Diabetes management system

Abstract

The invention relates to a diabetes management system, equipment and a medium. The method comprises the steps that a data processing module collects multi-modal medical data, dynamic time sequence features are extracted through a graph neural network to label semantic tags, and a structured patient information data set is formed; in the database construction module, a data fusion unit performs cross-modal feature alignment, redundancy elimination and feature enhancement on the structured data set, and establishes a dynamic and unified health record database; and the risk assessment unit extracts individual features by adopting an improved random forest algorithm based on the database, and generates a personalized health risk assessment report. And the health management module customizes a health management scheme through a generative adversarial network according to the evaluation report, and dynamically updates intervention measures according to real-time monitoring data. By adopting the system, the accuracy and individuation degree of risk prediction can be improved, accurate and dynamic whole-cycle health management service is provided for diabetics, and the management efficiency and the medical quality are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical health information technology, and in particular relates to a diabetes management system. Background Art

[0002] As a chronic disease, diabetes management requires the integration of multi-source medical data for precise intervention. Existing systems struggle with low data standardization and insufficient feature-association analysis when processing heterogeneous data such as electronic medical records, imaging reports, and real-time vital sign monitoring. This results in a lack of dynamism and semantic depth in the construction of health records. Traditional risk assessment methods are often based on single models, making it difficult to capture the temporal changes in physiological indicators and the synergistic effects between features. Furthermore, the generation of health management plans relies on fixed rules and cannot be adjusted in real time based on individual responses. Summary of the Invention

[0003] Based on this, it is necessary to address the above technical issues and provide a diabetes management system that can provide accurate and dynamic full-cycle health management services for diabetic patients and effectively improve management efficiency and medical quality.

[0004] In a first aspect, the present application provides a diabetes management system, comprising:

[0005] The data processing module is used to obtain multimodal medical data sources, perform dynamic semantic annotation using graph neural networks, and generate structured patient information datasets.

[0006] Database building blocks, including:

[0007] The data fusion unit is used to eliminate redundancy and enhance features of patient information data sets using an adaptive fusion model to build a dynamic and unified patient health record database.

[0008] The risk assessment unit is used to extract individual features from the health record database using an improved random forest algorithm to generate a personalized health risk assessment report.

[0009] The health management module is used to customize health management plans based on risk assessment reports using generative adversarial networks. If the real-time monitoring data after implementation exceeds the preset threshold, intervention measures are generated to update the health management plan.

[0010] In one embodiment, a multimodal medical data source is obtained, and dynamic semantic annotation is performed using a graph neural network to generate a structured patient information dataset, including:

[0011] Obtain examination results and monitoring data from multimodal medical data sources; multimodal medical data sources include imaging reports and real-time vital sign streams.

[0012] Extract dynamic time series features from the inspection results; dynamic time series features include fluctuations in inspection indicators and changes in medication records.

[0013] Construct a heterogeneous subgraph of dynamic time series features; the heterogeneous subgraph contains patient entity nodes and cross-source data relationship edges.

[0014] Node embedding vectors of heterogeneous subgraphs are generated through graph neural networks and annotated with dynamic semantic labels; dynamic semantic labels include disease stage and risk level.

[0015] Dynamic semantic labels and original data are mapped into structured fields, and the structured fields are fused to obtain the patient information dataset.

[0016] In one embodiment, the data fusion unit includes:

[0017] Acquire multi-source heterogeneous data from patient information datasets; multi-source heterogeneous data includes electronic medical records and vital sign monitoring records.

[0018] Based on multi-source heterogeneous data, a dynamic weight allocation algorithm is used to fuse the feature distributions of different data sources to generate cross-modal feature alignment results.

[0019] The feature similarity matrix is ​​calculated based on the cross-modal feature alignment results.

[0020] Redundant feature fields are eliminated according to the feature similarity matrix to obtain filtered non-redundant feature fields; the screening threshold of redundant feature fields is dynamically adjusted by the adaptive fusion model.

[0021] Non-redundant feature fields are used to enhance key health indicators using the attention mechanism.

[0022] Generate a dynamic unified patient health record index based on enhanced key health indicators.

[0023] Update the patient health record database based on the dynamic unified patient health record index.

[0024] In one embodiment, the feature similarity matrix is ​​calculated using the following formula:

[0025]

[0026] Among them, S i,j represents the feature similarity between samples i and j, GCK() represents the graph convolution kernel function, x i and x j represents the feature vector of the i-th and j-th samples, L represents the normalized Laplace matrix, K represents the number of graph convolution layers, α k represents the attention weight of the kth layer, W krepresents the feature transformation matrix, σ(·) represents the normalization function, ReLU(·) linear rectification activation function, b k Represents the bias vector of the kth layer.

