Pregnancy diabetes risk assessment and intervention decision method with potential metabolic state modeling

Abstract

The application discloses a kind of potential metabolic state modeling gestational diabetes risk assessment and intervention decision method, it is related to medical information technology field, including: first, obtain the minimum metabolic modeling data structure containing individual static information, follow-up observation and intervention data;Second, initialize the latent metabolic state vector formed by insulin sensitivity, beta cell compensatory capacity, fat burden and other latent variables;Then, an observation mapping and state updating mechanism is established, and the state vector is dynamically corrected and updated using follow-up observation and intervention data;Subsequently, based on the updated vector, a metabolic system stability index is calculated to assess patient risk stratification;Finally, by counterfactual reasoning, the state evolution under different hypothetical intervention scenarios is simulated, and the intervention priority and recommendations are output.The present application realizes early identification and precise intervention of risk through microscopic modeling and deduction of latent metabolic state, significantly improving the effectiveness of perinatal and postpartum health management.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology, specifically relating to a method for risk assessment and intervention decision-making of gestational diabetes based on potential metabolic state modeling. Background Technology

[0002] With the continuous advancement of precision medicine and chronic disease management technologies, the digital management of gestational diabetes has become an important component of perinatal healthcare. Traditional clinical practice primarily assesses a patient's metabolic status by regularly monitoring blood glucose levels, using this as the core basis for developing clinical intervention strategies. Especially in the mid-to-late stages of pregnancy, drastic fluctuations in the maternal metabolic environment directly impact fetal development and delivery safety, making full-cycle risk monitoring and dynamic management of gestational diabetes a crucial link in ensuring maternal and infant health.

[0003] Data-driven risk assessment and intervention decision-making methods are currently a hot research area in metabolic diseases, aiming to provide physicians with personalized treatment recommendations using multi-dimensional clinical data. These methods typically integrate individual vital signs, biochemical indicators, and behavioral data, constructing mathematical models to identify high-risk groups and optimize resource allocation. However, when facing the highly complex metabolic system of pregnancy, accurately characterizing individual physiological changes and simulating the long-term effects of interventions remains a core requirement for improving the precision of clinical management.

[0004] Current technologies primarily rely on macroscopic indicators such as fasting or postprandial blood glucose for diagnosis, making it difficult to reveal deeper metabolic system states such as insulin sensitivity and cellular compensatory capacity, resulting in a lag in characterizing potential pathological mechanisms. Furthermore, traditional assessment methods focus on achieving target blood glucose levels during pregnancy, lacking the ability to identify long-term postpartum complications and failing to effectively warn of metastable individuals whose blood glucose levels are temporarily normal but whose metabolic systems remain in a high-risk state. In addition, existing artificial intelligence models mostly employ end-to-end predictive logic, failing to explicitly model the state evolution process of the metabolic system. This leads to intervention plans relying heavily on static experience rather than dynamic quantitative deduction, making it difficult to achieve risk assessment and prioritization of complex intervention pathways.

[0005] Therefore, a risk assessment and intervention decision-making scheme for gestational diabetes based on potential metabolic state modeling is desired. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for risk assessment and intervention decision-making of gestational diabetes based on potential metabolic state modeling, which addresses the shortcomings of the prior art.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for risk assessment and intervention decision-making in gestational diabetes mellitus based on potential metabolic state modeling includes the following steps: Step 1: Obtain and construct the minimum metabolic modeling data structure: Obtain data on patients with gestational diabetes through a computer system. The data includes individual static information data, follow-up observation data, and intervention data. The individual static information data includes age, pre-pregnancy body mass index, previous pregnancy history, and family history of diabetes. The follow-up observation data includes blood glucose indicators, weight, waist circumference, blood pressure, and behavioral-related data at different time points. The intervention data includes dietary intervention, exercise intervention, sleep management, and compliance information. After standardization, the data forms a structured feature vector with temporal characteristics. Step 2, Initialize the latent metabolic state vector: Based on the individual static information data, the latent metabolic state vector of each patient is initialized by computer. The latent metabolic state vector includes insulin sensitivity state variables, beta cell compensation capacity state variables, fat burden state variables, and behavioral stability state variables. The above state variables are latent variables that cannot be directly measured. The nonlinear mapping relationship in the static features is processed by a deep learning network to estimate the initial probability distribution of the latent variables. Step 3, establish observation mapping and state update mechanism: establish the mapping relationship between follow-up observation data and potential metabolic state through computer, in which blood glucose index is used in combination with glucose kinetic equation to correct insulin sensitivity and beta cell compensation capacity, weight or waist circumference data is used to correct fat burden, and behavioral data is used to correct behavioral stability. Between adjacent follow-up time points, the distribution parameters of potential metabolic state vector are updated in real time through time-based discrete state transition model based on the marginal impact of intervention data on state variables. The mapping relationship includes the blood glucose observation equation: ; in, Let be the blood glucose observation value at time t; For basal glucose output from the liver; For insulin sensitivity state variables; For reference insulin sensitivity threshold; The nonlinear coefficient representing the effect of sensitivity on blood glucose; For the state variable of β-cell compensatory capacity; The contribution coefficient of compensatory capacity to blood glucose; To observe noise; The behavioral stability state variable is modified based on behavioral entropy, which is defined as: ; in, The normalized fluctuation frequency of the k-th type of behavior, including eating, exercising, and sleeping, within the time window; Between adjacent follow-up time points, the potential metabolic state vector is updated using a state transition model based on intervention data and observed changes. This state update mechanism employs a time-based discrete state transition model, and the state transition process follows the formula below: ; in, In order to be in The potential metabolic state vector at time t. In order to be in Intervention data vector applied at any given time, This is the state transition function, and its parameters are... It is obtained by pre-training a large amount of historical metabolic trajectory data through long short-term memory networks or gated recurrent units. For system random process noise that conforms to a normal distribution; Step 4, Metabolic state stratification and risk assessment: Based on the updated potential metabolic state vector, a potential energy function representing the energy topography of the metabolic system is constructed. The local optimal position of the potential metabolic state vector in the state space is evaluated. The Mahalanobis distance of the state point from the steady-state center is calculated as an index of metabolic system stability. Based on the Mahalanobis distance, patients are divided into metabolic homeostasis, metabolic metastasis, and high-risk persistent state, and the risk assessment results are output. The potential energy function is defined as: ; in, Let be the potential energy value under state s; This is the adjustment coefficient; It is the local Jacobian matrix in the state space, used to characterize the local curvature of the state points; Reflects the attractor strength of the system near that point; The stability index is defined as the Mahalanobis distance of the state point from the steady-state center, and its calculation formula is as follows: ; in, This represents the potential metabolic state vector at the current moment. This represents the mean metabolic state vector of a normal population at this gestational week. It is the inverse of the covariance matrix; Step 5, Evaluation of intervention decision based on counterfactual inference: Under the current potential metabolic state, at least one hypothetical intervention plan is constructed by computer. The hypothetical intervention plan is then subjected to counterfactual state inference using a structural causal model. The evolution path of the potential metabolic state under different intervention variables in the future prediction time period is simulated. By comparing the metabolic stability gains generated by the evolution path, the intervention priority or intervention recommendation is determined. In the counterfactual deduction, the posterior distribution of the exogenous noise variable is calculated as follows: ; in, State transition function process noise variance Assuming noise priors, we assume a Gaussian distribution. Step 6, Output Results: Output the risk assessment results and intervention decision results to the display terminal, and visualize the pathological mechanism focus by displaying a multi-dimensional potential metabolic state radar map.

[0008] Preferably, the process of obtaining and constructing the minimal metabolic modeling data structure in step 1 includes: By calling the electronic medical record system through a standardized interface, the baseline characteristics of the subject in early pregnancy or before pregnancy can be extracted. Convert age data into numerical features in a preset time unit; The pre-pregnancy body mass index was calculated by dividing the weight by the square of the height. Extract the total number of pregnancies and specific signs of fetal delivery from past pregnancy history, and then perform Boolean value processing on the records; The number of people with diabetes among relatives with a predetermined relationship is obtained, and weighted coefficients are assigned according to the closeness of the kinship, which are then quantified into a feature vector of family history of diabetes. The follow-up observation data were aligned to a time series, and the physiological indicators collected by the handheld blood glucose meter and wearable device were synchronized in real time via a wireless transmission protocol. Establish measurement error control logic to perform data smoothing when data fluctuations exceed the preset error range.

