A digital management method for early warning of diabetes complication risk

Abstract

The present application relates to the medical health information technology field, specifically to a kind of diabetic complication risk early warning digital management method, the present application collects hospital biochemical test, physiological time series of wearing equipment, three kinds of data of patient self-evaluation behavior, is preprocessed by event attribution weight reduction and credibility weighting;Based on the steady-state data of patient, individualized dynamic reference interval is constructed;Single-index slow deterioration characteristics and multi-organ physiological coupling network topological characteristics are extracted jointly;Through hierarchical bayesian network, the comprehensive risk score and the risk grade of complications such as nephropathy and cardiovascular disease are output;Based on causal matching, a personalized intervention plan is generated, and a closed-loop optimization mechanism of bidirectional iteration of risk model and intervention knowledge base is constructed.The present application can provide early warning window period, improve the accuracy of early warning and the efficiency of intervention, and is suitable for early warning and digital health management of diabetic complications.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, specifically a digital management method for early warning of diabetic complications. Background Technology

[0002] Diabetes mellitus is one of the most prevalent chronic metabolic diseases worldwide. The main danger of diabetes lies in the multi-system chronic complications caused by long-term metabolic disorders, encompassing various types such as diabetic nephropathy, diabetic cardiovascular disease, diabetic retinopathy, and diabetic peripheral neuropathy. These complications are generally characterized by insidious progression, early asymptomatic development, and strong irreversibility. Once they reach the clinically apparent stage, there are no effective reversible methods, making them a core cause of disability and death in diabetic patients. Numerous clinical evidence-based studies have confirmed that early identification and standardized intervention in the subclinical stage of complications can significantly delay disease progression and reduce the incidence of end-stage events. Therefore, accurate early risk warning is a core requirement for diabetes management.

[0003] Current technologies for early warning of diabetic complications have been widely researched and applied. The mainstream technical approaches can be divided into three categories: The first category is the traditional assessment method based on clinical scoring systems, represented by the UKPDS risk score and ADVANCE risk model. This method calculates the risk value by substituting discrete biochemical indicators such as glycated hemoglobin, blood pressure, blood lipids, and urinary microalbumin into a linear formula with fixed weights. The second category is the early warning model based on traditional machine learning. This model collects patient medical history, test results, and vital signs data, and uses algorithms such as logistic regression, random forest, and support vector machine to train binary or multi-class classification models to output the risk level. The third category is the dynamic early warning scheme based on wearable devices. This model uses continuous time-series data such as blood glucose and heart rate to achieve abnormal early warning through threshold judgment or simple time-series analysis.

[0004] However, existing technologies still face several insurmountable technical bottlenecks in practical clinical and health management applications: First, the early warning mechanism remains at the level of independent judgment of single indicators, ignoring the synergistic regulatory characteristics of the human physiological system, resulting in a delayed warning window. The essence of diabetic complications is the progressive instability of a multi-organ, multi-system regulatory network. Before a single biochemical indicator exceeds the clinical threshold, the coupling regulatory relationship between organs has often already undergone abnormal changes. Most existing solutions analyze each indicator independently, failing to pay attention to early changes in the correlation between indicators. They can only identify clinical-stage lesions with clearly abnormal indicators, but cannot capture network instability signals in the subclinical stage, leading to a generally late warning window and missing the optimal intervention opportunity.

[0005] Second, using a universal medical threshold as the risk assessment benchmark suffers from insufficient individual adaptability, leading to significant underreporting and false alarms. Significant differences exist in the physiological baselines of different patients. Some patients with long-term disease remain in a state of high metabolism, and although their indicators may not exceed the universal threshold, they have already shown significant deterioration relative to their own baseline, increasing their actual risk of complications and making them susceptible to underreporting by existing protocols. Conversely, some younger patients have lower basal metabolic rates, and even slight physiological fluctuations can trigger the universal threshold, easily causing false alarms. The inability of universal standards to match individual baseline differences directly limits the improvement of early warning accuracy.

[0006] Third, intervention plans rely on tiered rule matching, lacking personalized and precise derivation, resulting in a low intervention effectiveness rate. Most existing systems output standardized intervention recommendations based on risk levels, essentially matching and calling pre-set rule bases. This fails to predict the actual effects of different interventions based on individual patient characteristics, complication types, and lifestyle habits. In clinical practice, the same intervention plan often shows significant effects on some patients but no noticeable effect on others, leading to high trial-and-error costs and difficulty in improving overall intervention effectiveness.

[0007] Fourth, data preprocessing lacks differentiation between pathological and non-pathological abnormalities, and sporadic events can easily interfere with the stability of early warnings. Existing time-series data processing solutions mostly use general filtering and interpolation methods, which cannot distinguish between pathological abnormal fluctuations and physiological fluctuations caused by sporadic life events such as diet, exercise, emotions, and medication. The inclusion of non-pathological outliers in the feature calculation process can lead to feature distortion, trigger unnecessary early warnings, and reduce system reliability and user compliance.

[0008] Fifth, the risk model and intervention system are disconnected, failing to form a deeply collaborative closed-loop optimization mechanism. Most existing closed-loop solutions only optimize the risk model using feedback data, without feeding intervention effect data back into the intervention plan generation stage. The intervention knowledge base cannot continuously evolve with data accumulation, making it difficult to achieve a two-way improvement in both early warning accuracy and intervention effectiveness.

[0009] In summary, existing technologies cannot simultaneously solve multiple core problems such as network-level early signal recognition, individualized benchmark adaptation, precise intervention derivation, and non-pathological abnormality denoising. There is an urgent need for a brand-new digital management solution for early warning of diabetes complications, which requires technological breakthroughs in multiple dimensions such as early warning mechanism, judgment benchmark, intervention generation, and data processing. Summary of the Invention

[0010] To address the problems in existing technologies, this invention provides a digital management method for early warning of diabetes complications, which solves the technical problems in existing technologies such as delayed early warning window, poor individual adaptability, difficulty in constructing system coupling features from multi-resolution data, insufficient intervention accuracy, and easy interference of early warning stability by occasional events. It realizes full-process digital closed-loop management from network-level early identification to individualized precise intervention.

