the solution is adjusted, the optimization action is retraced to the previous state and the optimization direction is changed. Round 3-4: Perform adaptability optimization.

[0058] The action selection is adjusted based on the patient's preference feature vector. For patients who are sensitive to cooking time, complex recipes in the diet plan are replaced with 30-minute quick dishes, and the corresponding ingredient pre-processing actions are simplified. Through the exploration-utilization strategy, we balance the optimization effect and patient acceptance: 70% probability of selecting actions with high historical rewards, including adjusting the walking duration if it is known that the patient likes walking, and 30% probability of trying new actions, including recommending mild Tai Chi as an alternative exercise.

[0059] Round 5: The overall effect converges.

[0060] Calculate the marginal benefit of each intervention item, the reward increment brought by each parameter adjustment, and stop adjusting actions with a marginal benefit of less than 5 points; Output the optimized program parameter set, including the core parameters of the diet module including the upper limit of daily potassium intake, the execution parameters of the exercise module including the exercise intensity and heart rate threshold, and the monitoring parameters of the medical module.

[0061] S65: Check the optimization action constraints. Double check is performed after each round of optimization: Clinical compliance verification: Calling the health management rule library to verify whether the optimization actions comply with clinical guidelines, including whether the protein intake adjustment is within the safe range; Factor synergy verification: Ensure that the optimized movement does not produce factor conflicts, including checking whether increasing the duration of exercise will increase the risk of potential exacerbating factors related to joint injuries. If the verification fails, the action correction mechanism is automatically triggered, including reducing the exercise duration adjustment from +20 minutes to +10 minutes; Inject the optimization parameters output by reinforcement learning into the initial health management plan to form an optimized health management plan that includes core optimization items and auxiliary optimization items: Core Optimization: Adjustments to actions that contribute ≥30 points to a reward, including adding probiotic ingredients to increase the activation of intestinal flora improvement factors by 40%; Auxiliary optimization items: Fine-tuning actions with a reward value of 10-30 points, including changing the medication reminder time from 8:00 a.m. to 7:00 a.m., the patient's preferred time.

[0062] Conduct program effect prediction and verification, and simulate and predict the execution effect of the optimization program through historical data: Predicting the activation rate of improvement factors based on data from similar patients, including an expected increase in the achievement rate of regular exercise factors from 60% to 85%; Predicting the control rate of exacerbating factors, including a 50% reduction in exposure to factors that exceed the salt intake standard; If the prediction effect does not meet expectations, including the activation rate of the core improvement factor increasing by less than 15%, return to the reinforcement learning module and add 2 rounds of iterations.

[0063] Generate an optimized health management plan report with additional optimization trajectory description: List the key actions and reward changes for each round of optimization, including the increase of the preference matching reward of +25 in round 3 due to the addition of vegetarian recipes; The logic of associating the labeling plan with improvement factors and aggravation factors, including reducing red meat intake while responding to the need to control low protein improvement factors and high uric acid aggravation factors; Provides optimized flexible range, including exercise duration can be adjusted independently within ±5 minutes of the recommended value, leaving room for patients to make their own choices.

[0064] The health management system for kidney patients based on the neural network prediction model is used to implement the health management method for kidney patients based on the neural network prediction model, including a data acquisition module, a preprocessing module, a disease course division module, a disease factor analysis module, a neural network prediction model, a health management plan generation module, and a plan optimization module.

[0065] The data acquisition module is used to obtain multidimensional health data from target kidney disease patients at various stages. The preprocessing module is used to preprocess the multidimensional health data. The disease course classification module is used to obtain the target kidney disease patient's disease progression trends throughout the disease course based on the preprocessed multidimensional health data. Based on these progression trends, the disease course is divided into two categories: remission and exacerbation. The disease course classification module is used to obtain corresponding multidimensional health data for the remission and exacerbation stages, respectively. Based on the analysis of the multidimensional health data for the remission stage, improvement factors are obtained, and based on the analysis of the multidimensional health data for the exacerbation stage, exacerbation factors are obtained. The neural network prediction model is used to output a prediction of the target kidney disease patient's renal function status during a specified period based on the preprocessed multidimensional health data. Based on the renal function status prediction results, the neural network prediction model calls a preset health management rule library and combines them with the target patient's preferences to generate an initial health management plan. The plan optimization module is used to dynamically optimize the initial health management plan based on the improvement factors and exacerbation factors using a reinforcement learning algorithm to generate an optimized health management plan.