[0027] In one embodiment, an improved random forest algorithm is used to extract individual features from a health record database to generate a personalized health risk assessment report, including:

[0028] An individual health data set is obtained from a health record database; the individual health data set includes multidimensional physiological indicators and medical history records.

[0029] The feature weight matrix is ​​calculated using an improved random forest algorithm based on the individual health data set.

[0030] Dynamic feature screening rules are established based on the feature weight matrix to extract the key feature subset of the target individual.

[0031] An individual health profile is generated based on a subset of key features and input into a risk prediction model to output a multidimensional risk probability vector.

[0032] The multi-dimensional risk probability vector and the matching assessment dimensions are integrated to generate structured assessment data; the structured assessment data is converted into text paragraphs through a natural language processing engine.

[0033] The changing trends of temporal features in text paragraphs and individual health portraits are integrated to generate the final health risk assessment report.

[0034] In one embodiment, the feature weight matrix is ​​constructed using the following formula:

[0035]

[0036] Among them, W represents the feature weight matrix, Wi ,j represents the relative importance of feature i to feature j, T represents the total number of trees in the random forest, ΔImpurity t,i represents the reduction in impurity of the i-th tree on feature i, C t represents the feature co-occurrence matrix of the t-th tree, τ i , τ j represents the timestamps of feature i and feature j, γ represents the time decay coefficient, K represents the domain knowledge matrix, and the strength of the prior relationship between features is defined by medical experts.

[0037] In one embodiment, the health management module includes:

[0038] Health program building blocks for:

[0039] Obtain health risk indicators in the risk assessment report; health risk indicators include abnormal values ​​of physiological parameters and predicted probability of disease.

[0040] The health risk indicators are input into the generator model of the generative adversarial network, and an initial health plan including exercise intensity and nutritional ratio is output.

[0041] The evaluation results of the initial health plan by the discriminator model of the generative adversarial network are calculated based on the difference between the user's historical health data and the execution effect of the plan.

[0042] The weight parameters of the generator model are adjusted according to the evaluation results to generate a health management plan that includes dynamic intervention cycles and monitoring thresholds.

[0043] In one embodiment, the health management module further includes:

[0044] Health program optimization unit for:

[0045] The health management plan is matched and verified with the real-time physiological monitoring data to obtain a matching result; the matching verification is completed by calculating the deviation of physiological indicators during the execution of the plan.

[0046] If the deviation of the physiological indicators in the matching results exceeds a preset threshold, an abnormal indicator identification code set is generated.

[0047] The abnormal indicator identification code set is input into the intervention measure matching model, and the corresponding intervention parameter combination is output; the intervention measure matching model is constructed based on historical intervention effect data.

[0048] The executable intervention instruction sequence in the health management plan is updated according to the combination of intervention parameters to obtain an optimized health management plan.

[0049] In a second aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above system when executing the computer program.

[0050] In a third aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the above system when executed by a processor.

[0051] In the above-mentioned diabetes management system, computer equipment, and storage medium, the data processing module collects multimodal medical data, uses graph neural networks to extract dynamic temporal features and annotate semantic tags to form a structured patient information dataset. In the database construction module, the data fusion unit performs cross-modal feature alignment, redundancy elimination, and feature enhancement on the structured dataset to establish a dynamic and unified health record database; based on this database, the risk assessment unit uses an improved random forest algorithm to extract individual features and generate personalized health risk assessment reports. Based on the assessment report, the health management module customizes health management plans through generative adversarial networks and dynamically updates intervention measures based on real-time monitoring data. The use of this system can achieve efficient integration and standardized processing of multi-source heterogeneous medical data; improve the accuracy and personalization of risk prediction; and enhance the targetedness and timeliness of intervention measures. It provides accurate and dynamic full-cycle health management services for diabetic patients, effectively improving management efficiency and medical quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 A structural block diagram of a diabetes management system provided by an embodiment of the present invention;

[0054] Figure 2 A flowchart for obtaining a multimodal medical data source, using a graph neural network for dynamic semantic annotation, and generating a structured patient information dataset, provided by an embodiment of the present invention;

[0055] Figure 3 The present invention provides a flowchart for extracting individual features from a health record database using an improved random forest algorithm to generate a personalized health risk assessment report. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0057] In one embodiment, Figure 1 As shown, the present application provides a diabetes management system, which may include:

[0058] The data processing module 101 is used to obtain multimodal medical data sources, perform dynamic semantic annotation using graph neural networks, and generate structured patient information datasets.