[0011] The technical solution adopted by this invention to solve its technical problem is: a digital management method for early warning of diabetic complications, comprising the following steps: S1. Multi-source heterogeneous health data collection and credibility-weighted preprocessing with event attribution: Three types of data sources were collected: discrete biochemical test data from hospitals, continuous physiological time-series data from wearable devices, and self-reported behavioral data from patients. The cumulative sum control chart algorithm was used to identify abnormal fluctuation segments in the time-series data. Attribution was determined by matching the data with dietary, exercise, and medication events during the same period. Abnormal data confirmed as occasional and non-pathological were downweighted. The dynamic credibility factor of each data source was calculated. All data were preprocessed with time axis alignment and noise reduction standardization to obtain a unified fusion dataset. S2. Adaptive construction of individual longitudinal steady-state baseline: Physiological index data of patients' historical steady-state periods are screened according to preset quantitative standards. The kernel density estimation method is used to construct the individualized 95% normal baseline interval for each physiological index. New steady-state data are screened periodically, and the baseline interval is incrementally updated through weighted fitting to form a risk reference baseline that is dynamically adjusted with the patient's status. S3. Joint extraction of multi-dimensional risk features: On the one hand, the deviation of each physiological indicator relative to its own baseline interval is calculated to extract the slow deterioration features that characterize progressive lesions; on the other hand, multi-resolution data are aligned to a unified time granularity, and a multi-organ physiological coupling and correlation network is constructed with mutual information between nodes as edge weights to extract three types of network topology features: abnormal edge ratio, global coupling efficiency, and key node betweenness change rate. S4. Hierarchical Bayesian Network Dynamic Risk Assessment: Construct a hierarchical Bayesian network containing a basic attribute layer, a risk factor layer, and a complication risk layer. Initialize network parameters based on prior clinical data, input individual patient basic attributes, dynamically update the weight coefficients of each risk feature through probabilistic inference, integrate slow degradation features and network topology features, and output a comprehensive risk score and the risk level of each complication item. S5. Personalized intervention and closed-loop optimization based on causal matching: Retrieve the historical intervention case database, filter similar case sets through propensity score matching, calculate the expected risk improvement effect of different intervention programs using counterfactual causal inference, and output the optimal personalized intervention combination; continuously collect post-intervention data to calculate the intervention effectiveness, dynamically adjust the intervention program, and update the risk assessment model and intervention knowledge base through cumulative feedback data increment.

[0012] Specifically, in step S1, the specific rules for attributing abnormal fluctuation events are as follows: blood glucose data is detected using a 3-hour sliding window and heart rate data is detected using a 1-hour sliding window using a cumulative sum control chart. The drift parameter is set to 0.5 times the data standard deviation, and the mutation judgment threshold is set to 3 times the data standard deviation. If the mutation period corresponds to events such as high-sugar diet or strenuous exercise, and the mutation amplitude is within the physiological impact range of the corresponding event, it is marked as a non-pathological abnormality. The data for that period is multiplied by a weighting coefficient of 0.2 to 0.4 and is not included in the calculation of core risk features.

[0013] Specifically, in step S2, the quantitative screening criteria for steady-state periods are: no acute disease markers or major life event markers for more than 7 consecutive days, and the coefficient of variation of the daily average values ​​of core physiological indicators is ≤8%; the kernel density estimation uses the Gaussian kernel function, and the bandwidth is calculated using the Silverman criterion; the discrete biochemical indicators use the test values ​​of multiple steady-state periods combined with the bootstrap method to construct the benchmark interval; when the benchmark interval is incrementally updated, the weight of historical benchmark data is 0.6, and the weight of newly added steady-state data is 0.4.

[0014] Specifically, in step S3, the multi-resolution data alignment method is as follows: minute-level continuous physiological data are downsampled into daily statistics, and monthly discrete biochemical indicators are mapped into daily sequences through linear interpolation, thus unifying them into daily granular data; the coupling network is dynamically constructed with a 7-day sliding window and a 1-day step size, and mutual information is calculated using the equal-frequency binning method, with the number of bins set to 8.

[0015] Specifically, in step S3, the calculation rules for the three types of network topology features are as follows: the abnormal edge ratio is the proportion of the number of edges whose edge weights deviate from the individual baseline interval to the total number of edges; the global coupling efficiency is the average shortest path efficiency of the weighted network, and the edge weights are negatively correlated with the path length; the key node betweenness change rate is the magnitude of the change in the weighted betweenness of the current window node relative to the baseline betweenness.

[0016] Specifically, in step S3, the slow degradation characteristics include three categories: the slope of the benchmark offset trend, the coefficient of variation of the fluctuation amplitude, and the cumulative amount of exceeding the limit; wherein the slope of the benchmark offset trend is the linear regression slope of the index center value relative to the benchmark midline within the time window, and the cumulative amount of exceeding the limit is the cumulative area of ​​the index exceeding its own benchmark interval in terms of time and amplitude.

[0017] Specifically, in step S4, the basic attribute layer of the hierarchical Bayesian network includes four nodes: disease course stratification, age stratification, number of comorbidities, and medication adherence level, which serve as prior variables to adjust the conditional probability of the risk factor layer; the risk factor layer includes 10-14 nodes for slow-deterioration class and network topology class; the complication risk layer includes three output nodes for kidney disease risk, cardiovascular disease risk, and retinopathy risk; the model parameters are initialized using maximum likelihood estimation, and the probabilistic inference uses the joint tree precise inference algorithm.

[0018] Specifically, in step S5, propensity score matching is performed based on 12 baseline features, using a 1:3 nearest neighbor matching method, and the matching caliper is set to 0.2 times the standard deviation of the propensity score; counterfactual causal inference uses the inverse probability weighting method to estimate the average treatment effect of different intervention programs, and the outcome indicator is the improvement rate of the comprehensive risk score 3 months after intervention.

[0019] Specifically, in step S5, the intervention plan includes four dimensions: dietary guidance, exercise prescription, medication management, and follow-up monitoring. Each dimension is divided into three intensity levels: basic, enhanced, and enhanced upgrade. The intervention effectiveness calculation formula is: intervention effectiveness = 0.4 × slow change feature improvement rate + 0.3 × network topology recovery rate + 0.3 × biochemical indicator compliance rate.

[0020] Specifically, in step S5, the incremental update cycle of the risk model and the intervention knowledge base is 90-180 days; the Bayesian network parameter update adopts weighted incremental learning, with the weight of new data set to 0.6 and the weight of historical data set to 0.4; the intervention knowledge base synchronously adds new case data and updates the causal effect estimates of each scheme.

[0021] The beneficial effects of this invention are: The early warning window period has been further advanced, enabling network-level early identification. By extracting topological anomaly features from multi-organ coupled networks, early instability signals in the regulatory relationship between organs can be identified before a single indicator exceeds the normal range, upgrading from "single indicator anomaly early warning" to "system network instability early warning." At the same time, it solves the technical obstacle of not being able to uniformly construct networks from multi-resolution data, further advancing the early warning window period for complications and buying more time for early intervention.

[0022] Individual adaptability is significantly improved, resulting in a substantial reduction in missed and false alarm rates. A longitudinal baseline interval is constructed based on the patient's own steady-state data from quantitative screening, replacing general medical thresholds as the risk assessment benchmark. Simultaneously, a suitable benchmark construction method is provided for discrete low-sampling indicators, truly achieving individualized risk assessment "based on the patient's own experience." This effectively solves the problem of missed and false alarms caused by differences in the physiological baselines of different patients, fundamentally improving the accuracy of early warning.

