A personalized nutritional intervention system for metabolic comorbidities of the elderly based on multi-omics screening

Abstract

This invention discloses a personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening, belonging to the field of geriatric health technology. It includes a data acquisition and processing module, a unified feature extraction module, an environmental state identification module, a metabolic map construction module, a causal inference and analysis module, a twin modeling prediction module, a nutritional plan generation module, an intervention simulation evaluation module, and a safety joint evaluation module. This invention significantly enhances the comprehensiveness and accuracy of the analysis. Compared with traditional static evaluation methods, this invention can identify in real time whether the patient is in an adaptive or risky state, thus providing a reliable basis for the dynamic adjustment of subsequent intervention strategies. It improves the system's adaptability to complex environmental changes, avoids the spurious correlation problem that easily occurs in traditional correlation analysis, and makes the identification of intervention targets more accurate. This helps guide the design of nutritional plans from a mechanistic perspective, improves the scientificity and effectiveness of intervention, and enhances the safety and controllability of intervention.

Description

Technical Field

[0001] This invention relates to the field of geriatric health technology, and in particular to a personalized nutritional intervention system for geriatric metabolic comorbidities based on multi-omics screening. Background Technology

[0002] With the increasing aging of the population in high-altitude areas, metabolic diseases among the elderly in these regions exhibit a significant characteristic of "multiple co-occurrence and mutual coupling," typically manifesting as insulin resistance, abnormal uric acid metabolism, and bone metabolism disorders. These metabolic comorbidities are not only influenced by individual genetic background and lifestyle but are also closely related to the long-term effects of the hypoxic environment at high altitudes. Because the hypoxic environment has a continuous impact on energy metabolism, oxidative stress levels, and endocrine regulation, traditional nutritional intervention strategies based on single diseases or single indicators are difficult to achieve ideal results. Current technologies for nutritional intervention largely rely on empirical guidelines or static assessment methods, lacking the ability to jointly analyze multi-omics data (such as genomics, metabolomics, and microbiome) and dynamic physiological indicators, making it difficult to reveal the causal relationships between different metabolic abnormalities. Furthermore, existing methods lack systematic modeling tools for the unique "acclimatization-disaccimation" dynamic change mechanism of the high-altitude environment, failing to accurately assess individual differences in metabolic responses under different environmental exposure conditions. In addition, traditional intervention programs lack predictive validation stages during the development process, failing to assess in advance the potential impact of nutritional intake on blood pressure fluctuations, renal burden, and oxidative stress, posing certain safety risks. Therefore, there is an urgent need for a personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening.

[0003] Existing personalized nutritional intervention systems for metabolic comorbidities in the elderly have low comprehensiveness and accuracy. Traditional static assessment methods cannot identify whether patients are in an adaptive or risky state in real time, and cannot provide a reliable basis for the dynamic adjustment of subsequent intervention strategies. At the same time, the spurious correlation problem in correlation analysis reduces the accuracy of intervention target identification, which is not conducive to guiding the design of nutritional programs from the mechanism level, and reduces the scientificity and effectiveness of intervention. Therefore, we propose a personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening includes a data acquisition and processing module, a unified feature extraction module, an environmental state identification module, a metabolic map construction module, a causal inference and analysis module, a twin modeling prediction module, a nutritional plan generation module, an intervention simulation evaluation module, a safety joint evaluation module, a dietary conversion output module, and a dynamic scheduling and updating module. The data acquisition and processing module receives the patient's static and dynamic basic data, and performs missing data repair, anomaly removal, unit unification, centralization and normalization processing on the multi-source data. It also performs cross-device, cross-frequency and cross-scenario data alignment based on a unified time axis. The unified feature extraction module filters out multiple omics variables based on preprocessed multi-source data and extracts corresponding static multi-omics features. The environmental state recognition module is used to identify the patient's current environment, generate corresponding plateau environmental state parameters, and analyze the differences in the patient's environmental response under different altitudes and different living scenarios. The metabolic mapping module constructs a dynamic map of metabolic comorbidities in the elderly based on the screened corresponding variable characteristics, clinical phenotypes, and plateau environmental state parameters. The causal inference analysis module infers the causal relationship between changes in various indicators based on the dynamic map of metabolic comorbidities in the elderly, and identifies the direct and indirect effects of hypoxia exposure. The twin modeling prediction module constructs a corresponding virtual patient model based on the patient's static multi-omics characteristics, differences in environmental response, and causal relationships between changes in various indicators. It simulates the patient's physiological evolution trend and predicts metabolic responses under different nutritional intake conditions. The nutrition plan generation module automatically generates multiple candidate nutrition intervention plans based on the identified simulation results and causal inference results, combined with the patient's current habituation status, metabolic load level and nutritional deficit. The intervention simulation evaluation module uses a virtual patient model to reason and search for candidate nutritional intervention programs, simulates the effects of different dosage combinations, and screens feasible nutritional programs. The safety joint evaluation module performs dual judgment on the simulation results. When a candidate solution simultaneously meets the preset effectiveness threshold and safety constraints, it is marked as an output nutrition solution. The dietary conversion output module will convert the evaluated output-ready nutritional solutions into practical and implementable lifestyle solutions. The dynamic scheduling and update module combines the patient's behavioral responses, intake records, location status, sleep rhythm, and interactive feedback to dynamically adjust the reminder time window and push method.