[0059] Specifically, this module is responsible for collecting multimodal medical data such as imaging reports, real-time vital sign streams, and electronic medical records. Imaging reports contain image data, while real-time vital sign streams are continuous numerical sequences. The module uses graph neural network technology to deeply analyze the data. First, dynamic time series features such as fluctuations in test indicators and changes in medication records are extracted. Then, with the patient as the core, data from different sources are constructed into a heterogeneous subgraph containing entity nodes and relationship edges. Node embedding vectors are generated through graph neural networks, and dynamic semantic labels such as disease stage and risk level are annotated. Finally, the original data and semantic label mapping are integrated to form a structured patient information dataset.

[0060] The database construction module 102 includes:

[0061] The data fusion unit 1021 is used to eliminate redundancy and enhance features of the patient information data set using an adaptive fusion model to build a dynamic unified patient health record database.

[0062] Specifically, the data fusion unit receives a structured patient information dataset generated by the data processing module, which contains multi-source heterogeneous data such as electronic medical records and vital sign monitoring records. The unit uses an adaptive fusion model to analyze and fuse the feature distributions of different data sources through a dynamic weight allocation algorithm to achieve cross-modal feature alignment. Based on the alignment results, the feature similarity matrix is ​​calculated, and redundant feature fields are eliminated according to the matrix. The screening threshold is dynamically adjusted by the adaptive fusion model according to the characteristics of the data. Then, the attention mechanism is used to process the filtered non-redundant feature fields to enhance key health indicators, such as blood sugar, insulin dosage and other indicators closely related to diabetes. Finally, a dynamic unified patient health record index is generated based on the enhanced key health indicators, and the patient health record database is updated based on the index to ensure that the database can dynamically and accurately reflect the patient's health information.

[0063] The risk assessment unit 1022 is used to extract individual features from the health record database using an improved random forest algorithm to generate a personalized health risk assessment report.

[0064] The risk assessment unit obtains a set of individual health data from the established health record database, which covers a rich set of information such as multidimensional physiological indicators and medical history records. The unit uses an improved random forest algorithm to calculate the feature weight matrix through a specific formula, comprehensively considering multiple factors such as the reduction in impurity of the features in the decision tree, the frequency of feature co-occurrence, time factors, and domain knowledge, to establish dynamic feature screening rules, and extract a subset of key features that are highly relevant to the target individual's diabetes risk assessment from massive data. Based on the key feature subset, an individual health profile is generated and input into the risk prediction model. The model outputs a multidimensional risk probability vector, covering multiple dimensions such as the risk of diabetes complications and the risk of disease worsening. Finally, the multidimensional risk probability vector is fused with the corresponding assessment dimensions to generate structured assessment data, which is converted into text paragraphs through a natural language processing engine. At the same time, the changing trends of the temporal features in the individual health profile are integrated to generate a complete and personalized health risk assessment report.

[0065] The health management module 103 is used to customize a health management plan based on the risk assessment report using a generative adversarial network. If the real-time monitoring data after implementation exceeds a preset threshold, intervention measures are generated to update the health management plan.

[0066] Specifically, the health management module operates based on the risk assessment report. First, the health plan construction unit extracts health risk indicators from the report, including key information such as abnormal physiological parameter values ​​and disease prediction probabilities. These indicators are then fed into the generator model of a generative adversarial network (GAN) to generate an initial health plan, including exercise intensity and nutritional balance. Based on the discrepancy between the user's historical health data and the plan's execution results, the discriminator model of the GAN evaluates the initial plan. Based on the evaluation results, the weight parameters of the generator model are adjusted to further optimize the plan, generating a complete health management plan with dynamic intervention cycles and monitoring thresholds. During plan implementation, the health plan optimization unit continuously verifies the health management plan against real-time physiological monitoring data, determining its effectiveness by calculating the deviation of physiological indicators during plan execution. If the deviation exceeds a preset threshold, a set of abnormal indicator identification codes is generated and fed into an intervention measure matching model built based on historical intervention effect data. The resulting code outputs a corresponding combination of intervention parameters, which is then used to update the executable intervention instruction sequence in the health management plan, achieving dynamic optimization of the health management plan and ensuring effective health management services for patients.

[0067] In the above-mentioned diabetes management system, the data processing module collects multimodal medical data, uses graph neural networks to extract dynamic temporal features and annotate semantic tags to form a structured patient information dataset. In the database construction module, the data fusion unit performs cross-modal feature alignment, redundancy elimination, and feature enhancement on the structured dataset to establish a dynamic and unified health record database; based on this database, the risk assessment unit uses an improved random forest algorithm to extract individual features and generate personalized health risk assessment reports. Based on the assessment report, the health management module customizes health management plans through generative adversarial networks and dynamically updates intervention measures based on real-time monitoring data. The use of this system can achieve efficient integration and standardized processing of multi-source heterogeneous medical data; improve the accuracy and personalization of risk prediction; and enhance the pertinence and timeliness of intervention measures. It provides accurate and dynamic full-cycle health management services for diabetic patients, effectively improving management efficiency and medical quality.