[0023] The precision of intervention plans has been significantly improved, and the effectiveness of interventions has been greatly enhanced. The intervention generation method, based on propensity score matching and inverse probability weighted counterfactual inference, has upgraded from standardized rule matching to individualized optimal plan prediction. This enables quantitative assessment of the expected effects of different interventions on specific patients, significantly reducing the cost of trial and error in interventions and substantially improving the overall effectiveness of interventions.

[0024] The stability of early warning systems has been significantly enhanced, and user compliance has improved. By combining cumulative and control chart methods with a quantitative mechanism for event attribution, the system effectively distinguishes between pathological abnormalities and physiological fluctuations caused by occasional life events. Non-pathological abnormalities are downweighted rather than simply deleted, which eliminates interference while preserving data integrity, greatly reducing unnecessary false alarms and improving the system's reliability and long-term user compliance.

[0025] A deep, two-way closed loop is formed, enabling continuous self-evolution. A closed-loop mechanism for two-way iteration between the risk model and the intervention knowledge base has been constructed. Intervention feedback data is used to optimize the accuracy of risk assessment and improve the accuracy of intervention matching, achieving simultaneous and continuous improvement in early warning capabilities and management effectiveness. The system performance evolves continuously with the duration of use.

[0026] The above-mentioned technical means support each other and work together: individual benchmarks provide individualized standards for feature extraction and risk assessment, improving the recognition accuracy of slow-changing features and network features; network features and single-indicator features complement each other, covering early lesions from the system level and the indicator level, respectively; causal intervention relies on accurate risk features to achieve matching, while feedback data optimizes the risk model in reverse. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Figure 1 A flowchart of a digital management method for early warning of diabetes complications provided by the present invention; Figure 2 A schematic diagram of the hierarchical Bayesian network structure for a digital management method for early warning of diabetes complications provided by the present invention; Figure 3 A schematic diagram of the multi-organ physiological coupling and correlation network topology of a digital management method for early warning of diabetes complications provided by the present invention; Figure 4 A schematic diagram of the individual longitudinal steady-state baseline construction and updating process for a digital management method for early warning of diabetes complications provided by the present invention; Figure 5 This is a schematic diagram of the closed-loop intervention and bidirectional iterative optimization logic of a digital management method for early warning of diabetes complications provided by the present invention. Detailed Implementation

[0029] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0030] like Figures 1-5As shown, the digital management method for early warning of diabetes complications according to the present invention includes the following steps: S1. Collection of multi-source heterogeneous health data and confidence-weighted preprocessing with event attribution: A multi-dimensional data collection system consisting of three types of data sources is constructed to comprehensively cover three dimensions: clinical testing, dynamic physiology, and behavioral habits. The first type is discrete biochemical test data from the hospital, including standardized clinical indicators such as glycated hemoglobin, urine microalbumin / creatinine ratio, fasting blood glucose, four blood lipids, blood pressure, and liver and kidney function. The sampling period is 1 to 6 months depending on the patient's condition. The second type is continuous physiological time-series data from wearable devices, including dynamic data such as continuous blood glucose, heart rate, heart rate variability, daily steps, sleep duration and sleep efficiency. The sampling frequency is from minutes to hours. The third type is self-reported behavioral data from the patient, including life event data such as diet records, exercise records, medication adherence, symptom self-reporting, and water intake, which are self-reported by patients daily or when triggered by events.

[0031] To address abnormal fluctuations in time-series data and quantify the distinction between pathological and non-pathological abnormalities, the Cumulative Sum Control Chart (CUSUM) algorithm is used to identify amplitude abrupt changes in the data. A 3-hour sliding window is used for blood glucose data, and a 1-hour sliding window for heart rate data. The drift parameter is uniformly set to 0.5 times the standard deviation of the corresponding data, and the mutation threshold is set to 3 times the standard deviation of the corresponding data. After identifying the abrupt change period, patient-reported events such as diet, exercise, and medication within the same time window are matched. The physiological impact patterns corresponding to the event type are used to determine whether the abrupt change can be explained by the life event. For example, a high-sugar diet corresponds to a blood glucose increase ≥2 mmol / L 1-2 hours after a meal, and moderate-intensity exercise corresponds to a heart rate increase ≥20 beats / minute. If these ranges are met, the data for that period is marked as a non-pathological, occasional abnormality. This data is multiplied by a weighting factor of 0.3 and is excluded from the core risk feature calculation, serving only as an auxiliary reference to avoid interference from occasional life events in risk assessment.

[0032] Dynamic reliability factors were calculated for the three types of data sources to quantify the quality level of different data: The hospital-side biochemical test data adopts a standardized testing process, with a baseline confidence level set at 0.92~0.98, and a default value of 0.95. The reliability of wearable device data is based on the device's most recent calibration time. It decreases by 0.02 for every 7 days beyond the calibration cycle, and decreases by an additional 0.05 for every 10% increase in the daily data missing rate, with a minimum of 0.60. The reliability of patient-reported self-assessment data consists of two parts: completeness of completion and consistency of event attribution. A base score of 0.8 is given for completeness of completion above 90%, with a deduction of 0.1 points for every 10% decrease. A score of 0.05 is added for every 10% increase in event attribution consistency, with a maximum of 0.85.

[0033] After completing the attribution and credibility calculations, all data are preprocessed with time axis alignment and standardization: linear interpolation is used to fill in a small number of missing values ​​for continuous time series data, and high-frequency random noise is removed by a 5-point sliding window filter; linear interpolation is used to map discrete biochemical data to a unified daily time axis; finally, the three types of data are aligned by timestamp and combined with credibility weights to form a fused dataset.

[0034] S2. Adaptive construction of individual longitudinal steady-state baseline: Instead of relying on general medical thresholds, we establish a unique longitudinal steady-state baseline range for each patient as an individualized reference standard for risk assessment.

[0035] First, perform quantitative screening of steady-state time periods: review the patient's historical health data and screen time periods that simultaneously meet three conditions as steady-state time periods: First, the continuous duration must be no less than 7 days; Second, there were no signs of acute illness or major life events during the period (major life events include acute infection, trauma, severe sleep disorder, and abnormal eating for more than 3 consecutive days). Third, the coefficient of variation of the daily average values ​​of core physiological indicators is ≤8%. The data for this period represents the patient's normal physiological state.