[0006] As a further aspect of the present invention, the static and dynamic basic data include genomes, metagenomics, metabolomics, clinical laboratory indicators, continuous blood glucose, heart rate variability, sleep oxygenation, blood pressure, medication records, dietary records, and information on high-altitude exposure duration, altitude changes, and residence / migration; the multi-omics variables specifically include multi-omics variables such as genes, proteins, metabolites, and microbial abundance; the variable features are specifically high-weighted features related to insulin resistance, uric acid metabolism disorders, and bone metabolism abnormalities.

[0007] As a further aspect of the present invention, the environmental state recognition module identifies the patient's current environment, generates corresponding plateau environmental state parameters, and analyzes the differences in the patient's environmental response under different altitudes and living scenarios. The specific steps are as follows: S1.1: A unified system is used to read the patient's current location coordinates, real-time altitude, meteorological parameters, oxygen partial pressure monitoring values, activity trajectory points, and historical high-altitude exposure records. Time-attenuation weighted interpolation is then used to map various data from different sampling frequencies onto the same time axis. The specific calculation formula for time-attenuation weighted interpolation is as follows: ; In the formula, Representative moment No. Alignment values ​​for class data; Representing the The first class of data One original observation value; Representing the The first class of data The sampling time of each original observation; Representing the The number of observation points for this type of data; Represents the time decay coefficient; S1.2: After completing time alignment, the current altitude is standardized to obtain the corresponding altitude exposure factor. Then, the current change in oxygen partial pressure and the oxygen partial pressure decrease factor are calculated based on the current total atmospheric pressure. Next, the corresponding meteorological load factor is calculated based on the patient's current ambient temperature, relative humidity, and total atmospheric pressure to obtain the overall environmental pressure on the body. The specific calculation formula for the meteorological load factor is as follows: ; In the formula, Representative moment Meteorological load factors; Representative moment Temperature; Represents reference temperature; Represents the temperature normalization range; Representative moment relative humidity; Representative moment Total ambient air pressure; This represents the stable reference air pressure value; as well as These represent the weighting coefficients; S1.4: Based on the patient's trajectory speed and dwelling behavior at different times, and simultaneously calculating the spatial migration distance at each time point, the active-residential exposure factor is calculated. Then, based on the active-residential exposure factor, the patient's environmental exposure patterns in different living scenarios are distinguished. The specific calculation formula for the active-residential exposure factor is as follows: ; In the formula, Representative moment The activity and residential exposure factors; Represents the number of sampling points at the current moment; Representative sampling point The trajectory velocity; This represents the upper limit of the normalized velocity. Representative sampling point Duration of stay; Represents the normalized upper limit of the duration of stay; Representative moment Spatial migration distance; Represents the maximum migration distance within the window; as well as These represent the weighting coefficients; S1.5: The altitude exposure factor, oxygen partial pressure decrease factor, meteorological load factor, and activity / habitation exposure factor are weighted and summed to obtain the patient's current high-altitude environmental exposure intensity. Then, based on the patient's continuous exposure time, exposure fluctuation, and oxygen partial pressure change trend, the corresponding adaptation index is calculated. Finally, based on the environmental exposure intensity and the adaptation index, the patient's disadaptation risk value is calculated, and a corresponding risk threshold is set. The specific calculation formula for the high-altitude environmental exposure intensity is as follows: ; In the formula, For a moment The intensity of exposure to the high-altitude environment; For a moment Altitude exposure factors; For a moment The oxygen partial pressure decreasing factor; For a moment Meteorological load factors; For a moment The activity and residential exposure factors; For a moment Historical cumulative exposure items; These are the fusion weights, and the sum of the weights is 1; the specific calculation formula for the historical cumulative exposure item is as follows: ; In the formula, This represents the total length of the current time window. For a moment The oxygen partial pressure decreasing factor; The specific formula for calculating the adaptation index is as follows: ; In the formula, The acceptance index ranges from 0 to 1. This refers to the duration of continuous high-altitude exposure. The fluctuations exposed within the current time window; This represents the average value of the oxygen partial pressure change trend within the current time window; This represents the normalized upper limit of continuous high-altitude exposure time; This is the normalized upper limit for environmental exposure intensity; This represents the normalized upper limit of the rate of decrease in oxygen partial pressure. These are the weighting coefficients; S1.6: If the patient's acclimatization risk value is greater than or equal to the risk threshold, it indicates that the current exposure has exceeded the patient's tolerance boundary and the acclimatization risk is increased. Conversely, it indicates that the current environment is still within the patient's relatively tolerable range. At the same time, the average exposure intensity of the patient in the high-altitude pasture scenario and the average exposure intensity in the settled town scenario are collected in real time, and the difference between the two sets of data is obtained. Then, based on the difference at different times and the average exposure intensity in the settled town scenario, the scene response difference of the patient in different altitudes and residential scenarios is obtained, and the environmental response differences between different residential scenarios are recorded.