[0009] Preferably, the process of obtaining intervention data in step 1 includes: The dietary intervention records were quantified to analyze the types and weights of food recorded by the subjects and convert them into predetermined daily intake ratios of carbohydrates, protein, and fat. Record the predetermined duration of daily exercise at a specific intensity during exercise intervention, and calibrate the exercise intensity by combining heart rate data obtained from wearable devices; By monitoring and obtaining the total daily sleep duration and the planned percentage of deep sleep through sensors, sleep management data can be constructed. A comprehensive evaluation is conducted based on the interaction activity, record completeness, and suggestion execution rate of the mobile application to obtain compliance information, which is then mapped to a preset score range.

[0010] Preferably, the process of initializing the potential metabolic state vector in step 2 includes: The individual static information data is processed using a multilayer perceptron network, with at least three hidden layers, and the initial mean and variance of the potential metabolic state vector are output. The initial values ​​of the insulin sensitivity state variable were determined based on the nonlinear mapping relationship between pre-pregnancy body mass index and age. The initial values ​​of the state variable of beta cell compensatory capacity were determined based on the prior probabilities of previous pregnancy history and family genetic background. All state variables are assumed to conform to a multivariate normal distribution, and latent variable inference is performed using a variational inference mechanism to ensure the statistical independence of each state variable and eliminate the multicollinearity effect caused by redundant features in static information.

[0011] Preferably, the process of establishing the observation mapping relationship in step 3 includes: A probabilistic model based on the observation likelihood function is used to map unobservable latent variables to observable clinical indicators. The evolutionary relationship between blood glucose levels and latent variables was described using glucose kinetic equations, simulating the insulin-mediated glucose uptake process and the pulsatile secretion characteristics of beta cells. The impact of physiological noise on observation accuracy is eliminated by introducing a nonlinear mapping function; The state variables of fat burden are used to reflect the absolute amount of adipose tissue and the inhibitory effect of inflammatory factors produced by adipose tissue on overall metabolism. The fluctuations in exercise duration, sleep quality, and dietary intake in the follow-up observation data are defined as behavioral entropy, which reflects the impact of behavioral regularity on the metabolic system.

[0012] Preferably, the state update mechanism and the logic for correcting the behavioral stability state using behavioral data in step 3 include: A time-based discrete state transition model is adopted to determine the potential metabolic state at the next moment based on the potential metabolic state vector at the current moment, the intervention data vector applied at the current moment, and the state transition function parameters obtained through pre-training. System stochastic process noise that conforms to a normal distribution is introduced for calibration. The state transition function parameters are obtained by training historical metabolic trajectory data through a long short-term memory network or a gated recurrent unit. Set the synchronization period for the transition probability matrix in the state transition model, and periodically perform global synchronization to capture the metabolic environment fluctuations in the pregnant mother caused by changes in placental hormone levels. When the behavioral entropy increases and the duration exceeds a predetermined time threshold, the value of the behavioral stability state variable is reduced, and the change in behavioral stability state is fed back to other variables in the potential metabolic state vector through the system coupling coefficient.

[0013] Preferably, the process of calculating the metabolic system stability index in step 4 includes: Construct a potential energy function that characterizes the energy distribution of a metabolic system in the state space; Obtain the potential metabolic state vector at the current moment, and obtain the mean metabolic state vector of the normal population at the corresponding gestational week; Calculate the difference between the potential metabolic state vector and the mean metabolic state vector of the normal population; The difference is multiplied by the inverse of the covariance matrix, and a quadratic operation is performed using the transpose of the difference to obtain the Mahalanobis distance, which reflects the degree of deviation of the system from the steady-state center. The metabolic stratum to which the patient belongs is determined by comparing the Mahalanobis distance with the first preset threshold and the second preset threshold.

[0014] Preferably, the process of outputting the risk assessment result in step 4 includes: Predict the probability of severe metabolic disorders and related complications in late pregnancy; Calculate the long-term risk score of developing a specific type of diabetes within a predetermined time after delivery; ; in, The Mahalanobis distance threshold for the normal population at the corresponding time point; , As a risk attenuation factor; This is the end time of the follow-up visit; The long-term risk score is calculated by weighting the integral of the duration of metabolic metastability during pregnancy; When the long-term risk score reaches the preset risk warning threshold, a high-risk warning is triggered to the medical staff through an encrypted communication protocol.

[0015] Preferably, the intervention decision evaluation process based on counterfactual reasoning in step 5 includes: A directed acyclic graph (DAG) containing intervention variables, potential metabolic state mediator variables, and risk stratified outcome variables was constructed, and the path weights in the DAG were calibrated using observation sequences. Given the current timestamp, perform abductive operations to calculate the posterior distribution of the exogenous noise variables behind the current observations; Perform intervention operations to forcibly modify the values ​​of intervention variable nodes in the model in order to construct a hypothetical intervention plan; Perform prediction operations to deduce the evolution path of potential metabolic states within a preset prediction time period while keeping exogenous noise constant; Multiple evolutionary pathways were generated using the Monte Carlo sampling method, and all evolutionary pathways were aggregated with probability weighting to calculate the metabolic stability gain and confidence interval for each hypothetical intervention scheme.

[0016] Preferably, the logic for intervention decision-making and outcome output in steps 5 and 6 includes: The benefit-cost ratio of all simulated hypothetical intervention plans was calculated using a multi-objective optimization algorithm, where the optimization objectives included maximizing the improvement of metabolic stability, maximizing the probability of compliance, and minimizing the degree of disturbance to quality of life. Benefit is defined as the extent of the reduction in metabolic system risk, and cost is defined as the expected intensity of the intervention on the patient's quality of life. ; in, The decrease in Mahalanobis distance resulting from intervention i; The probability of compliance predicted by the model; The intensity of the intervention includes the duration of exercise and the degree of dietary restriction; Rate the loss in quality of life; These are the weighting coefficients; Through Pareto optimal search, a predetermined number of intervention strategies that meet the balance requirements are recommended first. The expected effects and predicted uncertainty range of each strategy are displayed on the display terminal in the form of dynamic evolution curves; By utilizing the discrepancy between the actual postpartum metabolic recovery trajectory and the predicted trajectory during pregnancy, the parameters in the state transition model are automatically optimized through reinforcement learning algorithms, thereby achieving self-evolution of the model's accuracy. A radar map displaying the patient's potential metabolic status in real time is provided. The dimensions of the radar map correspond to insulin sensitivity, beta cell compensatory capacity, fat burden, and behavioral stability, respectively.

[0017] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 0. This invention breaks through the limitations of traditional methods that rely solely on blood glucose test results: by abstracting the metabolic system into a state-space system, it achieves quantitative modeling of deep physiological functions such as insulin sensitivity and beta cell compensatory ability, and can identify metastable individuals whose blood glucose levels are within the normal range but whose internal regulatory mechanisms have undergone structural shifts. The advance time for risk identification is significantly increased compared to traditional methods. 0. This invention addresses the pain point of difficulty in assessing long-term postpartum risks: by explicitly modeling the state evolution process of the metabolic system, the system can capture subtle pathological features based on the potential metabolic trajectory during pregnancy, significantly improving the accuracy of predicting postpartum metabolic disease risks and providing a scientific basis for the early prevention of postpartum chronic disease transformation. 0. This invention realizes the transformation of intervention programs from experience-driven to evidence-driven: the counterfactual inference-based technical solution can simulate the long-term benefits of multiple hypothetical programs before actual intervention, avoiding the clinical risks and time costs caused by blindly trying different intervention measures, greatly improving the targeting of intervention programs, and significantly improving the achievement rate of patients' metabolic indicators. 0. This invention has good engineering feasibility and scalability: The system adopts a modular design, supports seamless integration with existing electronic medical records and wearable devices, the automated processing flow significantly reduces the workload of clinical medical staff, and the model framework is universal and can be quickly extended to other metabolic disease management scenarios with latent variable characteristics. Attached Figure Description