[0036] For indicators with different time resolutions, an appropriate benchmark construction method is adopted: For continuous physiological time-series data, based on daily statistics of steady-state periods, the kernel density estimation method is used to calculate the 95% confidence interval for each indicator. A Gaussian kernel is used as the kernel function, and the bandwidth is adaptively calculated using the Silverman criterion. The final confidence interval is the individualized normal benchmark interval for that indicator. For discrete biochemical indicators with low sampling frequency, such as glycated hemoglobin and urinary microalbumin, test values ​​from multiple steady-state periods are extracted, and expanded samples are generated by bootstrap resampling. Then, nonparametric estimation is used to construct the 95% benchmark interval to solve the problem of inaccurate benchmark estimation under small sample conditions.

[0037] An adaptive weighted incremental update mechanism is set up for the baseline interval, with an update cycle of 30-90 days. At each update cycle, newly added steady-state period data are selected, and the baseline interval is updated using a weighted kernel density fitting method. Historical baseline data has a weight of 0.6, and newly added steady-state data has a weight of 0.4. During the update process, data from abnormal periods and acute events are not included in the baseline fitting to avoid abnormal data contaminating the baseline. Through dynamic updates, the baseline interval can be adjusted synchronously with the slow changes in the patient's long-term physiological state, always maintaining an accurate representation of the patient's current normal state.

[0038] S3. Joint extraction of multi-dimensional risk features: Risk characteristics are extracted from two dimensions: gradual degradation of a single indicator and system network instability, to comprehensively cover different manifestations of early complications.

[0039] The first category is the slow degradation feature of a single indicator, which characterizes the progressive pathological trend of a single physiological indicator. Using the individual longitudinal baseline interval as a reference, the degree of deviation of each physiological indicator within a 30-day time window is calculated. Combined with ensemble empirical mode decomposition to remove high-frequency fluctuation components, three types of slow degradation features are extracted: First, the baseline offset trend slope, which is the linear regression slope of the indicator's central value relative to its own baseline midline, characterizes the overall drift direction and rate of the indicator. A continuously positive and increasing slope indicates progressive deterioration of the indicator. Second, the coefficient of variation of the fluctuation range, that is, the coefficient of variation of the daily fluctuation range of the indicator, which characterizes the degree of decline in physiological stability. Third is the cumulative amount of out-of-bounds, which is the cumulative area of ​​time and amplitude when the indicator exceeds its own baseline range, representing the cumulative degree of deterioration damage.

[0040] The second category is the topological anomaly features of multi-organ coupled networks, which characterize the overall instability of physiological regulatory systems. First, the alignment problem of multi-resolution data is addressed: minute-level continuous physiological data is downsampled to daily statistics (daily average, daily fluctuation range), and monthly discrete biochemical indicators are mapped to daily sequences through linear interpolation, unifying all node data to daily granularity. Then, a multi-organ coupled network is constructed. Five key physiological indicators are selected for network nodes: fasting blood glucose, systolic blood pressure, heart rate variability, sleep efficiency, and estimated urinary microalbumin, corresponding to core regulatory systems such as glucose metabolism, cardiovascular system, autonomic nervous system, and kidneys, respectively. A 7-day sliding window and a 1-day step size are used to dynamically calculate the mutual information value between pairs of nodes as edge weights. Mutual information is calculated using an equal-frequency binning method, with 8 bins. Higher mutual information indicates a stronger coupling regulatory relationship between the two physiological systems.

[0041] Three types of topological features are extracted based on dynamic coupling networks: first, the proportion of anomalous edges, i.e., the proportion of edges whose weights deviate from the individual baseline interval to the total number of edges; a higher proportion indicates more abnormalities in the regulatory relationships between organs; second, the weighted global coupling efficiency, calculated using the following formula: ; Where n is the number of network nodes, Let the weight of the edge between nodes i and j be . The first is the shortest path length between two nodes; a decrease in efficiency indicates a reduction in the overall coordination capability of the regulation network. The second is the rate of change of the betweenness of key nodes. The weighted betweenness algorithm is used to calculate the betweenness value of each node and to calculate the change in the betweenness of the current window relative to the baseline betweenness, reflecting the degree of functional abnormality of the core regulation nodes.

[0042] S4. Dynamic Risk Assessment of Hierarchical Bayesian Networks: A three-layer hierarchical Bayesian network model is constructed to fuse two types of features for personalized dynamic risk assessment. The network structure consists of three layers: The first layer is the basic attribute layer, which includes four nodes: disease course stratification, age stratification, number of comorbidities, and medication adherence level. These nodes serve as prior variables to adjust the conditional probability distribution of the risk factor layer nodes, reflecting the impact of individual patient differences on risk weights. The second layer is the risk factor layer, which contains 12 specific factor nodes, including slow degradation feature class and network topology feature class, corresponding to the two types of multi-dimensional features extracted in step S3. The third layer is the complication risk layer, which includes three output nodes: diabetic nephropathy risk, diabetic cardiovascular disease risk, and diabetic retinopathy risk, each corresponding to a different type of complication risk.

[0043] During the model initialization phase, based on a large clinical dataset, the maximum likelihood estimation method is used to train the prior conditional probability table for each node, forming general initial parameters. When assessing a single patient, the patient's basic attributes are input into a Bayesian network, and a joint tree precise inference algorithm is used for probabilistic inference, dynamically updating the posterior weight coefficients of each risk feature. For example, for patients with a long disease duration and hypertension, the weights of renal-cardiovascular related coupling features and urinary microalbumin shift features are automatically increased; for newly diagnosed young patients, the weights of glucose metabolism-autonomic nervous system coupling features and lifestyle-related features are automatically increased.

[0044] In the risk calculation phase, all feature values ​​are standardized to 0-1 and multiplied by the corresponding dynamic weight coefficients. The weighted sum is then used to obtain the comprehensive risk score. Simultaneously, the sub-risk scores for the three main complications are calculated separately. Based on preset thresholds, the risk is divided into three levels: low risk (0-30 points), medium risk (31-70 points), and high risk (71 points and above). The corresponding graded warning results are then output.

[0045] S5. Personalized intervention and closed-loop optimization based on causal matching: Abandoning the rule-matching intervention generation model, we adopt causal inference to accurately deduce personalized optimal intervention plans, while constructing a deep closed-loop mechanism for bidirectional iteration of risk assessment and intervention optimization.

[0046] First, a historical intervention case knowledge base is constructed. Each record in the base contains information such as 12 baseline characteristics of the patient, details of the intervention plan, intervention period, changes in risk characteristics before and after the intervention, and clinical outcomes. For the current patient to be intervened, their baseline feature vector and risk feature vector are extracted. A propensity score matching algorithm is used to screen a set of similar cases in the case base. The matching dimensions include 12 features such as age, disease duration, BMI, baseline glycated hemoglobin, number of comorbidities, and risk level. A 1:3 nearest neighbor matching method is used, and the matching caliper is set to 0.2 times the propensity score standard deviation to ensure that the baselines of the matched cases are highly similar.