[0008] As a further aspect of the present invention, the specific steps of the metabolic mapping construction module in constructing a dynamic map of metabolic comorbidities in the elderly are as follows: S2.1: Extract the selected multi-omics features, clinical phenotypes, and plateau environmental state parameters from the original data records of omics data, clinical records, and environmental observations. Use the multi-omics features, clinical phenotypes, and plateau environmental state parameters as entities to be mapped. Map synonyms, abbreviations, and different naming methods across platforms in a unified manner to construct a standard entity set. S2.2: After completing the standardized mapping, knowledge nodes are constructed based on the standard entity set for genes, proteins, metabolites, clinical phenotypes and plateau environment state parameters, respectively. The source attributes, functional attributes and context attributes of each node are integrated into a unified feature vector, and the node features are mapped to a unified semantic space through linear encoding. S2.3: After the nodes are established, the number of evidence supports, cross-validation consistency and literature or database reliability of each node in different data sources are counted respectively. The results are normalized and weighted according to a uniform scale to obtain the initial importance of each node. From the node set, two nodes are randomly selected and potential associations are simultaneously screened from multi-omics correlation, clinical co-occurrence pattern and environmental response pattern. Then, the initial strength of the connection edge between the two corresponding nodes is calculated through multi-evidence fusion. S2.4: When the initial strength is higher than the preset threshold, retain the connection edge; otherwise, delete it. Calculate the matching score between the node pair corresponding to each connection edge and each relation semantic. After normalizing the obtained matching scores, select the relation semantic with the highest normalized matching score as the relation type of the corresponding connection edge. Then, combine the candidate edge strength and relation type coefficient to determine the initial edge weight of the corresponding connection edge to complete the construction of the dynamic map of metabolic comorbidities in the elderly. S2.5: When new omics data, clinical records and environmental observations are integrated, the entities to be built in the graph are generated again, and the existing nodes and edges in the dynamic graph of metabolic comorbidities in the elderly are incrementally updated to reflect the latest knowledge of the patient's status after the evolution over time in real time. At the same time, the change amplitude of the entire dynamic graph of metabolic comorbidities in the elderly between adjacent rounds is statistically analyzed. S2.6: If the change exceeds the preset threshold, the relation reorganization process is triggered to re-evaluate the corresponding connection edges, reduce the weight of edges lacking new evidence, or delete them. Then, the activation intensity of each disease-related subgraph in the dynamic graph of metabolic comorbidities in the elderly is calculated, and the disease subgraph with an activation intensity higher than the preset selection value is output.