[0018] Figure 1 A flowchart of the gestational diabetes risk assessment and intervention decision-making method for modeling potential metabolic states according to the present invention; Figure 2 A schematic diagram illustrating the principle of initializing the potential metabolic state vector in the gestational diabetes risk assessment and intervention decision-making method for modeling potential metabolic states in this invention. Figure 3 The logical flowchart of the gestational diabetes risk assessment and intervention decision-making method based on observation mapping and state transition model in the modeling of potential metabolic states of the present invention is shown in the following diagram. Figure 4 A flowchart illustrating the logical process of risk stratification and assessment based on metabolic system stability indicators in the gestational diabetes risk assessment and intervention decision-making method for modeling potential metabolic states in this invention. Figure 5 The logical flowchart of the intervention program evaluation and path evolution based on counterfactual deduction in the gestational diabetes risk assessment and intervention decision-making method for modeling potential metabolic states of the present invention is shown in the figure. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0020] like Figures 1 to 5As shown, a method for risk assessment and intervention decision-making for gestational diabetes based on potential metabolic state modeling is presented, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain and construct the minimum metabolic modeling data structure: Obtain data on patients with gestational diabetes through a computer system. The data includes individual static information data, follow-up observation data, and intervention data. The individual static information data includes age, pre-pregnancy body mass index, previous pregnancy history, and family history of diabetes. The follow-up observation data includes blood glucose indicators, weight, waist circumference, blood pressure, and behavioral-related data at different time points. The intervention data includes dietary intervention, exercise intervention, sleep management, and compliance information. After standardization, the data forms a structured feature vector with temporal characteristics. Step 2, Initialize the latent metabolic state vector: Based on the individual static information data, the latent metabolic state vector of each patient is initialized by computer. The latent metabolic state vector includes insulin sensitivity state variables, beta cell compensation capacity state variables, fat burden state variables, and behavioral stability state variables. The above state variables are latent variables that cannot be directly measured. The nonlinear mapping relationship in the static features is processed by a deep learning network to estimate the initial probability distribution of the latent variables. Step 3, establish observation mapping and state update mechanism: establish the mapping relationship between follow-up observation data and potential metabolic state through computer, in which blood glucose index is used in combination with glucose kinetic equation to correct insulin sensitivity and beta cell compensation capacity, weight or waist circumference data is used to correct fat burden, and behavioral data is used to correct behavioral stability. Between adjacent follow-up time points, the distribution parameters of potential metabolic state vector are updated in real time through time-based discrete state transition model based on the marginal impact of intervention data on state variables. The mapping relationship includes the blood glucose observation equation: ; in, Let be the blood glucose observation value at time t; For basal glucose output from the liver; For insulin sensitivity state variables; For reference insulin sensitivity threshold; The nonlinear coefficient representing the effect of sensitivity on blood glucose; For the state variable of β-cell compensatory capacity; The contribution coefficient of compensatory capacity to blood glucose; To observe noise; The behavioral stability state variable is modified based on behavioral entropy, which is defined as: ; in, The normalized fluctuation frequency of the k-th type of behavior, including eating, exercising, and sleeping, within the time window; Between adjacent follow-up time points, the potential metabolic state vector is updated using a state transition model based on intervention data and observed changes. This state update mechanism employs a time-based discrete state transition model, and the state transition process follows the formula below: ; in, In order to be in The potential metabolic state vector at time t. In order to be in Intervention data vector applied at any given time, This is the state transition function, and its parameters are... It is obtained by pre-training a large amount of historical metabolic trajectory data through long short-term memory networks or gated recurrent units. For system random process noise that conforms to a normal distribution; Step 4, Metabolic state stratification and risk assessment: Based on the updated potential metabolic state vector, a potential energy function representing the energy topography of the metabolic system is constructed. The local optimal position of the potential metabolic state vector in the state space is evaluated. The Mahalanobis distance of the state point from the steady-state center is calculated as an index of metabolic system stability. Based on the Mahalanobis distance, patients are divided into metabolic homeostasis, metabolic metastasis, and high-risk persistent state, and the risk assessment results are output. The potential energy function is defined as: ; in, Let be the potential energy value under state s; This is the adjustment coefficient; It is the local Jacobian matrix in the state space, used to characterize the local curvature of the state points; Reflects the attractor strength of the system near that point; The stability index is defined as the Mahalanobis distance of the state point from the steady-state center, and its calculation formula is as follows: in, This represents the potential metabolic state vector at the current moment. This represents the mean metabolic state vector of a normal population at this gestational week. It is the inverse of the covariance matrix; Step 5, Evaluation of intervention decision based on counterfactual inference: Under the current potential metabolic state, at least one hypothetical intervention plan is constructed by computer. The hypothetical intervention plan is then subjected to counterfactual state inference using a structural causal model. The evolution path of the potential metabolic state under different intervention variables in the future prediction time period is simulated. By comparing the metabolic stability gains generated by the evolution path, the intervention priority or intervention recommendation is determined. In the counterfactual deduction, the posterior distribution of the exogenous noise variable is calculated as follows: in, State transition function process noise variance Assuming noise priors, we assume a Gaussian distribution. Step 6, Output Results: Output the risk assessment results and intervention decision results to the display terminal, and visualize the pathological mechanism focus by displaying a multi-dimensional potential metabolic state radar map.

[0021] Example 1: To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0022] This embodiment provides a method for risk assessment and intervention decision-making in gestational diabetes mellitus based on potential metabolic state modeling. This method is executed by a computer system and is used to perform in-depth quantitative evaluation of metabolic homeostasis in pregnant women and provide individualized intervention pathway simulation. The method specifically includes the following steps:

[0023] Step S1: Acquire and construct a minimal metabolic modeling data structure: Acquire data on patients with gestational diabetes mellitus through a computer system. The data acquisition process is integrated with the electronic medical record system of medical institutions through a standardized application programming interface, supporting distributed data collection. The data includes at least individual static information data, follow-up observation data, and intervention data.

[0024] When acquiring individual static information data, structured fields from the electronic medical record system are accessed through a distributed data acquisition interface. Baseline characteristics of the subjects before or during early pregnancy are extracted. Age data is converted into numerical features in days or weeks as preset time units for subsequent alignment in time series models. Pre-pregnancy body mass index (BMI) is calculated by obtaining the subject's pre-pregnancy weight and height, and then dividing weight by the square of height. The extraction of past pregnancy history data includes quantifying the total number of pregnancies, and Booleanizing delivery records of fetuses with specific physical characteristics such as preterm birth and macrosomia. Family history of diabetes is quantified as the number of relatives with the disease within a predetermined relationship, and different weighting coefficients are assigned according to the closeness of the kinship to construct a quantitative feature vector of individual genetic background.

[0025] When acquiring follow-up observation data, the time span covers the entire period from the time of diagnosis of gestational diabetes to a predetermined time point, such as 42 days or 180 days postpartum. Blood glucose level collection includes at least fasting blood glucose, 1-hour postprandial blood glucose, and 2-hour postprandial blood glucose. In the mid-to-late stages of pregnancy, blood glucose collection is set at a preset frequency of at least four times per day, with data from the subject's handheld blood glucose meter synchronized in real-time to the computer system via a wireless transmission protocol. Simultaneously, the subject's weight change data is synchronized in real-time via a wearable device worn by the subject, utilizing gravity and pressure sensors. To ensure data quality, the system incorporates measurement error control logic; when data fluctuations exceed a preset error range, a retest reminder or data smoothing is triggered.

[0026] When acquiring intervention data, dietary intervention records were quantified. By analyzing the types and quantities of food recorded by the subjects, these were converted into predetermined daily intake ratios of carbohydrates, protein, and fat, as well as total energy values. Exercise intervention records included predetermined durations of daily exercise at specific intensities, with exercise intensity automatically calibrated using heart rate data acquired through wearable devices. Sleep management data included total daily sleep duration and the proportion of deep sleep, monitored by sensors. Compliance information was comprehensively assessed through interactive activity on the mobile application, record completeness, and recommendation execution rate, with compliance scores set within a preset range of 0 to 100.

[0027] Step S2: Initialize the latent metabolic state vector: Based on the individual static information data, initialize the latent metabolic state vector for each patient using a computer system. The latent metabolic state vector consists of a set of latent variables that cannot be directly measured, characterizing the internal structural features of the patient's metabolic system. The latent metabolic state vector includes at least the following state variables: insulin sensitivity, β-cell compensation capacity, fat burden, and behavioral stability.