[0047] Based on a similar case set, a counterfactual causal inference was performed using an inverse probability weighted method. The average treatment effect of different intensities of interventions across four intervention dimensions—dietary adjustment, exercise prescription, medication management, and follow-up monitoring—was calculated, representing the expected improvement rate in the overall risk score. Each intervention dimension was divided into three intensity levels: basic, enhanced, and upgraded, resulting in multiple intervention combinations. The intervention dimension combination with the highest expected overall improvement was selected to form a personalized optimal intervention plan for the patient, specifying the exact intensity, frequency of implementation, and follow-up period for each dimension.

[0048] Establish a dynamic evaluation and adjustment mechanism for intervention effectiveness: Continuously collect patients' health data after intervention, and calculate the intervention effectiveness using a weighted average of the improvement rate of slow deterioration characteristics, the recovery rate of network topology characteristics, and the rate of biochemical indicator compliance. The calculation formula is as follows: Intervention effectiveness = 0.4 × slow change characteristic improvement rate + 0.3 × network topology recovery rate + 0.3 × biochemical indicator compliance rate; When the intervention efficacy is below 30%, similar case matching and causal inference are repeated, and the dimensional focus and intensity of the intervention plan are adjusted; when the intervention efficacy is above 50%, the current intervention plan is maintained and the follow-up interval is appropriately extended to avoid over-intervention.

[0049] An incremental update mechanism for the model and knowledge base is established, with an update cycle of 90-180 days. At each update cycle, accumulated intervention feedback data and clinical outcome data are added to the risk model training set and the intervention knowledge base, respectively. Bayesian network parameters are updated using a weighted incremental learning approach, with a weight of 0.6 for new data and 0.4 for historical data. Simultaneously, new case data is added to the intervention knowledge base, and the causal effect values ​​of each intervention are re-estimated. Through incremental updates, the accuracy of the risk assessment model continuously improves, and the matching accuracy of the intervention knowledge base continuously optimizes, forming a virtuous cycle of bidirectional evolution.

[0050] Example 1: Complication Early Warning Management in Patients with Long-Term Type 2 Diabetes and Hypertension: This example selected 50 patients with type 2 diabetes and hypertension followed up in the endocrinology department of a tertiary hospital as the research subjects. All patients had a disease duration of more than 8 years, had primary hypertension, and had not yet developed any clear clinical complications. Their ages ranged from 52 to 72 years. The patients were randomly divided into an experimental group and a control group, with 25 patients in each group. The experimental group received complication early warning management using the method described in this invention for 6 months, while the control group received routine management using the traditional fixed-weight risk scoring method.

[0051] The specific management process for the experimental group is as follows: First, multi-source data collection and confidence-weighted preprocessing with event attribution were performed. Hospital-side data was collected every 3 months, including indicators such as glycated hemoglobin, urinary microalbumin / creatinine ratio, fasting blood glucose, blood lipids, and blood pressure, with a baseline confidence level set at 0.95. Wearable devices, using a medical-grade continuous glucose monitor and a multi-functional smart bracelet, continuously sampled blood glucose every 5 minutes, and summarized heart rate, heart rate variability, and sleep data hourly. The initial confidence level of the wearable devices was 0.90, decreasing by 0.02 every 7 days beyond the calibration period, with an additional decrease of 0.05 for a single day with a missing data rate exceeding 10%. Patients reported their diet, exercise, and medication information daily via mobile devices, and could mark special events such as high-sugar diets and strenuous exercise during the reporting process.

[0052] In the abnormal fluctuation attribution phase, a cumulative sum control chart algorithm was used to detect mutations: blood glucose data were processed using a 3-hour window with a drift parameter of 0.5 times the standard deviation and a threshold of 3 times the standard deviation; heart rate data were processed using a 1-hour window with the same parameters. After identifying mutations, matching them with concurrent events was performed. For example, if a patient's blood glucose level increased by 2.8 mmol / L 1.5 hours after lunch, corresponding to a reported high-sugar, high-fat lunch event, this falls within the range of postprandial blood glucose elevation and is marked as a non-pathological abnormality. Data from this time period was multiplied by a 0.3 weighting factor and excluded from core feature calculations. The reliability of patient self-reported data was calculated by combining completeness and event matching. A completeness of over 90% was assigned a base score of 0.8, which could be increased to 0.85 for higher matching scores. All data were uniformly aligned to a daily timeline, forming a fused dataset with reliability weights.

[0053] The second step was to construct individual longitudinal steady-state baselines. Historical data from the three months prior to enrollment for each patient were retrospectively analyzed, and periods of more than seven consecutive days without acute events and with a daily mean coefficient of variation of core indicators ≤8% were selected as steady-state periods. For example, a 62-year-old patient with an 11-year disease duration had a daily mean coefficient of variation of 6.2% for their steady-state fasting blood glucose, meeting the steady-state criteria. Using Gaussian kernel density estimation and Silverman bandwidth calculation, their individualized 95% baseline range for fasting blood glucose was determined to be 5.3–6.1 mmol / L. Although this was generally within the high end of the general normal range, it represented the patient's normal condition. Urinary microalbumin, a discrete indicator measured quarterly, had three steady-state test values ​​extracted. After resampling using the bootstrap method, a baseline range of 12–22 mg / g was constructed. The baseline ranges were updated every 60 days, with historical data weighted at 0.6 and new steady-state data weighted at 0.4.

[0054] The third step involves the joint extraction of multi-dimensional risk features. For single-indicator slow-varying features, the slope of the deviation trend, coefficient of variation of fluctuation, and cumulative out-of-bounds amount of indicators such as blood glucose, blood pressure, and heart rate variability relative to their baselines over the past 30 days were calculated, totaling nine slow-varying features. For coupled network features, all data were standardized to a daily granularity. Using five indicators—fasting blood glucose, systolic blood pressure, heart rate variability, sleep efficiency, and estimated urinary microalbumin—as nodes, a 7-day sliding window and a 1-day step size were used. An 8-bin equal-frequency binning method was employed to calculate pairwise mutual information, constructing a dynamic coupled network. Three network features were extracted: the proportion of abnormal edges, weighted global coupling efficiency, and the rate of change of betweenness in core nodes. The calculation results for this patient showed that the mutual information of the blood glucose-systolic blood pressure coupling edge decreased by 28% compared to the baseline, the proportion of abnormal edges reached 40%, and the global coupling efficiency decreased by 19% compared to the baseline, suggesting early instability in the cardiovascular-renal regulatory network, but individual biochemical indicators remained within the generally normal range.