[0068] In one embodiment, Figure 2 As shown in the figure, obtaining a multimodal medical data source, using a graph neural network for dynamic semantic annotation, and generating a structured patient information dataset can include the following steps:

[0069] Step S201: Obtain examination results and monitoring data from a multimodal medical data source; the multimodal medical data source includes imaging reports and real-time vital sign streams.

[0070] Step S202: extract dynamic time series features from the inspection results; dynamic time series features include fluctuations in inspection indicators and changes in medication records.

[0071] Step S203: construct a heterogeneous subgraph of dynamic time series features; the heterogeneous subgraph includes patient entity nodes and cross-source data relationship edges.

[0072] In step S204, a node embedding vector of the heterogeneous subgraph is generated through a graph neural network, and dynamic semantic labels are annotated; the dynamic semantic labels include disease stage and risk level.

[0073] Step S205 : Mapping the dynamic semantic tags and the original data into structured fields, and fusing the structured fields to obtain a patient information dataset.

[0074] This medical data processing pipeline achieves data structuring and semanticization through a multi-stage operation. First, examination results and monitoring data from multimodal medical data sources, such as imaging reports and real-time vital sign streams, are acquired. Dynamic time series features, such as fluctuations in test indicators and changes in medication records, are extracted. Based on these features, a heterogeneous subgraph is constructed with patients as entity nodes and cross-source data associations as edges. A graph neural network is used to generate node embedding vectors and annotate dynamic semantic labels, such as disease stage and risk level. Finally, the semantic labels are mapped to standardized fields with the raw data, which are then fused to generate a structured patient information dataset.

[0075] This implementation captures the temporal variations of medical data by extracting dynamic time series features. The construction of heterogeneous subgraphs and the application of graph neural networks enable semantic association and in-depth analysis of cross-source data. The generation of structured datasets addresses the standardization of multimodal medical data and improves data usability. This overall process enhances the semantic expression and time series analysis capabilities of medical data, providing a high-quality data foundation for subsequent applications such as clinical decision support and personalized medical plan development.

[0076] In one embodiment, the data fusion unit 1021 may include:

[0077] Step S301: Acquire multi-source heterogeneous data in a patient information data set; the multi-source heterogeneous data includes electronic medical records and vital sign monitoring records.

[0078] Step S302 : A cross-modal feature alignment result is generated by fusing feature distributions of different data sources using a dynamic weight allocation algorithm based on multi-source heterogeneous data.

[0079] Preferably, a dynamic weight allocation algorithm is used for different modal data such as text data in electronic medical records and time series values ​​of vital sign monitoring records, and weights are assigned according to the differences in the contribution of each data source in health assessment. The algorithm first analyzes the integrity, update frequency and clinical relevance of the data. For example, a higher weight is assigned to real-time blood glucose monitoring data because it has immediate reference value for diabetes management; a lower weight is assigned to relatively stable data such as medical history records. Through weighted summation and normalization processing, the feature distributions of data of different modalities are aligned, and the metric space and semantic expression of the data are unified, thereby generating cross-modal feature alignment results and eliminating fusion barriers caused by differences in data structure and format.

[0080] Step S303: Calculate a feature similarity matrix based on the cross-modal feature alignment result.

[0081] Step S304: Eliminate redundant feature fields according to the feature similarity matrix to obtain filtered non-redundant feature fields; the screening threshold of the redundant feature fields is dynamically adjusted by the adaptive fusion model.

[0082] In step S305, the non-redundant feature fields are used to enhance the key health indicators using the attention mechanism.

[0083] Step S306: Generate a dynamic unified patient health record index based on the enhanced key health indicators.

[0084] Step S307: updating the patient health record database based on the dynamic unified patient health record index.

[0085] Specifically, for multi-source heterogeneous data in patient information datasets, such as electronic medical records and vital sign monitoring records, a dynamic weight allocation algorithm is used to fuse the feature distributions of different data sources to generate cross-modal feature alignment results. Based on the alignment results, a feature similarity matrix is ​​calculated. An adaptive fusion model dynamically adjusts the screening threshold to remove redundant feature fields. Key health indicators are then enhanced through an attention mechanism. Finally, a dynamic and unified patient health record index is generated based on these enhanced indicators, completing the update of the health record database.

[0086] This embodiment solves the semantic difference problem of multi-source data through cross-modal feature alignment to improve data consistency; dynamically eliminates redundant features to reduce data noise and improve feature quality; the attention mechanism strengthens key indicators and highlights core information related to health assessment; and the dynamically updated health record database can reflect the patient's health status in real time, providing reliable data support for precision medicine.