[0028] The initialization process employs a multilayer perceptron network to process individual static information data. Static features such as age, pre-pregnancy body mass index, and medical history are used as input layers. A deep neural network with at least three hidden layers outputs the initial mean and variance of the potential metabolic state vector. Specifically, the initial value of the insulin sensitivity state variable is based on a non-linear mapping of pre-pregnancy body mass index and age, reflecting the individual's tissue response potential to insulin at baseline. The initial value of the β-cell compensatory capacity state variable is probabilistically initialized based on past pregnancy history and family genetic background.

[0029] All state variables are assumed to follow a multivariate normal distribution during the initialization phase. Through a latent variable inference algorithm and variational inference mechanism, the statistical independence of each state variable is ensured, eliminating the multicollinearity effects caused by redundant features in static information. The initialized state vector provides the initial probability distribution basis for subsequent state transition calculations.

[0030] Step S3: Establishing an observation mapping and state update mechanism: A computer-based mapping relationship is established between follow-up observation data and potential metabolic states. A probabilistic model based on the observation likelihood function is used to map unobservable latent variables to observable clinical indicators. Specifically, blood glucose levels are used to correct insulin sensitivity and β-cell compensatory capacity. The relationship between blood glucose levels and the aforementioned latent variables follows the glucose kinetic equation. By introducing a nonlinear mapping function, the insulin-mediated glucose uptake process and the pulse-like secretion characteristics of β-cells are simulated.

[0031] Fat burden status is corrected using weight or waist circumference data. Fat burden status variables reflect not only the absolute amount of adipose tissue but also the inhibitory effect of inflammatory factors produced by adipose tissue on overall metabolism. Behavioral stability status is corrected using behavioral data, defining the fluctuations in exercise duration, sleep quality, and dietary intake as behavioral entropy.

[0032] Between adjacent follow-up time points, the potential metabolic state vector is updated using a state transition model based on intervention data and observed changes. The state update mechanism employs a time-based discrete state transition model. The state transition process follows the formula below:

[0033] in, Representative at The potential metabolic state vector at time t. Representative at Intervention data vector applied at any given time, This is the state transition function, and its parameters are... It is obtained by pre-training a large amount of historical metabolic trajectory data through long short-term memory networks or gated recurrent units. This refers to the noise of a system random process that conforms to a normal distribution.

[0034] The transition probability matrix in the state transition model is globally synchronized every 24 or 48 hours according to a preset period to ensure that it can capture the metabolic environment fluctuations caused by drastic changes in placental hormone levels during pregnancy. The logic for correcting behavioral stability using behavioral data is as follows: when behavioral entropy increases and its duration exceeds a preset time threshold of 3 days, the value of the behavioral stability state variable is automatically reduced, and this is synchronously fed back to other variables in the metabolic state vector. For example, a decrease in behavioral stability will lead to a decrease in the predicted value of insulin sensitivity through the system coupling coefficient.

[0035] Step S4: Metabolic State Stratification and Risk Assessment: Based on the updated potential metabolic state vector, a metabolic system stability index is calculated, and patients are stratified. The calculation is performed by constructing a potential energy function for the metabolic system. This potential energy function characterizes the energy topography of the metabolic system in the state space, assessing the local optimum of the current potential metabolic state vector in the state space.

[0036] The stability index is defined as the Mahalanobis distance of the state point from the steady-state center, and its calculation formula is as follows: in, This represents the potential metabolic state vector at the current moment. This represents the mean metabolic state vector of a normal population at this gestational week. It is the inverse of the covariance matrix.

[0037] Preferably, the process of outputting the risk assessment result in step 4 includes: Predict the probability of severe metabolic disorders and related complications in late pregnancy; Calculate the long-term risk score of developing a specific type of diabetes within a predetermined time after delivery; ; in, The Mahalanobis distance threshold for the normal population at the corresponding time point; , As a risk attenuation factor; This is the end time of the follow-up visit; The long-term risk score is calculated by weighting the integral of the duration of metabolic metastability during pregnancy; When the long-term risk score reaches the preset risk warning threshold, a high-risk warning is triggered to the medical staff through an encrypted communication protocol.

[0038] When the Mahalanobis distance is less than the first preset threshold, it is determined to be a metabolic homeostasis; when the Mahalanobis distance is between the first preset threshold and the second preset threshold, it is determined to be a metabolic metastable state; when the Mahalanobis distance is greater than the second preset threshold, it is determined to be a high-risk persistent state.

[0039] The system outputs corresponding risk assessment results. These results include a predicted probability of severe metabolic disorders in late pregnancy (such as the risk of macrosomia, preeclampsia complications, etc.) and a long-term risk score for developing type 2 diabetes within 12 weeks to 1 year postpartum. The long-term risk score is calculated based on a weighted average of the duration of metabolic metastability during pregnancy. When the score reaches a preset risk warning threshold, the system automatically triggers a high-risk warning to healthcare professionals via an encrypted communication protocol.

[0040] Step S5: Counterfactual Intervention Decision Evaluation: Under the current potential metabolic state, at least one hypothetical intervention plan is constructed using a computer. The hypothetical intervention plan may include adjustments to dietary proportions, increases in exercise intensity, or optimization of sleep patterns.

[0041] At a given current timestamp, freeze all potential metabolic state components except for the intervention variable. Using a structural causal model to change the variable values ​​in the intervention scheme, simulate the evolutionary paths of potential metabolic states over preset prediction time periods of 4 or 8 weeks under different intervention combinations. Calculate the metabolic stability gain brought by each evolutionary path, i.e., the expected magnitude of Mahalanobis distance reduction.

[0042] Benefit-cost ratios were calculated for all simulated hypothetical interventions. Benefits were defined as the extent of the reduction in metabolic system risk, and costs were defined as the expected intensity of the intervention on the patient's quality of life (e.g., the psychological cost of very low carbohydrate intake, the physical cost of high-intensity exercise).

[0043] ; in, The decrease in Mahalanobis distance resulting from intervention i; The probability of compliance predicted by the model; The intensity of the intervention includes the duration of exercise and the degree of dietary restriction; Rate the loss in quality of life; These are the weighting coefficients; Based on the calculation results, the system prioritizes and recommends 3 to 5 intervention strategies that offer the greatest benefit and the highest probability of patient compliance. On the display terminal, the system presents the expected effect of each strategy in the form of a dynamic evolution curve.

[0044] Step S6: Output Results: Output the risk assessment results and intervention decision results to a display terminal or medical information system. The display terminal supports interactive visualization, showing a radar chart of the patient's potential metabolic status in real time. The dimensions of the radar chart correspond to insulin sensitivity, β-cell compensation capacity, fat burden, and behavioral stability, respectively. The area and shape of the radar chart help clinicians intuitively determine the focus of the pathological mechanism.

[0045] This embodiment also provides a system for performing the above method, including a data acquisition module, a potential metabolic state modeling module, a state transition calculation module, a risk assessment module, a counterfactual deduction and intervention decision-making module, and a result output module.

[0046] The data acquisition module is equipped with a distributed data acquisition interface, supporting the synchronization of structured data from multiple medical institutions, and features data anonymization and encrypted storage. The latent metabolic state modeling module is deployed on a high-performance computing server, supporting parallel inference of latent variables based on a graphics processor. The state transition calculation module adopts a low-latency streaming processing framework, ensuring that once patient follow-up data is updated, the system can complete the iterative update of the state vector within seconds.

[0047] The system is deployed on a cloud server and connects to the hospital information system and patient handheld terminals via an SSL encrypted link. The database employs a hybrid storage architecture, with static patient data stored in a relational database and high-frequency follow-up data stored in a time-series database. The system's maximum concurrent processing capacity meets the requirement of over 10,000 concurrent requests, and the time taken for a single risk assessment is within a preset time threshold of 200 milliseconds.

[0048] The system supports continuous integration of postpartum follow-up data. By comparing the deviation between the actual postpartum metabolic recovery trajectory and the predicted trajectory during pregnancy, reinforcement learning algorithms are used to automatically optimize the parameters in the state transition model, achieving self-evolution of model accuracy. Through this self-evolution, the accuracy rate of metabolic risk identification within one year postpartum reaches a preset accuracy target of over 90%.

[0049] Example 2: Based on Example 1, this example further refines the initialization and counterfactual deduction logic of potential metabolic states for the special metabolic background of pregnant women of advanced maternal age (age greater than 35 years).