[0055] The fourth step involved performing a hierarchical Bayesian network dynamic risk assessment. Basic attributes such as the patient's 11-year disease duration, age of 62, presence of one type of hypertension, and good medication adherence were input into the Bayesian network. After inference using a joint tree algorithm, renal-cardiovascular coupling features and urinary microalbumin shift features received higher weights. The weighted calculation yielded a comprehensive score of 65 points for the patient, placing them in the intermediate-risk category; the cardiovascular sub-score was 73 points, and the renal sub-score was 68 points, both falling within the intermediate-to-high-risk range.

[0056] The fifth step involves generating a personalized intervention plan and implementing closed-loop management. Based on the patient's 12 baseline characteristics, a 1:3 nearest neighbor matching was performed in the historical case database, with a caliper deviation of 0.2, resulting in 32 highly similar cases. The average treatment effect of different intervention combinations was calculated using an inverse probability weighted method. The results showed that the combination of intensive blood pressure control and a low-salt, high-quality protein diet had the highest expected risk improvement rate, reaching 47%. The system outputs a targeted intervention plan: in terms of diet, strict control of sodium intake (≤5g per day) and a high-quality, low-protein diet are recommended; in terms of exercise, low-intensity aerobic exercise is recommended, 5 times a week for 30 minutes each time; in terms of medication, reminders for antihypertensive medication are strengthened, and the blood sugar control plan is maintained; in terms of follow-up, blood pressure and urinary microalbumin are monitored monthly.

[0057] Intervention effectiveness is calculated every 30 days using the formula: 0.4 × slow-change characteristic improvement rate + 0.3 × network recovery rate + 0.3 × biochemical target achievement rate. If the effectiveness is below 30%, the treatment plan is readjusted; if it is above 50%, the follow-up interval is extended. Within the 6-month management period, the risk model and intervention knowledge base are incrementally updated every 3 months, with new data having a weight of 0.6.

[0058] Six months later, the statistical results showed that the experimental group identified 10 patients with early network instability. Among them, 7 patients showed significant improvement in network and slow characteristics after timely intervention and did not progress to abnormal indicators. The early warning window for complications in the experimental group was on average 38 days earlier than that in the control group, with a warning accuracy rate of 88.0% and a false alarm rate of 8.0%. In contrast, the control group identified only 4 patients with abnormal indicators, with a warning accuracy rate of 48.0% and a false alarm rate of 28.0%. At the same time, the intervention effectiveness rate in the experimental group was 51% higher than that in the control group, verifying the significant advantages of this method in high-risk populations with long disease duration.

[0059] Example 2: Individualized Risk Warning for Young Patients with Newly Diagnosed Type 2 Diabetes: This example selected 40 newly diagnosed type 2 diabetes patients (aged 28-42 years, with a disease duration of less than one year and no clear complications) registered at a community health service center. They were randomly divided into an experimental group and a control group, with 20 patients in each group. The experimental group received 12 months of health management using the method of this invention, while the control group received routine annual physical examinations plus a general threshold warning model. This verified the individualized warning capability of this method in low-risk young populations.

[0060] The experimental group management process is as follows: During the data collection phase, the hospital conducted biochemical tests every 6 months, including glycated hemoglobin, liver and kidney function, blood lipids, and urinary microalbumin, with a baseline reliability of 0.95. Wearable devices used consumer-grade smart bracelets and fingertip blood glucose monitoring. Patients self-tested their fasting and postprandial blood glucose daily and entered the data into the system, while also collecting heart rate, steps, and sleep data. The reliability of the fingertip blood glucose data was dynamically adjusted based on operational standardization and numerical consistency, ranging from 0.70 to 0.85; the reliability of the bracelet data was 0.75 to 0.85. Patients reported their diet, exercise, and medication status weekly, with self-assessed data reliability ranging from 0.60 to 0.80.

[0061] In the abnormal fluctuation attribution phase, considering the irregular eating and exercising habits of young patients, the focus was on event matching of fluctuations in postprandial blood glucose and post-exercise heart rate. For example, if a patient's blood glucose increased by 3.1 mmol / L after a weekend dinner party, this would be reported as a high-sugar dinner party event, meeting the criteria for non-pathological abnormality and thus being downweighted to avoid over-alarming young patients. This phase eliminated a large number of non-pathological abnormal fluctuations caused by dinner parties, staying up late, or sudden exercise.

[0062] The core advantage of this implementation is the establishment of individual baselines. A unique baseline range is established for each patient: data from the first two months after diagnosis when the condition is stable are selected, and the steady-state period is chosen based on a coefficient of variation ≤8%. For example, Patient A, 32 years old, diagnosed 6 months ago, has a steady-state fasting blood glucose level consistently around 4.0 mmol / L. Kernel density estimation yields a baseline range of 3.8–4.4 mmol / L. If fasting blood glucose subsequently rises to 5.0 mmol / L, although it does not exceed the general normal threshold of 6.1 mmol / L, it deviates by 13.6% relative to the baseline, and the system will identify it as abnormal. Patient B, 38 years old, has a baseline fasting blood glucose level consistently around 5.8 mmol / L. The baseline range is 5.5–6.0 mmol / L, close to the upper limit of the general threshold, but without significant deviation relative to the baseline, thus avoiding misjudgment as high-risk. Discrete biochemical indicators are constructed using multiple steady-state test values ​​combined with a bootstrap method. The baseline range is updated every 90 days.

[0063] In the feature extraction phase, focusing on the characteristics of newly diagnosed patients, the key focus is on changes in the glucose metabolism-autonomic nervous system coupling network and the slow-changing characteristics of blood glucose fluctuation stability. The core early manifestation of complication progression in newly diagnosed patients is a decline in metabolic regulation capacity, rather than an increase in absolute values ​​of indicators. Coupled network features and baseline deviation features can accurately capture these changes. For example, in patient A at the 6th month of follow-up, fasting blood glucose remained within the generally normal range, but the mutual information of the blood glucose-heart rate variability coupling decreased by 24% compared to the baseline, the global coupling efficiency decreased by 17%, and the coefficient of variation of fluctuation amplitude increased by 21%, indicating a weakened autonomic nervous system regulation of blood glucose, which is a very early signal of complications.

[0064] In the risk assessment phase, hierarchical Bayesian networks automatically increase the weight of network topology features and behavioral-related features for young, newly diagnosed patients, while reducing the weight of traditional biochemical indicators, focusing more on early changes in regulatory capacity. Risk grading uses a standard based on the degree of relative deviation rather than a universal absolute value standard. In the intervention plan generation phase, lifestyle interventions are emphasized for young patients, and causal matching is used to screen the intervention plan most suitable for the patient's lifestyle. For example, for patients who sit for long periods at work, causal inference shows that fragmented exercise plans have better adherence and improvement effects than concentrated exercise plans, and the system outputs a fragmented exercise intervention plan of 6 times a day, 5 minutes each time; for patients who frequently order takeout, an optimized takeout meal selection plan is output to improve patient adherence.