[0087] In one embodiment, the feature similarity matrix can be calculated using the following formula:

[0088]

[0089] Among them, Si ,j represents the feature similarity between samples i and j, GCK() represents the graph convolution kernel function, x i and x j represents the feature vector of the i-th and j-th samples, L represents the normalized Laplace matrix, K represents the number of graph convolution layers, α k represents the attention weight of the kth layer, W k represents the feature transformation matrix, σ(·) represents the normalization function, ReLU(·) linear rectification activation function, b k Represents the bias vector of the kth layer.

[0090] Preferably, S i,j The feature similarity matrix reflects the semantic relevance of cross-modal features such as electronic medical records and vital sign data (e.g., the strength of association between blood glucose values ​​and the diagnosis of “diabetes”).

[0091] The σ(·) normalization function is used to ensure the interpretability of similarity values ​​and to facilitate the setting of redundant feature screening thresholds (such as S i,j >0.7 is considered redundant).

[0092] Specifically, the normalized Laplace matrix L is constructed based on the adjacency matrix A, which converts medical domain knowledge (such as the association between disease diagnosis and examination indicators) into graph structure information, so that the sample feature similarity S i,j The calculation not only depends on the data value, but also integrates the pathological logic; the feature transformation matrix W kCombined with the linear rectification function ReLU, hierarchical feature extraction is achieved, which gradually abstracts the original features (such as blood sugar value) into high-level semantics (such as pre-diabetes risk). At the same time, the attention weight α of the kth layer k Dynamically adjust the fusion ratio of each layer of features to strengthen the key information; finally, the normalization function σ maps the multi-layer feature interaction results into a probability form, so that S i,j Accurately reflect the semantic association strength between features in diabetes management scenarios, such as the correlation between insulin dosage and hypoglycemia risk.

[0093] This implementation introduces a priori data graph structures (e.g., pathological associations in medical knowledge graphs) through a normalized Laplacian matrix, avoiding the limitations of relying solely on numerical distances. Multi-layer graph convolution combined with an attention mechanism captures feature interactions at varying granularities (e.g., cross-modal associations between electronic medical record text and vital sign time series data). Dynamically adjusted attention weights and feature transformation matrices enable the model to adapt to distributional differences across different data sources. This overall approach improves the semantic accuracy of feature similarity calculations.

[0094] In one embodiment, Figure 3 As shown in FIG, extracting individual features from a health record database using an improved random forest algorithm to generate a personalized health risk assessment report may include the following steps:

[0095] Step S401: Acquire an individual health data set from a health record database; the individual health data set includes multi-dimensional physiological indicators and medical history records.

[0096] Step S402 : Calculate a feature weight matrix using an improved random forest algorithm based on the individual health data set.

[0097] Step S403: establishing dynamic feature screening rules based on the feature weight matrix to extract a key feature subset of the target individual.

[0098] Specifically, the system first constructs a feature importance score vector based on the feature weight matrix, which comprehensively considers the interdependence between features, time sensitivity, and domain expert knowledge. Subsequently, the dynamic feature screening rule adopts an adaptive threshold strategy: for highly variable features (such as blood sugar fluctuation values), a lower screening threshold is set to retain subtle changes; for stable features (such as blood type), the threshold is raised to exclude redundant information. A sliding window mechanism is introduced in the screening process to continuously evaluate the weight change trend of features in different time windows to ensure that dynamic features with predictive value are captured. Finally, the optimal feature subset is iteratively selected through a greedy algorithm, which minimizes information loss while maximizing the complementarity between features.

[0099] Step S404: Generate an individual health profile based on the key feature subset and input it into the risk prediction model to output a multi-dimensional risk probability vector.

[0100] Step S405: The multi-dimensional risk probability vector and the matching assessment dimensions are integrated to generate structured assessment data; the structured assessment data is converted into text paragraphs through a natural language processing engine.

[0101] Step S406: Integrate the temporal feature change trends in the text paragraphs and the individual health portrait to generate a final health risk assessment report.

[0102] Specifically, a set of individual health data containing multidimensional physiological indicators and medical history records is extracted from the health record database, and the feature weight matrix is ​​calculated using an improved random forest algorithm. This matrix comprehensively considers the reduction in feature impurity in the decision tree, feature co-occurrence frequency, time factors, and domain knowledge, and accordingly establishes dynamic feature screening rules to extract the key feature subset of the target individual. Based on the key feature subset, an individual health profile is generated, which is input into the risk prediction model to output a multidimensional risk probability vector, covering risk dimensions such as diabetic complications and disease progression. The multidimensional risk probability vector is fused with the corresponding evaluation dimension, converted into a text paragraph by a natural language processing engine, and the trend of changes in the temporal features in the individual health profile is integrated to finally generate a complete personalized health risk assessment report.