[0050] In the context of advanced maternal age pregnancy, the nonlinear mapping function with age weighting in the individual static information data obtained in step S1 is given higher sensitivity. A computer system performs in-depth analysis of the past pregnancy history of older pregnant women to extract surrogate indicators related to vascular endothelial function and renal function as supplementary dimensions to the potential metabolic state vector.

[0051] In step S2, when initializing the potential metabolic state vector, the initial weights of the multilayer perceptron network take into account ovarian reserve function and age-related decline in mitochondrial activity. The initial distribution of the insulin sensitivity state variable is set to have a wider variance to reflect the high individual variability in metabolic elasticity in the elderly.

[0052] In the state update mechanism of step S3, the state transition model introduces a gestational age-dependent parameter matrix. Because the impact of placental lactogen secretion on the metabolic system is more significant in older pregnant women after 28 weeks of gestation, the parameters in the state transition function... The system dynamically segments the data based on gestational age. Between adjacent follow-up points, the state transition model incorporates blood pressure data to assist in correcting the microcirculation status. If an abnormal increase in blood pressure fluctuations is observed, the system automatically lowers the update threshold for the fat burden state variable, reflecting the coupling effect between fat metabolism and vascular pressure.

[0053] In the risk assessment of step S4, a penalty term based on energy metabolism bottlenecks was introduced into the calculation of the metabolic system stability index. For older adults, the bottom region of their metabolic potential function is narrower, meaning that their system has a higher risk of collapse after deviating from steady state. The system's threshold for determining metabolic metastability was narrowed by 15% compared to Example 1, enabling earlier intervention and warning.

[0054] In the counterfactual reasoning process of step S5, a constraint on "intervention tolerance" was added when constructing the hypothetical intervention plan. Considering that older pregnant women generally have lower compliance with high-intensity exercise interventions, the structural causal model automatically sets a higher cost penalty for high-intensity exercise plans when simulating the evolutionary path. The recommended intervention focuses more on fine-tuning dietary structure and deep optimization of sleep quality.

[0055] In the output of step S6, the physician decision support interface adds a "metabolic age" comparison dimension for elderly patients. By comparing the patient's potential metabolic state vector with the mean of pregnant women of the same gestational age, the degree of premature aging of their metabolic system is quantified, providing a more intuitive pathophysiological background reference for clinical diagnosis and treatment.

[0056] When processing data from the elderly population, the system's reinforcement learning algorithm employs transfer learning techniques. It uses a large-sample pre-trained model from the entire population as a foundation, and fine-tunes the prediction accuracy for the specific elderly subgroup by adjusting the data accordingly. The distributed data acquisition interface, when acquiring data from these patients, simultaneously retrieves higher-frequency fetal heart rate monitoring and placental perfusion indicators, enabling coordinated risk monitoring of maternal metabolism and fetal development.

[0057] Example 3: Based on Example 1, this example strengthens the dynamic modeling process of fat burden state variables for gestational diabetes patients with obesity (pre-pregnancy BMI greater than 28).

[0058] In step S1, the data acquisition module synchronizes the subject's body composition analysis data, including body fat percentage, basal metabolic rate, and visceral fat level, using a wearable body fat scale. This high-level observational data is directly integrated into the follow-up observation data stream.

[0059] In step S2 initialization, the initial mean of the fat burden state variable is directly correlated with pre-pregnancy BMI and waist-to-hip ratio. Using a Bayesian inference algorithm, baseline levels of obesity-related inflammatory factors are incorporated as prior probabilities into the state vector. The initialization of the β-cell compensation capacity state variable incorporates the subject's prime ministerial secretion performance in the glucose tolerance test for a more precise characterization.

[0060] In step S3, the observed mapping relationship adds a nonlinear feedback loop of adipose tissue on blood glucose. The attenuation coefficient of adipokines is introduced into the glucose kinetic equation to simulate the interference of free fatty acids on the insulin signaling pathway. In high-risk individuals for obesity, the time step of the state transition model is shortened to 6 hours to more sensitively capture the alternating effects of postprandial fat metabolism and glucose metabolism.

[0061] In step S4, the calculation of metabolic system stability indicators pays particular attention to the Mahalanobis distance between the fat burden variable and the insulin sensitivity variable. If the deviation of the linkage between the two continues to increase, the system will directly upgrade the stratification result to a high-risk persistent state. In the long-term risk score, the calculation of the "metabolic memory effect" has been added, which assesses the extent of irreversible damage to pancreatic tissue after delivery caused by long-term high fat burden during pregnancy.

[0062] In the counterfactual deduction in step S5, the intervention decision evaluation module constructs several "weight control-oriented" hypothetical intervention plans. The system uses a structural causal model to simulate the evolution of body fat distribution and its effect on the recovery of insulin sensitivity in patients under strict restriction of fat intake and maintenance of moderate-intensity aerobic exercise. The intervention priority output logic will prioritize and recommend customized dietary plans that can effectively reduce fat burden without inducing the risk of ketosis for this population.

[0063] The system includes a dedicated image analysis module that uses convolutional neural networks to automatically identify food types and approximate portions from food photos uploaded by patients, thereby improving the accuracy of dietary intervention data acquisition. The distributed data acquisition interface supports low-level protocol integration with mainstream smart scale brands on the market, ensuring lossless and real-time synchronization of fat composition data.

[0064] Example 4: Based on Example 1, this example describes in detail the implementation details of the system in a multi-center collaborative mode, especially the application scenario in a regional medical and health management platform.

[0065] In the data construction phase of step S1, the system uses a distributed transaction processing mechanism to ensure the consistency of timestamps in data obtained from hospitals at different levels (such as community health service centers and tertiary hospitals). The data acquisition module includes a medical text parsing engine based on natural language processing, which can extract intervention suggestions and subjects' subjective compliance statements from unstructured outpatient medical records and transform them into structured intervention data.

[0066] In step S2, the initialization of the potential metabolic state vector incorporates geographical and seasonal factors. Big data on dietary habits from the subjects' residential areas is acquired via a computer system, serving as a population baseline for initializing the insulin sensitivity state variable. This population-prior-based initialization method significantly improves the model's inference stability in the early stages of data sparsity.

[0067] During the state update process in step S3, the system employs a federated learning framework. Each medical center trains the branch parameters of the state transition model on its local server without compromising patient privacy, and then encrypts and uploads the parameter increments to the cloud master server. The cloud server performs global aggregation of the model parameters, enabling the state update model to rapidly adapt to the metabolic characteristics of cross-regional and multi-ethnic populations.

[0068] In addition to outputting individual risk, the risk assessment results in step S4 also generate a regional metabolic homeostasis distribution map. For areas with a dense distribution of metabolically metastable populations, the system automatically sends an early warning to the management platform, indicating the necessity of public health intervention.

[0069] Step S5, the counterfactual deduction, adds the variable constraint of "community intervention resources." When simulating intervention plans, the system automatically searches whether the patient's community has corresponding sports facilities or food delivery support. If relevant resources are lacking, the system will increase the intervention cost accordingly when calculating the benefit-cost ratio, thereby ensuring that the output intervention recommendations are highly implementable in real-life scenarios.

[0070] The output interface of step S6 provides a social incentive module for patients. Based on the expected gains derived from counterfactual reasoning, the system sets phased metabolic homeostasis goals for patients. By displaying metabolic homeostasis recovery paths of patients in the same category, social feedback is used to enhance patient compliance.

[0071] The system's database is deployed in a highly available cluster, and the storage architecture is specifically optimized for time-series data. High-frequency follow-up data is stored in shards on a weekly basis, and compression algorithms are used to reduce the storage cost of long-term follow-up data. The encrypted link uses a two-way authentication mechanism to ensure data security throughout the entire link from wearable devices to the cloud, and then from the cloud to the doctor's terminal.

[0072] Example 5: This example further illustrates the specific application of the above-mentioned gestational diabetes risk assessment and intervention decision-making method in postpartum follow-up and early warning of chronic disease conversion.

[0073] In step S1, the timeline for data acquisition was extended to 1 to 5 years postpartum. Follow-up observation data included key variables such as glucose tolerance test data, lipid analysis data, and duration of breastfeeding at different postpartum time points. Intervention data incorporated postpartum weight management protocols and records of specific nutritional interventions during lactation.