[0009] As a further aspect of the present invention, the specific steps of the causal inference analysis module inferring the causal relationship between changes in various indicators and identifying the direct and indirect effects of hypoxia exposure are as follows: S3.1: Candidate causal pathways between high-altitude hypoxia exposure, acclimatization status, and metabolic comorbidity outcomes were identified from the latest dynamic map of metabolic comorbidity in the elderly and its disease sub-maps. High-altitude hypoxia exposure was designated as the exposure node, acclimatization status as the mediating node, and metabolic comorbidity outcome as the outcome node. Insulin resistance, uric acid metabolism disorders, and bone metabolism abnormalities were identified as the outcome types for the outcome nodes to be analyzed in each candidate causal pathway. The cumulative score for each candidate causal pathway was calculated using the following formula: ; in, The cumulative score representing each candidate causal path; Represents the total number of candidate causal paths; Represents either side of the path; This represents the strength of the support relationship between corresponding edges; S3.2: Select upstream nodes that are connected to both the exposure node and the outcome node from the dynamic map of metabolic comorbidities in the elderly, and evaluate the pseudo-association between the corresponding variables of each upstream node and the weakening exposure and the outcome by the cumulative score of each candidate causal path. Based on the evaluation results, establish a backdoor adjustment set. After screening the backdoor adjustment set, quantitatively estimate the total causal effect of high altitude hypoxia exposure on each outcome to obtain the overall impact strength of hypoxia exposure on the three types of metabolic comorbidities, i.e., the total effect. S3.3: Validate any mediator node in each candidate causal path and calculate the validity score of each mediator node. If the validity score is higher than the preset screening value, the corresponding mediator node is included in the front door mediator set. The front door criterion is used to calculate the estimated expectation of the corresponding outcome node under the given exposure node and mediator node, i.e. the mediator transmission effect. S3.4: Based on the total effect and the mediating transmission effect, the direct and indirect effects of hypoxia exposure on each outcome are decomposed. After decomposition, the total effect, direct effect and indirect effect of the three types of outcomes are statistically analyzed, and the magnitude of the total effect, direct effect and indirect effect of the three types of outcomes are compared. The mediating proportion is used to determine whether hypoxia exposure mainly works through the direct damage pathway or mainly through the habituation pathway. The stability and reliability of the inference results are verified by using resampling confidence intervals, and the causal relationship between the changes of each indicator is output based on the verification results.

[0010] As a further aspect of the present invention, the twin modeling prediction module constructs a corresponding virtual patient model, simulates the physiological evolution trend of the patient, and predicts the metabolic response under different nutritional intake conditions. The specific steps are as follows: S4.1: Obtain the patient's multi-omics features, environmental state parameters, and causal inference results, and map data with different sampling frequencies onto the same time axis to unify the prediction time scale of various types of data. Then, in the order of "multi-omics features - environmental state - causal results", the time-corresponding data are concatenated into a unified input vector. Then, the input vector is mapped to a low-dimensional latent space to obtain the initial hidden state of the virtual patient at different times. S4.2: After obtaining the initial hidden state, construct a continuous-time dynamic equation for simulating the physiological evolution of the patient, and enable it to output the state result at any prediction time to complete the construction of the corresponding virtual patient model. Encode each candidate nutritional intervention plan into a structured plan vector, then map each plan into a continuous-time control signal, and input each continuous-time control signal into the corresponding virtual patient model. S4.3: Run the virtual patient model under different nutritional conditions, based on the continuous-time dynamic equation, gradually advance the virtual patient state, and then convert the hidden state into various interpretable physiological indicators and form short-term, medium-term and long-term evolution trajectories to obtain the dynamic impact of different nutritional conditions on the patient's physiological state. S4.4: Verify the consistency and stability of the dynamic effects of various nutritional conditions on the patient's physiological state. After the verification is passed, based on the output physiological indicators and evolution trajectories, compare the degree of deviation of different plans at each time point and each indicator, calculate the overall response score under different nutritional conditions, and then arrange each candidate nutritional intervention plan in descending order of overall response score, while outputting various physiological indicators and corresponding evolution trajectories.

[0011] As a further aspect of the present invention, the physiological indicators mentioned in S4.3 specifically include various data such as blood glucose, metabolic level, and circulatory indicators; the short-term, medium-term, and long-term are typically set to 7 days, 30 days, and 90 days, respectively.

[0012] As a further aspect of the present invention, the specific steps of the nutrition plan generation module in automatically generating multiple candidate nutrition intervention plans are as follows: S5.1: Based on the causal inference results, all paths are expanded according to the "exposure node - mediator node - outcome node" link, and the node relationships, directional consistency and evidence strength on each link are jointly evaluated. At the same time, the comprehensive support of each link is generated, and the link with the highest comprehensive support and directional consistency is selected as the key pathway for the corresponding patient. S5.2: After the key pathways are identified, the acclimatization stability, volatility and recovery capacity are extracted based on the corresponding status information of high altitude adaptation, and the corresponding acclimatization status index is calculated. Then, the three types of loads, namely insulin resistance, abnormal uric acid and abnormal bone metabolism, are weighted and fused to obtain the corresponding current status information of the patients. S5.3: Based on the patient's current status information, the intake of protein, essential amino acids, vitamins, minerals and corresponding functional nutrients is statistically analyzed. The gap between these intakes and the preset target nutritional needs is calculated item by item to form a nutritional gap vector. First, the type of nutritional support that each pathological pathway relies on is determined. Then, the needs are processed in combination with the patient's current high status information and habituation level to form the demand intensity of each nutrient. S5.4: Based on the intensity of demand for each nutrient, perform a combination search from the preset nutrient library and food combination library to generate multiple candidate nutritional intervention plans. Calculate the comprehensive score of each candidate nutritional intervention plan based on demand matching degree, feasibility, and plan differentiation. Sort the generated multiple candidate nutritional intervention plans in descending order of comprehensive score. Then, combine the plan differentiation to screen out multiple sets of nutritional intervention plans that do not overlap with each other and cover different nutritional focuses, and output multiple sets of candidate nutritional intervention plans.