[0103] In this embodiment, the improved random forest algorithm combines multiple factors to construct a feature weight matrix, thereby enhancing the accuracy and comprehensiveness of the assessment of the importance of individual features; the dynamic feature screening rules effectively remove redundant information and improve data processing efficiency; the health portraits and multi-dimensional risk predictions generated based on key features realize the personalization and refinement of risk assessment; through the integration of natural language processing and time series features, the assessment report is both professional and readable, providing a scientific and intuitive decision-making basis for the health management of diabetic patients, and helping to improve the accuracy of disease prevention and intervention.

[0104] In one embodiment, the feature weight matrix can be constructed using the following formula:

[0105]

[0106] Among them, W represents the feature weight matrix, Wi ,j represents the relative importance of feature i to feature j, T represents the total number of trees in the random forest, ΔImpurity t,i represents the reduction in impurity of the i-th tree on feature i, C t represents the feature co-occurrence matrix of the t-th tree, τ i , τ jrepresents the timestamps of feature i and feature j, γ represents the time decay coefficient, K represents the domain knowledge matrix, and the strength of the prior relationship between features is defined by medical experts.

[0107] Preferably, W i,j The relative importance of feature i to feature j can, for example, reflect the strength of the association between features such as "blood sugar level" and "risk of diabetes".

[0108] C t The feature co-occurrence matrix of the t-th tree has a dimension of m×n, C t,ij It represents the frequency of feature i and feature j appearing in the same split path in the t-th tree, capturing the synergy between features. For example, “insulin usage” and “blood sugar fluctuation” often jointly influence decision-making, so the co-occurrence frequency is high.

[0109] K domain knowledge matrix, dimension is m×n, K ij It represents the strength of the prior relationship between feature i and feature j, and incorporates professional knowledge in the medical field. For example, the strength of the association between "blood pressure" and "cardiovascular complications" is pre-set by expert knowledge.

[0110] This implementation breaks through the limitations of traditional algorithms that rely solely on single-feature contributions by integrating the feature importance, feature co-occurrence relationships, temporal dynamics, and domain prior knowledge of random forests, enabling a multi-dimensional assessment of individual health data. The feature co-occurrence matrix captures feature synergies (such as the combined diagnostic value of blood glucose and glycosylated hemoglobin), the time decay factor increases the weight of recent data (such as prioritizing the latest physical examination indicators), and the domain knowledge matrix ensures that the algorithm complies with clinical logic (such as the strong correlation between hypertension and diabetic complications).

[0111] In one embodiment, the health management module 103 may include:

[0112] The health plan construction unit 1031 is used to:

[0113] Step S501, obtaining health risk indicators in the risk assessment report; health risk indicators include abnormal values ​​of physiological parameters and disease prediction probabilities.

[0114] Step S502: Input the health risk index into the generator model of the generative adversarial network, and output an initial health plan including exercise intensity and nutritional ratio.

[0115] Step S503 , based on the difference between the user's historical health data and the execution effect of the plan, the evaluation result of the discriminator model of the generative adversarial network on the initial health plan is calculated.

[0116] Preferably, the evaluation process of the discriminator model optimizes the generator model by quantifying the difference between historical data and program effects. Specifically, the discriminator first constructs a time series of the user's health status, including basic indicators before the execution of the program (such as fasting blood glucose values) and real-time monitoring data after execution (such as postprandial blood glucose fluctuations), and calculates the statistical distance between the two (such as mean square error, Kullback-Leibler divergence) as a preliminary difference metric. Domain knowledge constraints are then introduced to map the difference values ​​to clinical evaluation dimensions (such as blood glucose control target rate, complication risk change rate) to form a multidimensional evaluation vector. The discriminator performs a nonlinear transformation on the vector through a multi-layer perceptron and outputs a comprehensive scoring matrix that includes program rationality, intervention intensity matching, and time series adaptability. This matrix not only reflects the deviation between the current program and historical effects, but also provides parameter adjustment direction for the generator through the gradient backpropagation mechanism, ensuring that the subsequently generated health management program is closer to individual needs in terms of dynamic intervention cycle and monitoring threshold setting.

[0117] Step S504: adjust the weight parameters of the generator model according to the evaluation results to generate a health management plan including a dynamic intervention cycle and a monitoring threshold.