[0065] The 12-month follow-up results showed that the experimental group identified 9 patients with early-stage metabolic regulatory impairment. After intensive lifestyle intervention, 8 of these patients showed significant improvement in network and chronic features without progressing to abnormal biochemical indicators. In contrast, the control group only had 2 patients with mildly abnormal biochemical indicators detected during annual physical examinations, at which point metabolic regulatory capacity had already shown clear impairment. Furthermore, the false alarm rate in the experimental group was only 5.0%, significantly lower than the 35.0% in the control group, greatly reducing unnecessary interference with young patients. This embodiment validates that combining individual longitudinal benchmarks with network features can achieve accurate early warning in newly diagnosed young individuals while effectively controlling the false alarm rate.

[0066] Example 3: Batch Digital Management and Closed-Loop Iterative Validation of Community Diabetic Population: This example was conducted in a community health service center in a certain street. A total of 240 patients with type 2 diabetes were included in the management, with an age range of 35 to 78 years and a disease duration range of 1 to 18 years. The applicability of this method in batch population management and the effect of the two-way closed-loop iterative mechanism on improving system performance were fully validated.

[0067] Management was divided into two phases. The first phase, lasting three months, used an initial version of the risk model and intervention knowledge base for early warning management. The second phase, lasting three months, used the intervention feedback and follow-up data accumulated in the first phase to incrementally update the risk model and intervention knowledge base before continuing operation. The early warning accuracy, false alarm rate, intervention effectiveness, and management efficiency were compared between the two phases to verify the effectiveness of the closed-loop iteration.

[0068] During batch management, the system automatically completes data collection, event attribution, baseline updates, feature extraction, risk grading, and intervention plan generation. Community doctors only need to manually review and provide intervention guidance for medium- and high-risk patients. The system automatically generates a monthly risk report for each patient, marking the risk level, core abnormal characteristics (single indicator deviation or network instability), and intervention recommendations, significantly reducing the workload of doctors. Tiered management is implemented for different risk levels: low-risk patients are automatically followed up by the system every 3 months; medium-risk patients are followed up monthly by their family doctors; and high-risk patients are jointly assessed by specialists to develop intensive intervention plans.

[0069] The causal intervention matching mechanism significantly improves intervention efficiency in batch management: the system automatically generates personalized intervention plans for each patient, replacing traditional uniform health education. For example, it automatically matches a low-salt, high-quality, low-protein diet plan for patients with combined kidney disease risk, and automatically matches a sodium control plus moderate-intensity aerobic exercise plan for patients with combined cardiovascular risk, eliminating the need for doctors to develop individual plans. The system also automatically adjusts the intervention method according to the patient's lifestyle, improving adherence.

[0070] During the incremental update phase, follow-up data, intervention effect data, and clinical outcome data of 240 patients accumulated over the first three months were added to the risk model training set and the intervention knowledge base, respectively. The Bayesian network parameters were updated using weighted incremental learning, with a weight of 0.6 for new data and 0.4 for historical data. Simultaneously, 240 new case data were added to the intervention knowledge base to re-estimate the causal effect values ​​of each intervention regimen and optimize the accuracy of intervention matching. Furthermore, the risk grading thresholds were fine-tuned based on actual follow-up results, and the criteria for classifying intermediate- and high-risk intervals were optimized.

[0071] Comparing the operational data from the two phases: In the first phase, the initial model achieved a comprehensive early warning accuracy of 79.2%, a false alarm rate of 20.8%, and an intervention effectiveness rate of 62.5%, with community doctors managing an average of 80 patients per patient. In the second phase, after incremental updates, the comprehensive early warning accuracy improved to 90.8%, the false alarm rate decreased to 9.2%, and the intervention effectiveness rate increased to 78.3%. Simultaneously, the system's automated intervention capabilities were enhanced, and the average doctor's management efficiency increased to 130 patients per doctor. Furthermore, overall patient satisfaction in the second phase increased by 34% compared to the first phase, and the reduction in false alarms significantly improved user compliance. This embodiment verifies that a bidirectional closed-loop iterative mechanism can sustainably improve system performance and management efficiency, making it highly suitable for large-scale community diabetes management scenarios.

[0072] Comparative Example 1, Traditional Static Indicator Fixed Weight Early Warning Method: This comparative example corresponds to the existing mainstream technical solution, which only uses static biochemical indicators such as glycated hemoglobin, urinary microalbumin, blood pressure, and blood lipids to calculate the risk of complications through a fixed weight risk scoring formula, without involving individual benchmarks, network characteristics, causal intervention and other technical means.

[0073] Comparing Control Example 1 with the experimental group of Example 1, the results showed that the warning window of the control example was 38 days later than that of the experimental group, the warning accuracy was 40 percentage points lower, the false alarm rate was 20 percentage points higher, and the intervention effectiveness was 51% lower. This control example illustrates that traditional methods rely solely on static indicators and fixed weights, which cannot identify early network instability signals or adapt to individual differences, resulting in significant limitations in both warning accuracy and intervention effectiveness.

[0074] Comparative Example 2, an early warning method with single-indicator slow-change features but using a general threshold: This comparative example retains the single-indicator slow-change deterioration feature extraction steps, but uses a general medical threshold as the judgment benchmark, does not construct individual longitudinal benchmarks, and does not extract coupled network features, and uses fixed values ​​for risk weights.

[0075] Comparing Control Example 2 with the experimental group of Example 1, the results showed that the warning window period of Control Example 2 was improved compared with the traditional method, 16 days earlier than Control Example 1, but still 22 days later than the experimental group; the warning accuracy rate was 64.0%, lower than the 88.0% of the experimental group; among them, patients with lower baseline values ​​had a higher false negative rate, while patients with higher baseline values ​​had a higher false positive rate. This control example illustrates that relying solely on the slow-varying characteristics of a single indicator can provide early warning to a certain extent, but lacks the support of individual baselines and network features, resulting in limited overall performance improvement, thus verifying the synergistic gain of combining multiple technologies.

[0076] Comparative Example 3, an early warning method with coupled network features but using a general threshold: This comparative example retains the multi-organ coupled network feature extraction step, but uses a general population network threshold to judge anomalies, does not construct an individual longitudinal baseline, and does not use causal intervention matching.