Claims

1. A health management method for kidney disease patients based on a neural network prediction model, characterized in that: The following steps are involved: S1: Obtain multidimensional health data of target kidney disease patients at various stages, preprocess the data, and obtain the disease condition change trend of the target kidney disease patients throughout the entire disease course based on the preprocessed multidimensional health data; S2: The entire course of the disease is divided into different stages based on the trend of disease changes. Different stages are divided into two categories, including remission stage and exacerbation stage; S3: Obtain corresponding multidimensional health data based on the remission stage and the exacerbation stage, obtain an improvement factor based on the analysis of the multidimensional health data in the remission stage, and obtain a deterioration factor based on the analysis of the multidimensional health data in the exacerbation stage; S4: Build a neural network prediction model that integrates the attention mechanism and knowledge graph, input preprocessed multi-dimensional health data after training, and output the predicted renal function status of the target kidney disease patient during the specified period; S5: Based on the predicted results of renal function status, the preset health management rule library is called and combined with the preferences of the target patient to generate an initial health management plan; S6: Based on the improvement factors and deterioration factors, the initial health management plan is dynamically optimized through the reinforcement learning algorithm to generate an optimized health management plan.

2. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: It also includes a dynamic feedback process, as follows: S7: Execute the optimized health management plan for the target kidney disease patient and collect multi-dimensional health data of the target kidney disease patient within a specified time period starting from the execution time node; S8: obtaining the condition change trend of the target kidney disease patient within the specified time period based on the multi-dimensional health data within the specified time period; S9: Based on the trend of disease changes in the specified time period, determine whether it belongs to the stage of disease remission or the stage of disease worsening. If it is the stage of disease remission, continue to execute the optimized health management plan. If it is the stage of disease worsening, re-analyze the corresponding multi-dimensional health data based on the trend of disease changes in the specified stage to obtain the feedback deterioration factor. Based on the feedback deterioration factor, further optimize the optimized health management plan and execute it on the target kidney disease patient. After execution, iteratively execute steps S7-S9 to achieve iterative optimization of the optimized health management plan.

3. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: The multidimensional health data includes designated clinical indicator data, medical intervention record data, diet data, associated impact data, and exercise data. The specific process of preprocessing the multidimensional health data in step S1 is as follows: S11: dividing the multidimensional health data into continuous indicator data, classification data, and time series data according to the first division dimension, and dividing the multidimensional health data into initial diagnosis data, treatment period data, and follow-up period data according to the second division dimension; S12: Z-score normalization is used for continuous indicator data, one-hot encoding is performed on categorical data, and time series data is uniformly converted into time series with weekly units; S13: A stage-by-stage adaptive preprocessing strategy was used to process missing values ​​for data at different disease stages. S14: Through the 3σ principle combined with clinical rationality verification, true abnormal values ​​are distinguished from false abnormal values. For true abnormal values, the original value is retained and the cause of the abnormality is marked. For false abnormal values, the moving average method is used to correct them.

4. The health management method for kidney disease patients based on a neural network prediction model according to claim 2, characterized in that: The specific process of obtaining the target kidney disease patient's condition change trend throughout the course of the disease based on the pre-processed multi-dimensional health data in step S1 is as follows: S15: Constructing a time series feature matrix: Arrange the preprocessed data along the time axis to generate a feature matrix with time-index as the dimension; S16: Extract trend features: Calculate the slope of change of each specified indicator within a continuous time window, and generate a smooth trend line by combining spline curve fitting; determine the indicator's rising and falling trend through the first-order derivative of the trend line, and determine the trend change rate through the second-order derivative; S17: Perform multi-indicator fusion: Perform dimensionality reduction processing on the multi-indicator trends, extract the top three principal components with the highest contribution rate, and construct a comprehensive disease trend index.

5. The health management method for kidney disease patients based on a neural network prediction model according to claim 4, characterized in that: The specific process of step S2 is as follows: S21: Determine the stage division indicators: use the comprehensive disease trend index as the basic indicator, combine the changing trends of designated key clinical indicators as auxiliary verification indicators, and set the weights of each indicator; S22: Setting stage division thresholds: including thresholds for determining the stage of disease remission, thresholds for determining the stage of disease exacerbation, and transition stage processing; S23: Perform dynamic window division to obtain the remission stage and the exacerbation stage, and classify the short-term indicator fluctuations that do not meet the duration requirements into adjacent major stages; S24: Verify and revise the stage boundaries: Verify the preliminary stage boundaries based on clinical event records and introduce stage consistency verification rules for revision; S25: Generate a disease stage division report, including the time interval of each stage, the core indicator change curve, the basis for stage division and the annotation of key clinical events.

6. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: The specific process of step S3 is as follows: S31: extracting time interval data corresponding to the remission stage and the exacerbation stage from the multidimensional health data of the entire disease course, forming two independent data sets including a remission stage data set and a exacerbation stage data set; S32: For the dataset of the remission stage, the temporal relationship between intervention measures and indicator changes is retained, and a time alignment algorithm is used to match the execution time of intervention measures with the indicator change time; S33: Perform improvement factor analysis to screen potential related factors from the data of the disease remission stage and form a candidate improvement factor pool; Bivariate correlation analysis combined with multivariate regression model was used to conduct correlation analysis of improvement factors and classification of improvement factors: S34: Perform exacerbation factor analysis and extract candidate exacerbation factors to screen potential risk factors from the disease exacerbation stage data to form a candidate exacerbation factor pool; Causal inference combined with risk accumulation model was used to conduct correlation analysis and classification of exacerbation factors; S35: Generate a factor analysis report, including a list of improvement factors, a list of deterioration factors, and a factor correlation graph.

7. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: The specific process of step S4 is as follows: S41: Constructing a knowledge graph for kidney disease: Construct a knowledge graph containing core entities and associated entities, define entity relationships, convert entities and relationships in the knowledge graph into low-dimensional vectors, and generate a knowledge graph embedding vector; S42: The input layer of the neural network prediction model receives the preprocessed multi-dimensional health data and loads the knowledge graph embedding vector as auxiliary input; S43: performing feature extraction to obtain a feature vector; S44: The feature vector output by the feature extraction layer is integrated with the knowledge graph embedding vector through a gated fusion mechanism. The fusion ratio of the two types of features is dynamically adjusted through the gated unit. After fusion, a data-knowledge joint feature vector is generated. S45: An improved bidirectional long short-term memory network is used to capture the long-term temporal dependencies of the joint feature vector. By introducing a temporal attention mechanism, a higher temporal weight is assigned to recent data, and the prediction results of renal function status are output separately: including continuous value prediction and classification prediction.

8. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: The specific process of step S5 is as follows: S51: Rule base matching and candidate solution generation: Matching basic intervention framework: Call corresponding rules based on the predicted results of renal function status to generate a basic intervention framework; Expand the candidate solution set: For each intervention item in the basic framework, retrieve 3-5 alternative solutions from the rule base to obtain candidate solutions; S52: Patient preference adaptation and program screening: Preference matching calculation: For each candidate solution, the matching score with the patient’s preference is calculated; Screening the best candidate plan: select the candidate plan with a matching score of ≥80 points. If there are multiple plans that meet the standard, select one comprehensively based on the expected intervention effect; S53: Integrate the screened candidate solutions into the initial health management plan; S54: Check the rationality of the initial health management plan.

9. The health management method for kidney disease patients based on a neural network prediction model according to claim 1, characterized in that: The specific process of step S6 is as follows: S61: Define the core elements of reinforcement learning and map the state space to the action space: S62: Perform initial adjustments directed towards improvement factors; S63: Make risk aversion adjustments guided by deterioration factors; S64: Iteratively optimize the initial solution based on reinforcement learning; S65: Check the optimization action constraints: After each round of optimization, double verification is performed: clinical compliance verification and factor synergy verification; The optimization parameters output by reinforcement learning are injected into the initial health management plan to form an optimized health management plan that includes core optimization items and auxiliary optimization items.

10. A health management system for kidney disease patients based on a neural network prediction model, used to implement the health management method for kidney disease patients based on a neural network prediction model according to any one of claims 1 to 9, characterized in that: It includes data acquisition module, preprocessing module, disease course classification module, disease factor analysis module, neural network prediction model, health management plan generation module, and plan optimization module; A data acquisition module is used to obtain multi-dimensional health data of target kidney disease patients at various stages; Preprocessing module, used to preprocess multidimensional health data; The disease course classification module is used to obtain the disease course change trend of the target kidney disease patient throughout the disease course based on the pre-processed multi-dimensional health data. The disease course is divided into different stages based on the disease course change trend. The different stages are divided into two categories: remission stage and exacerbation stage. A disease course classification module is used to obtain corresponding multidimensional health data based on the remission stage and the exacerbation stage, obtain improvement factors based on the analysis of the multidimensional health data in the remission stage, and obtain exacerbation factors based on the analysis of the multidimensional health data in the exacerbation stage; A neural network prediction model is used to output a prediction result of the renal function status of a target kidney disease patient during a specified period based on the input of pre-processed multi-dimensional health data; A neural network prediction model is used to call a preset health management rule library based on the predicted results of renal function status and combine it with the preferences of the target patient to generate an initial health management plan; The solution optimization module is used to dynamically optimize the initial health management plan based on improvement factors and deterioration factors through reinforcement learning algorithms to generate an optimized health management plan.

Images