[0118] The health plan construction unit generates personalized health management plans through a multi-stage process. First, health risk indicators such as abnormal physiological parameter values ​​and disease prediction probabilities from the risk assessment report are input into the generator model of the generative adversarial network to generate an initial health plan that includes exercise intensity and nutritional balance. Subsequently, the discriminator model evaluates the initial plan based on the user's historical health data and the differences in the plan's execution results. Finally, the weight parameters of the generator model are adjusted based on the evaluation results to generate a health management plan that includes dynamic intervention cycles and monitoring thresholds.

[0119] In this example, the application of a generative adversarial network enables personalized health plan customization, generating targeted intervention measures based on individual health risk profiles. By evaluating the differences between historical data and implementation results, plan parameters are continuously optimized, improving the adaptability and effectiveness of the health management plan. Dynamic intervention cycles and monitoring thresholds enable timely adjustments to the plan based on changes in the patient's health status, enhancing the flexibility and precision of health management. This overall process ensures the scientific nature and personalization of the health management plan, helping to improve the health management outcomes for diabetic patients.

[0120] In one embodiment, the health management module 103 further includes:

[0121] The health plan optimization unit 1032 is used to:

[0122] Step S601 , matching and verifying the health management plan with the real-time physiological monitoring data to obtain a matching result; the matching verification is completed by calculating the deviation of the physiological indicators during the execution of the plan.

[0123] Step S602: If the deviation of the physiological indicators in the matching result exceeds a preset threshold, an abnormal indicator identification code set is generated.

[0124] Step S603: Input the abnormal indicator identification code set into the intervention measure matching model, and output the corresponding intervention parameter combination; the intervention measure matching model is constructed based on historical intervention effect data.

[0125] Preferably, after receiving the abnormal indicator identification code set, the intervention measure matching model conducts an in-depth analysis of the abnormal indicators based on the decision rules constructed based on the historical intervention effect data. The model first performs feature matching on the identification code with similar abnormal scenarios in the historical data, extracts the patient's basic information (such as age, medical history), physiological indicator combination (such as the correlation between blood sugar and insulin resistance) and the efficacy data of the corresponding intervention measures (such as the blood sugar recovery curve after the medication dose is adjusted) when the abnormality occurs. Through machine learning algorithms (such as gradient boosting trees or association rule mining), the correlation weights of different intervention parameters (such as the proportion of dietary structure adjustment, the increase or decrease of drug dosage, and the change in exercise intensity) and the improvement effect are calculated to generate an intervention parameter combination containing multi-dimensional adjustment suggestions. This combination not only takes into account the current status of the abnormal indicators, but also combines historical experience to predict the potential effects of different intervention measures to ensure that the output parameters can effectively respond to fluctuations in the patient's health status.

[0126] Step S604: updating the executable intervention instruction sequence in the health management plan according to the intervention parameter combination to obtain an optimized health management plan.

[0127] Specifically, real-time optimization of health management plans is achieved through a data-driven dynamic adjustment mechanism. This unit first verifies the health management plan by matching it with real-time physiological monitoring data, quantifying the effectiveness of the plan by calculating the deviation of physiological indicators during the execution of the plan. When the matching results show that the deviation of physiological indicators exceeds the preset threshold, the system generates a set of abnormal indicator identification codes, inputs them into the intervention measure matching model built based on historical intervention effect data, and outputs the corresponding intervention parameter combination. Finally, based on this parameter combination, the executable intervention instruction sequence in the health management plan is updated to complete the plan optimization.

[0128] This embodiment uses real-time monitoring and dynamic assessment to ensure that health management plans can promptly respond to changes in the patient's health status, avoiding delayed intervention measures. A matching model built based on historical intervention effect data provides scientific solutions for abnormal situations, improving the effectiveness of intervention measures. Automatic updates to intervention instruction sequences achieve closed-loop feedback on health management, enhancing the system's adaptability and intelligence. This overall process can effectively improve the accuracy and sustainability of diabetes management and enhance patient health management outcomes.

[0129] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0130] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the diabetes management system as described above when executing the computer program.

[0131] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0132] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0133] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A diabetes management system, characterized in that: The system comprises: The data processing module is used to obtain multimodal medical data sources, perform dynamic semantic annotation using graph neural networks, and generate structured patient information datasets; Database building blocks, including: a data fusion unit, configured to eliminate redundancy and enhance features of the patient information dataset using an adaptive fusion model, and construct a dynamic unified patient health record database; a risk assessment unit, configured to extract individual features from the health record database using an improved random forest algorithm to generate a personalized health risk assessment report; A health management module is used to customize a health management plan based on the risk assessment report using a generative adversarial network, and to generate intervention measures to update the health management plan if the real-time monitoring data after implementation exceeds a preset threshold.