[0077] Comparing Control Example 3 with the experimental group of Example 2, the results showed that the warning window period of Control Example 3 was significantly improved compared with the traditional method, but the false alarm rate was as high as 27.5%, far higher than the 5.0% of the experimental group. In particular, for young patients with special baselines, the adaptability of the general network threshold was very poor, frequently triggering false alarms. This control example illustrates that coupled network features must be combined with individual longitudinal benchmarks to achieve the best effect. The general threshold will significantly offset the advantages of network features, further verifying the necessity of the synergistic effect of the various technical means of this invention.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital management method for early warning of diabetes complications, characterized in that, Includes the following steps: S1. Multi-source heterogeneous health data collection and credibility-weighted preprocessing with event attribution: Three types of data sources were collected: discrete biochemical test data from hospitals, continuous physiological time-series data from wearable devices, and self-reported behavioral data from patients. The cumulative sum control chart algorithm was used to identify abnormal fluctuation segments in the time-series data. Attribution was determined by matching the data with dietary, exercise, and medication events during the same period. Abnormal data confirmed as occasional and non-pathological were downweighted. The dynamic credibility factor of each data source was calculated. All data were preprocessed with time axis alignment and noise reduction standardization to obtain a unified fusion dataset. S2. Adaptive construction of individual longitudinal steady-state baseline: Physiological index data of patients' historical steady-state periods are screened according to preset quantitative standards. The kernel density estimation method is used to construct the individualized 95% normal baseline interval for each physiological index. New steady-state data are screened periodically, and the baseline interval is incrementally updated through weighted fitting to form a risk reference baseline that is dynamically adjusted with the patient's status. S3. Joint extraction of multi-dimensional risk features: On the one hand, the deviation of each physiological indicator relative to its own baseline range is calculated, and the slow deterioration features that characterize progressive lesions are extracted. On the other hand, multi-resolution data are aligned to a unified time granularity, and a multi-organ physiological coupling and correlation network is constructed with mutual information between nodes as edge weights. Three types of network topology features are extracted: the proportion of abnormal edges, global coupling efficiency, and the rate of change of betweenness of key nodes. S4. Hierarchical Bayesian Network Dynamic Risk Assessment: Construct a hierarchical Bayesian network containing a basic attribute layer, a risk factor layer, and a complication risk layer. Initialize network parameters based on prior clinical data, input individual patient basic attributes, dynamically update the weight coefficients of each risk feature through probabilistic inference, integrate slow degradation features and network topology features, and output a comprehensive risk score and the risk level of each complication item. S5. Personalized intervention and closed-loop optimization based on causal matching: Retrieve the historical intervention case database, filter similar case sets through propensity score matching, calculate the expected risk improvement effect of different intervention programs using counterfactual causal inference, and output the optimal personalized intervention combination; continuously collect post-intervention data to calculate the intervention effectiveness, dynamically adjust the intervention program, and update the risk assessment model and intervention knowledge base through cumulative feedback data increment.

2. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S1, the specific rules for attributing abnormal fluctuation events are as follows: blood glucose data is detected using a 3-hour sliding window and heart rate data is detected using a 1-hour sliding window. The drift parameter is set to 0.5 times the data standard deviation, and the mutation judgment threshold is set to 3 times the data standard deviation. If the period of mutation corresponds to events such as high-sugar diet or strenuous exercise, and the magnitude of the mutation is within the physiological impact range of the corresponding event, it is marked as a non-pathological abnormality. The data for this period is multiplied by a weighting factor of 0.2 to 0.4 and is not included in the calculation of core risk features.

3. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S2, the quantitative screening criteria for steady-state periods are: no acute disease markers or major life event markers for more than 7 consecutive days, and the coefficient of variation of the daily average values ​​of core physiological indicators is ≤8%. Kernel density estimation uses a Gaussian kernel function, and bandwidth is calculated using the Silverman criterion; The baseline interval for discrete biochemical indicators is constructed by combining test values ​​from multiple steady-state periods with the bootstrap method. When updating the baseline interval incrementally, the weight of historical baseline data is 0.6, and the weight of newly added steady-state data is 0.

4.

4. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S3, the multi-resolution data alignment method is as follows: minute-level continuous physiological data are downsampled to daily-level statistics, and monthly-level discrete biochemical indicators are mapped to daily-level sequences through linear interpolation, thus unifying them into daily-level data. The coupled network is dynamically constructed with a 7-day sliding window and a 1-day step size. Mutual information is calculated using the equal-frequency binning method, with the number of bins set to 8.

5. The digital management method for early warning of diabetes complications according to claim 4, characterized in that: In step S3, the calculation rules for the three types of network topology features are as follows: the abnormal edge ratio is the proportion of the number of edges whose edge weights deviate from the individual baseline interval to the total number of edges; The global coupling efficiency is the average shortest path efficiency of the weighted network, and the edge weights are negatively correlated with the path length. The rate of change of critical node betweenness is the magnitude of the change in the weighted betweenness of the current window node relative to the baseline betweenness.

6. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S3, the slow degradation characteristics include three categories: the slope of the baseline offset trend, the coefficient of variation of the fluctuation amplitude, and the cumulative amount of out-of-bounds movement. The slope of the benchmark offset trend is the linear regression slope of the indicator's center value relative to the benchmark midline within the time window, and the cumulative out-of-bounds amount is the cumulative area of ​​the indicator's time-amplitude deviation from its own benchmark range.

7. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S4, the basic attribute layer of the hierarchical Bayesian network includes four nodes: disease course stratification, age stratification, number of comorbidities, and medication adherence level, which serve as prior variables to adjust the conditional probability of the risk factor layer. The risk factor layer contains 10-14 nodes, including those for slow degradation and network topology. The complication risk layer includes three output nodes: kidney disease risk, cardiovascular disease risk, and retinopathy risk. The model parameters are initialized using maximum likelihood estimation, and the probabilistic inference uses the joint tree exact inference algorithm.

8. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S5, the propensity score matching is performed based on 12 baseline features, using a 1:3 nearest neighbor matching method, and the matching caliper is set to 0.2 times the standard deviation of the propensity score. Counterfactual causal inference uses the inverse probability weighting method to estimate the average treatment effect of different intervention programs, with the outcome indicator being the improvement rate of the overall risk score 3 months after intervention.

9. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S5, the intervention plan includes four dimensions: dietary guidance, exercise prescription, medication management, and follow-up monitoring. Each dimension is divided into three intensity levels: basic, enhanced, and upgraded. The formula for calculating the effectiveness of intervention is: Effectiveness of intervention = 0.4 × Improvement rate of slow-change characteristics + 0.3 × Network topology recovery rate + 0.3 × Biochemical indicator compliance rate.

10. The digital management method for early warning of diabetes complications according to claim 1, characterized in that: In step S5, the incremental update cycle for the risk model and intervention knowledge base is 90-180 days. The Bayesian network parameter update uses weighted incremental learning, with the weight of new data set to 0.6 and the weight of historical data set to 0.

4. The intervention knowledge base is updated with new case data and causal effect estimates for each treatment plan.

Images