[0074] The initialization of the latent metabolic state vector in step S2 is reset after delivery. Based on static indicators from the first week postpartum, the system initializes a new "metabolic repair state vector." This vector incorporates latent variables such as the rate of tissue damage repair and the speed of hormone level decline.

[0075] The state update mechanism in step S3 focuses on capturing the "metabolic regression path." The state transition model calculates the residual between postpartum observation data and the pregnancy trajectory to determine whether the patient's metabolic system is returning to homeostasis or has formed new pathological stagnation points in a metastable region. Using a reinforcement learning algorithm, the system continuously adjusts the weights of the positive impacts of different lactation intensities on fat burden and insulin sensitivity.

[0076] Step S4's risk assessment focuses on early warning of type 2 diabetes conversion. By calculating the decay slope of Mahalanobis distance on the postpartum time axis, the system can accurately identify individuals whose blood glucose levels are within target range, but whose metabolic energy topography still shows significant instability. The risk assessment results will provide a specific probability distribution of impaired glucose tolerance within the next 3 to 5 years.

[0077] The counterfactual analysis in step S5 is used to develop a postpartum prevention plan. The system simulates different pathways, such as "continuous breastfeeding + structured diet" and "non-breastfeeding + intensive exercise," to assess which approach offers greater long-term diabetes prevention benefits for the patient. The output intervention recommendations are directly integrated into the patient's chronic disease management plan.

[0078] The system integrates a patient lifecycle management module. When the system detects a continuous rise in postpartum risk scores, it automatically adds a "high risk of chronic disease conversion" tag to the medical information system and automatically schedules the next outpatient follow-up. The results output module supports the generation of multi-dimensional postpartum metabolic recovery reports, providing quantitative evidence for general practitioners' interventions.

[0079] Example 6: Based on Example 1, this example provides a detailed description of the structural causal model and its engineering implementation in the counterfactual deduction and intervention decision-making module.

[0080] In step S5, the structural causal model constructs a directed acyclic graph (DAG) to describe the causal relationships between intervention variables, mediating variables (potential metabolic state), and outcome variables (glucose levels, risk stratification). The computer system uses the patient observation sequence accumulated over the past three months to automatically calibrate the path weights in the DAG using a causal discovery algorithm.

[0081] When performing counterfactual reasoning, the system first performs a "causal" operation. Given the current observations, it calculates the posterior distribution of the underlying exogenous noise variable. Next, it performs an "intervention" operation by forcibly modifying the values ​​of the intervention variable nodes in the model. Finally, it performs a "prediction" operation, predicting the system's evolution after the intervention variable change, while keeping the exogenous noise constant.

[0082] To improve extrapolation efficiency, the system employs a Monte Carlo sampling method to generate thousands of possible evolutionary paths. Each path corresponds to a subtle fluctuation in the intervention parameter (e.g., a fluctuation of ±5% in carbohydrate intake). The system performs probability-weighted aggregation on all sampled paths to derive the expected benefit and confidence interval of the intervention plan.

[0083] When outputting intervention recommendations, the logical judgment is implemented through a multi-objective optimization algorithm. The optimization objective functions include: A) maximizing the improvement of metabolic stability; B) maximizing the probability of compliance; and C) minimizing the degree of disruption to quality of life. The system uses Pareto optimality search to find a predetermined number of intervention strategy combinations that achieve the best balance among the above objectives.

[0084] The dynamic evolution curve displayed on the terminal not only shows the mean path but also indicates the range of uncertainty in the system's predictions through shaded areas. When the prediction uncertainty is too high, the system will prompt the doctor that the current intervention recommendations are for reference only and suggest increasing the follow-up observation frequency to further calibrate the model. This transparent decision support mechanism significantly enhances clinicians' confidence in the system's output results.

[0085] Example 7: Based on Example 1, this example describes how the system handles the fault tolerance mechanism for missing data and abnormal behavioral data to ensure robustness in a real clinical environment.

[0086] If data is missing due to device offlineness or missed patient entries during the acquisition of follow-up observation data in step S1, the system initiates an interpolation mechanism based on latent state vectors. The data acquisition module uses a generative adversarial network to generate pseudo-observation data that is highly consistent with the patient's historical metabolic patterns to fill data gaps in short time periods (less than 48 hours) and ensure the continuity of state transition calculations.

[0087] In the state update step S3, the system incorporates an outlier detection algorithm. When the wearable device's synchronized behavioral data (such as steps or sleep duration) exhibits impulsive jumps exceeding physiological limits, the system automatically identifies it as a device error. At this point, the state update logic temporarily suspends the correction operation based on the outlier data and instead relies on the prior predictions of the state transition model for further steps.

[0088] The calculation logic for behavioral stability state variables incorporates a dynamic time window mechanism. The system considers not only behavioral fluctuations at the current time point but also compares them with the baseline fluctuation levels over the past 14 days. If a sudden increase in behavioral entropy is detected as being caused by changes in the external environment (such as hospitalization or travel), the system automatically adjusts the weighting coefficients to avoid erroneous metabolic risk warnings due to short-term environmental disturbances.

[0089] In the risk assessment module, the calculation of the stability index will simultaneously output a "credibility score." This score reflects the quality of the original data upon which the current assessment result depends. If the data missing rate exceeds a preset threshold, the credibility score will decrease, and the system will display a prompt "Insufficient data, please supplement records" when outputting the risk assessment result to prevent misjudgments due to incomplete information.

[0090] The system adopts a microservice architecture, with data imputation services, anomaly detection services, and core modeling services isolated from each other. This design ensures that even in extreme data environments, the core risk assessment logic can maintain minimal operation. When displaying risk warnings, the results output module automatically adjusts the alarm tone and visual intensity based on the credibility score, ensuring that medical staff can accurately perceive the effectiveness of the assessment results.

[0091] Example 8: This example illustrates the integration and implementation details of the method of the present invention in a hospital intelligent decision support system.

[0092] The system seamlessly integrates with the hospital's middleware system via the FHIR protocol. In step S1, the data acquisition module subscribes to the hospital's laboratory information system in real time. Once the test results of obstetric outpatients, such as blood glucose and glycated hemoglobin, are released, the system immediately and automatically triggers a complete risk assessment process.

[0093] The potential metabolic state modeling module is encapsulated as a high-performance container service. It supports dynamic scaling based on fluctuations in hospital outpatient volume. During a pregnant woman's visit, the doctor's decision support interface automatically pops up in the sidebar of the patient's electronic medical record. The interface displays the patient's metabolic system potential energy map in real time, visually demonstrating the patient's metabolic reserve capacity through the depth of the potential energy trap.

[0094] The counterfactual analysis function in step S5 is integrated into the prescription writing process. When a doctor enters a suggested diet or exercise prescription, the system immediately performs a counterfactual state analysis. If the analysis shows that the prescription cannot effectively reduce the probability of a high-risk persistent state, the system will issue a soft-interception alert and simultaneously display more effective alternative intervention options for the doctor's reference.

[0095] The results output module provides automated clinical report generation. The reports detail the evolution of latent variables such as insulin sensitivity and β-cell compensatory capacity throughout pregnancy, and correlate these trends with the implementation milestones of clinical interventions. This not only assists physicians in immediate decision-making but also accumulates high-quality structured data for departmental academic research.

[0096] The system's reinforcement learning self-optimization function collects data on doctors' final prescription choices and patients' actual outcomes, feeding this information back to the model as a reward function. With increased clinical application time, the intervention plans recommended by the system will increasingly align with the clinical experience of the hospital's experts, achieving a deep integration of medical expert knowledge and the artificial intelligence model.

[0097] Example 9: Physical derivation and principle verification of the core model. This example elaborates on the physical / physiological derivation basis of the key mathematical model in this invention, and verifies its clinical rationality and technical superiority through simulation experiments.

[0098] Physiological kinetic derivation of the blood glucose observation equation: The blood glucose observation equation mentioned in step S3 is as follows: ; This equation is not an empirical formula, but a mathematical abstraction based on the physiological regulatory mechanism of glucose-insulin homeostasis. Its derivation logic is as follows:

[0099] Insulin dependence of hepatic glucose output (HGP): Under physiological conditions, basal hepatic glucose output Insulin sensitivity Strict regulation. When insulin sensitivity is high, HGP is efficiently inhibited; conversely, HGP increases. This mechanism exhibits typical saturation nonlinearity: in Below the threshold When HGP is insensitive to changes in sensitivity, when Exceed At this point, the inhibitory effect rapidly manifests and tends to saturate. This S-shaped curve relationship can be precisely represented by the Sigmoid function. Precise description. Therefore, the first item... This represents the regulatory term of insulin sensitivity on hepatic endogenous glucose output.