2. The system according to claim 1, wherein: The method of obtaining a multimodal medical data source, performing dynamic semantic annotation using a graph neural network, and generating a structured patient information dataset includes: Obtaining examination results and monitoring data from the multimodal medical data source; the multimodal medical data source includes imaging reports and real-time vital sign streams; Extracting dynamic time series features from the inspection results; the dynamic time series features include fluctuations in inspection indicators and changes in medication records; Constructing a heterogeneous subgraph of the dynamic time series features; the heterogeneous subgraph includes patient entity nodes and cross-source data relationship edges; Generate node embedding vectors of the heterogeneous subgraphs through a graph neural network and annotate them with dynamic semantic labels; the dynamic semantic labels include disease stage and risk level; The dynamic semantic tags and the original data are mapped into structured fields, and the structured fields are fused to obtain a patient information dataset.

3. The system according to claim 1, wherein: The data fusion unit includes: Acquire multi-source heterogeneous data in the patient information data set; the multi-source heterogeneous data includes electronic medical records and vital sign monitoring records; According to the multi-source heterogeneous data, a dynamic weight allocation algorithm is used to fuse the feature distributions of different data sources to generate a cross-modal feature alignment result; Calculating a feature similarity matrix based on the cross-modal feature alignment result; Eliminating redundant feature fields according to the feature similarity matrix to obtain filtered non-redundant feature fields; the filtering threshold of the redundant feature fields is dynamically adjusted by the adaptive fusion model; The non-redundant feature fields are enhanced with an attention mechanism to enhance key health indicators; generating a dynamic unified patient health record index based on the enhanced key health indicators; A patient health record database is updated based on the dynamic unified patient health record index.

4. The system according to claim 3, characterized in that The feature similarity matrix is ​​calculated using the following formula: Among them, S i,j represents the feature similarity between samples i and j, GCK() represents the graph convolution kernel function, x i and x j represents the feature vector of the i-th and j-th samples, L represents the normalized Laplace matrix, K represents the number of graph convolution layers, α k represents the attention weight of the lth layer, W k represents the feature transformation matrix, σ(·) represents the normalization function, ReLU(·) linear rectification activation function, b k Represents the bias vector of the kth layer.

5. The system according to claim 1, wherein: The method of extracting individual features from the health record database using an improved random forest algorithm to generate a personalized health risk assessment report includes: Acquire an individual health data set from the health record database; the individual health data set includes multidimensional physiological indicators and medical history records; Calculating a feature weight matrix using an improved random forest algorithm according to the individual health data set; Establishing dynamic feature screening rules based on the feature weight matrix to extract a key feature subset of the target individual; Generate an individual health profile based on the key feature subset and input it into a risk prediction model to output a multidimensional risk probability vector; Fusion of the multi-dimensional risk probability vector and the matching assessment dimensions to generate structured assessment data; the structured assessment data is converted into text paragraphs through a natural language processing engine; The text paragraphs are integrated with the temporal feature change trends in the individual health portrait to generate a final health risk assessment report.

6. The system according to claim 5, characterized in that The feature weight matrix is ​​constructed by the following formula: Among them, W represents the feature weight matrix, W i,j represents the relative importance of feature i to feature j, T represents the total number of trees in the random forest, ΔImpurity t,i represents the reduction in impurity of the i-th tree on feature i, C t represents the feature co-occurrence matrix of the t-th tree, τ i , τ j represents the timestamps of feature i and feature j, γ represents the time decay coefficient, K represents the domain knowledge matrix, and the strength of the prior relationship between features is defined by medical experts.

7. The system according to claim 1, wherein: The health management module includes: Health program building blocks for: Obtain health risk indicators from the risk assessment report; the health risk indicators include abnormal values ​​of physiological parameters and predicted probability of disease; Inputting the health risk indicator into a generator model of a generative adversarial network to output an initial health plan including exercise intensity and nutritional ratio; An evaluation result of the discriminator model of the generative adversarial network on the initial health plan is calculated based on the difference between the user's historical health data and the execution effect of the plan; The weight parameters of the generator model are adjusted according to the evaluation results to generate a health management plan including a dynamic intervention cycle and a monitoring threshold.

8. The system according to claim 1, wherein: The health management module further includes: Health program optimization unit for: Matching and verifying the health management plan with the real-time physiological monitoring data to obtain a matching result; the matching verification is completed by calculating the deviation of physiological indicators during the execution of the plan; If the deviation of the physiological indicator in the matching result exceeds a preset threshold, generating an abnormal indicator identification code set; Inputting the abnormal indicator identification code set into an intervention measure matching model and outputting a corresponding intervention parameter combination; the intervention measure matching model is constructed based on historical intervention effect data; The executable intervention instruction sequence in the health management plan is updated according to the intervention parameter combination to obtain an optimized health management plan.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the system according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the system according to any one of claims 1 to 8 are implemented.

Images