[0100] The direct contribution of β-cell compensation to blood glucose: β-cell compensatory capacity This represents the insulin secretory reserve. Upon glucose stimulation, β-cells secrete insulin in phases one and two to lower blood glucose. However, in pathological conditions such as gestational diabetes, compensatory hyperinsulinemia itself does not directly lower blood glucose, but rather reflects the degree of systemic struggle against insulin resistance. Therefore, the second item... Introduced as a positive correlation term, it quantifies the steady-state blood glucose level indirectly boosted by increased compensatory demand.

[0101] Noise item It covers random factors not characterized by the model, including hormonal fluctuations (such as glucagon and cortisol) and measurement errors.

[0102] This equation connects the latent variables ( , With observable blood glucose levels, it enables non-invasive "vision" of deep metabolic functions.

[0103] A statistical mechanical analogy between the potential energy function and metabolic stability: The potential energy function defined in step S4: ; Its design is inspired by Langevin's equations and energy landscape theory in statistical mechanics.

[0104] First item (Hooke's potential energy): This represents changing the system from its current state. Pull back to steady-state center The "restorative" potential energy. In healthy individuals, metabolic state revolves around the mean. Random fluctuations, the distribution of which is determined by the covariance matrix Characterization. The further the system deviates from steady state (the larger the Mahalanobis distance), the higher the "energy barrier" the system needs to overcome, and the worse its stability. This quantitative relationship has been verified in a cohort study involving 500 normal pregnant women: their Mahalanobis distance... The 95th percentile corresponds precisely to the clinically defined threshold for impaired glucose tolerance.

[0105] The second item (local curvature penalty): Local geometric information from the state space is introduced. Jacobian matrix. The determinant reflects the "attractor strength" of the system near that point. When the value is very large, it indicates that the system is trapped in a local energy trap (such as a metastable state that is difficult to escape), even if and The distance is not large, but the actual risk of collapse is very high. This allows the risk assessment of this invention to identify high-risk individuals whose blood sugar is temporarily normal but whose metabolic system has become "rigid." Clinical data shows that approximately 15% of patients with target blood sugar levels have high risk of metabolic dysfunction. Those correctly classified as having metabolic metastable conditions and subsequently developing gestational diabetes requiring insulin therapy are 3.2 times more likely to be affected than the general population with normal blood sugar levels.

[0106] Simulation experiments and quantitative verification: To verify the effectiveness of the model, we constructed a simulation testing environment based on real clinical data characteristics.

[0107] Experimental setup: Dataset: 1000 simulated pregnancy metabolic trajectories were generated, covering the period from 12 weeks of gestation to delivery. Among them, 500 were "normal trajectories" generated based on parameters from healthy individuals, and 500 were "high-risk trajectories" infused with pathological features such as progressively declining insulin sensitivity and limited β-cell compensatory capacity.

[0108] Comparison method: The method of the present invention is compared with the traditional fasting blood glucose threshold method, logistic regression model and end-to-end LSTM model without counterfactual inference.

[0109] Evaluation metrics: the lead time of risk identification (how many weeks before clinical diagnosis is the warning issued) and the limitations of the intervention program (the root mean square error between the expected and actual reduction in Mahalanobis distance).

[0110] Experimental results: Risk identification lead time: The average warning time for high-risk trajectories using the method of this invention is 6.2 ± 1.5 weeks before diagnosis. In contrast, the average warning time using the fasting blood glucose threshold method is only 1.5 ± 0.8 weeks before diagnosis (depending on blood glucose exceeding the threshold). This demonstrates the early identification capability of this invention through deep latent variable modeling.

[0111] Metastable state identification capability: In a simulated population with "normal blood glucose but metabolic metastable state" (a total of 200 trajectories), the method of this invention correctly identified 162 of them (sensitivity 81%), while the traditional method failed to identify any. This is consistent with the aforementioned clinical observation data.

[0112] Intervention effectiveness prediction: For high-risk trajectories, the system recommended a TOP-3 intervention. The root mean square error (RMSE) between the predicted Mahalanobis distance decrease and the actual simulation results was 0.21, significantly lower than that of the LSTM model without counterfactual inference (RMSE 0.58). This indicates that structural causal models and counterfactual inference can accurately quantify the long-term benefits of interventions, providing high-confidence evidence for clinical decision-making.

[0113] Conclusion: The above physical derivation and simulation verification data fully demonstrate that the potential metabolic state modeling method proposed in this invention can identify the risk of gestational diabetes earlier and more accurately, and can effectively quantify the intervention plan. Its theoretical basis is solid and its technical effect is significantly better than the existing plan.

[0114] 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 the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for risk assessment and intervention decision-making in gestational diabetes mellitus based on potential metabolic state modeling, characterized in that, Includes the following steps: Step 1, Obtain and construct the minimum metabolic modeling data structure: Obtain data on gestational diabetes patients through a computer system, including individual static information data, follow-up observation data, and intervention data; The individual static information data includes age, pre-pregnancy body mass index, previous pregnancy history, and family history of diabetes. The follow-up observation data includes blood glucose indicators, weight, waist circumference, blood pressure, and behavioral-related data at different time points. The intervention data includes dietary intervention, exercise intervention, sleep management, and compliance information. The data are standardized to form a structured feature vector with temporal characteristics. Step 2, Initialize the latent metabolic state vector: Based on the individual static information data, the latent metabolic state vector of each patient is initialized by computer. The latent metabolic state vector includes insulin sensitivity state variables, beta cell compensation capacity state variables, fat burden state variables, and behavioral stability state variables. The above state variables are latent variables that cannot be directly measured. The nonlinear mapping relationship in the static features is processed by a deep learning network to estimate the initial probability distribution of the latent variables. Step 3, establish observation mapping and state update mechanism: establish the mapping relationship between follow-up observation data and potential metabolic state through computer, in which blood glucose index is used in combination with glucose kinetic equation to correct insulin sensitivity and beta cell compensation capacity, weight or waist circumference data is used to correct fat burden, and behavioral data is used to correct behavioral stability. Between adjacent follow-up time points, the distribution parameters of potential metabolic state vector are updated in real time through time-based discrete state transition model based on the marginal impact of intervention data on state variables. The mapping relationship includes the blood glucose observation equation: ； in, Let be the blood glucose observation value at time t; For basal glucose output from the liver; For insulin sensitivity state variables; For reference insulin sensitivity threshold; The nonlinear coefficient representing the effect of sensitivity on blood glucose; For the state variable of β-cell compensatory capacity; The contribution coefficient of compensatory capacity to blood glucose; To observe noise; The behavioral stability state variable is modified based on behavioral entropy, which is defined as: ； in, The normalized fluctuation frequency of the k-th type of behavior, including eating, exercising, and sleeping, within the time window; Between adjacent follow-up time points, the potential metabolic state vector is updated using a state transition model based on intervention data and observed changes. This state update mechanism employs a time-based discrete state transition model, and the state transition process follows the formula below: ； in, In order to be in The potential metabolic state vector at time t. In order to be in The data vector of the intervention applied at any given time. This is the state transition function, and its parameters are... It is obtained by pre-training a large amount of historical metabolic trajectory data through long short-term memory networks or gated recurrent units. For system random process noise that conforms to a normal distribution; Step 4, Metabolic state stratification and risk assessment: Based on the updated potential metabolic state vector, a potential energy function representing the energy topography of the metabolic system is constructed. The local optimal position of the potential metabolic state vector in the state space is evaluated. The Mahalanobis distance of the state point from the steady-state center is calculated as an index of metabolic system stability. Based on the Mahalanobis distance, patients are divided into metabolic homeostasis, metabolic metastasis, and high-risk persistent state, and the risk assessment results are output. The potential energy function is defined as: ； in, Let be the potential energy value under state s; This is the adjustment coefficient; It is the local Jacobian matrix in the state space, used to characterize the local curvature of the state points; Reflects the attractor strength of the system near that point; The stability index is defined as the Mahalanobis distance of the state point from the steady-state center, and its calculation formula is as follows: ； in, This represents the potential metabolic state vector at the current moment. This represents the mean metabolic state vector of a normal population at this gestational week. It is the inverse of the covariance matrix; Step 5, Evaluation of intervention decision based on counterfactual inference: Under the current potential metabolic state, at least one hypothetical intervention plan is constructed by computer. The hypothetical intervention plan is then subjected to counterfactual state inference using a structural causal model. The evolution path of the potential metabolic state under different intervention variables in the future prediction time period is simulated. By comparing the metabolic stability gains generated by the evolution path, the intervention priority or intervention recommendation is determined. In the counterfactual deduction, the posterior distribution of the exogenous noise variable is calculated as follows: ； in, State transition function process noise variance Assuming noise priors, we assume a Gaussian distribution. Step 6, Output Results: Output the risk assessment results and intervention decision results to the display terminal, and visualize the pathological mechanism focus by displaying a multi-dimensional potential metabolic state radar map.

2. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of obtaining and constructing the minimal metabolic modeling data structure in step 1 includes: By calling the electronic medical record system through a standardized interface, the baseline characteristics of the subject in early pregnancy or before pregnancy can be extracted. Convert age data into numerical features in a preset time unit; The pre-pregnancy body mass index was calculated by dividing the weight by the square of the height. Extract the total number of pregnancies and specific signs of fetal delivery from past pregnancy history, and then perform Boolean value processing on the records; The number of people with diabetes among relatives with a predetermined relationship is obtained, and weighted coefficients are assigned according to the closeness of the kinship, which are then quantified into a feature vector of family history of diabetes. The follow-up observation data were aligned to a time series, and the physiological indicators collected by the handheld blood glucose meter and wearable device were synchronized in real time via a wireless transmission protocol. Establish measurement error control logic to perform data smoothing when data fluctuations exceed the preset error range.

3. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of obtaining intervention data in step 1 includes: The dietary intervention records were quantified to analyze the types and weights of food recorded by the subjects and convert them into predetermined daily intake ratios of carbohydrates, protein, and fat. Record the predetermined duration of daily exercise at a specific intensity during exercise intervention, and calibrate the exercise intensity by combining heart rate data obtained from wearable devices; By monitoring and obtaining the total daily sleep duration and the planned percentage of deep sleep through sensors, sleep management data can be constructed. A comprehensive evaluation is conducted based on the interaction activity, record completeness, and suggestion execution rate of the mobile application to obtain compliance information, which is then mapped to a preset score range.

4. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of initializing the potential metabolic state vector in step 2 includes: The individual static information data is processed using a multilayer perceptron network, with at least three hidden layers, and the initial mean and variance of the potential metabolic state vector are output. The initial values ​​of the insulin sensitivity state variable were determined based on the nonlinear mapping relationship between pre-pregnancy body mass index and age. The initial values ​​of the state variable of beta cell compensatory capacity were determined based on the prior probabilities of previous pregnancy history and family genetic background. All state variables are assumed to conform to a multivariate normal distribution, and latent variable inference is performed using a variational inference mechanism to ensure the statistical independence of each state variable and eliminate the multicollinearity effect caused by redundant features in static information.

5. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of establishing the observation mapping relationship in step 3 includes: A probabilistic model based on the observation likelihood function is used to map unobservable latent variables to observable clinical indicators. The evolutionary relationship between blood glucose levels and latent variables was described using glucose kinetic equations, simulating the insulin-mediated glucose uptake process and the pulsatile secretion characteristics of beta cells. The impact of physiological noise on observation accuracy is eliminated by introducing a nonlinear mapping function; The state variables of fat burden are used to reflect the absolute amount of adipose tissue and the inhibitory effect of inflammatory factors produced by adipose tissue on overall metabolism. The fluctuations in exercise duration, sleep quality, and dietary intake in the follow-up observation data are defined as behavioral entropy, which reflects the impact of behavioral regularity on the metabolic system.

6. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The state update mechanism and the logic for correcting the behavioral stability state using behavioral data in step 3 include: A time-based discrete state transition model is adopted to determine the potential metabolic state at the next moment based on the potential metabolic state vector at the current moment, the intervention data vector applied at the current moment, and the state transition function parameters obtained through pre-training. System stochastic process noise that conforms to a normal distribution is introduced for calibration. The state transition function parameters are obtained by training historical metabolic trajectory data through a long short-term memory network or a gated recurrent unit. Set the synchronization period for the transition probability matrix in the state transition model, and periodically perform global synchronization to capture the metabolic environment fluctuations in the pregnant mother caused by changes in placental hormone levels. When the behavioral entropy increases and the duration exceeds a predetermined time threshold, the value of the behavioral stability state variable is reduced, and the change in behavioral stability state is fed back to other variables in the potential metabolic state vector through the system coupling coefficient.

7. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of calculating the metabolic system stability index in step 4 includes: Construct a potential energy function that characterizes the energy distribution of a metabolic system in the state space; Obtain the potential metabolic state vector at the current moment, and obtain the mean metabolic state vector of the normal population at the corresponding gestational week; Calculate the difference between the potential metabolic state vector and the mean metabolic state vector of the normal population; The difference is multiplied by the inverse of the covariance matrix, and a quadratic operation is performed using the transpose of the difference to obtain the Mahalanobis distance, which reflects the degree of deviation of the system from the steady-state center. The metabolic stratum to which the patient belongs is determined by comparing the Mahalanobis distance with the first preset threshold and the second preset threshold.

8. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The process of outputting the risk assessment results in step 4 includes: Predict the probability of severe metabolic disorders and related complications in late pregnancy; Calculate the long-term risk score of developing a specific type of diabetes within a predetermined time after delivery; ； in, The Mahalanobis distance threshold for the normal population at the corresponding time point; , As a risk attenuation factor; This is the end time of the follow-up visit; The long-term risk score is calculated by weighting the integral of the duration of metabolic metastability during pregnancy; When the long-term risk score reaches the preset risk warning threshold, a high-risk warning is triggered to the medical staff through an encrypted communication protocol.

9. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The intervention decision evaluation process based on counterfactual reasoning in step 5 includes: A directed acyclic graph (DAG) containing intervention variables, potential metabolic state mediator variables, and risk stratified outcome variables was constructed, and the path weights in the DAG were calibrated using observation sequences. Given the current timestamp, perform abductive operations to calculate the posterior distribution of the exogenous noise variables behind the current observations; Perform intervention operations to forcibly modify the values ​​of intervention variable nodes in the model in order to construct a hypothetical intervention plan; Perform prediction operations to deduce the evolution path of potential metabolic states within a preset prediction time period while keeping exogenous noise constant; Multiple evolutionary pathways were generated using the Monte Carlo sampling method, and all evolutionary pathways were aggregated with probability weighting to calculate the metabolic stability gain and confidence interval for each hypothetical intervention scheme.

10. The method for risk assessment and intervention decision-making of gestational diabetes mellitus based on potential metabolic state modeling according to claim 1, characterized in that, The logic for intervention decision-making and outcome output in steps 5 and 6 includes: The benefit-cost ratio of all simulated hypothetical intervention plans was calculated using a multi-objective optimization algorithm, where the optimization objectives included maximizing the improvement of metabolic stability, maximizing the probability of compliance, and minimizing the degree of disturbance to quality of life. Benefit is defined as the extent of the reduction in metabolic system risk, and cost is defined as the expected intensity of the intervention on the patient's quality of life. ； in, The decrease in Mahalanobis distance resulting from intervention i; The probability of compliance predicted by the model; The intensity of the intervention includes the duration of exercise and the degree of dietary restriction; Rate the loss in quality of life; These are the weighting coefficients; Through Pareto optimal search, a predetermined number of intervention strategies that meet the balance requirements are recommended first. The expected effects and predicted uncertainty range of each strategy are displayed on the display terminal in the form of dynamic evolution curves; By utilizing the discrepancy between the actual postpartum metabolic recovery trajectory and the predicted trajectory during pregnancy, the parameters in the state transition model are automatically optimized through reinforcement learning algorithms, thereby achieving self-evolution of the model's accuracy. A radar map displaying the patient's potential metabolic status in real time is provided. The dimensions of the radar map correspond to insulin sensitivity, beta cell compensatory capacity, fat burden, and behavioral stability, respectively.

Images