[0013] As a further aspect of the present invention, the intervention simulation evaluation module simulates the effects of different dosage combinations, and the specific steps for screening feasible nutritional regimens are as follows: S6.1: Based on the generated candidate nutritional intervention plans, the available dose of each nutrient is divided into multiple discrete levels according to the type of nutrient. Then, the dose levels of different nutrients are combined in a Cartesian combination to generate multiple intake combinations in order to establish a candidate intervention path space. S6.2: Decompose each group of candidate nutritional intervention schemes in the candidate intervention path space into continuous inputs over multiple time periods, and transform them into counterfactual control sequences that act over time. Then, input the counterfactual control sequences into the corresponding virtual patient model to generate state trajectories under different schemes. S6.3: After generating the state trajectory, map it to metabolic indicators, blood pressure fluctuation indicators, kidney burden indicators and oxidative stress indicators, and generate the prediction observation vector of each candidate nutritional intervention program. At the same time, calculate the benefit and safety penalty value of each candidate nutritional intervention program at different times. S6.4: Combine the benefits and safety penalty values ​​of each candidate nutritional intervention plan at different times to form its corresponding comprehensive plan value, and perform enhanced search and screening in the candidate intervention path space. If the benefit is greater than or equal to the preset effectiveness threshold and the safety penalty value is less than or equal to the preset safety penalty value, the corresponding candidate nutritional intervention plan is retained; otherwise, it is removed. The remaining candidate nutritional intervention plans are then input into the corresponding virtual patient model in descending order of comprehensive plan value.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention collects multi-source data, including patient location, altitude, meteorological parameters, oxygen partial pressure, and activity trajectory. Through time alignment and standardization, it calculates the intensity of exposure to the high-altitude environment, the adaptation index, and the risk of non-adaptation, obtaining the patient's environmental response characteristics in different scenarios. Subsequently, it establishes a dynamic atlas of metabolic comorbidities in the elderly, incorporating genes, proteins, metabolites, and environmental factors, and continuously optimizes and updates the atlas structure in real time. From this dynamic atlas, it extracts causal pathways between high-altitude hypoxia exposure, adaptation status, and metabolic comorbidity outcomes, using backdoor and frontdoor criteria for causal inference and quantitatively decomposing direct and indirect effects. Then, it inputs multi-omics characteristics, environmental conditions, and causal results into a digital twin model to construct a virtual patient and simulates the dynamic impact of different nutritional interventions on physiological states within a continuous time frame, generating multi-timescale evolutionary trajectories. Based on key pathological pathways, patient metabolic load, and nutritional deficits, this invention automatically generates multiple candidate nutritional intervention programs. These programs are then ranked and screened through needs matching and feasibility assessment. Finally, counterfactual reasoning and reinforcement search are performed on the candidate programs based on a virtual patient model to comprehensively evaluate the metabolic improvement effect and safety risks. This process selects personalized nutritional intervention programs that are both effective and safe, significantly enhancing the comprehensiveness and accuracy of the analysis. Compared to traditional static assessment methods, this invention can identify in real time whether a patient is in an adaptive or risky state, thus providing a reliable basis for the dynamic adjustment of subsequent intervention strategies. It improves the system's adaptability to complex environmental changes, avoids the spurious correlation problem that easily occurs in traditional correlation analysis, and makes the identification of intervention targets more accurate. This helps guide the design of nutritional programs from a mechanistic perspective, improving the scientific nature and effectiveness of interventions, as well as enhancing the safety and controllability of interventions. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a system block diagram of a personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening proposed in this invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening includes a data acquisition and processing module, a unified feature extraction module, an environmental state identification module, a metabolic map construction module, a causal inference and analysis module, a twin modeling prediction module, a nutrition plan generation module, an intervention simulation evaluation module, a safety joint evaluation module, a diet conversion output module, and a dynamic scheduling and updating module.

[0019] The data acquisition and processing module receives static and dynamic basic data of patients and performs missing data repair, anomaly removal, unit unification, centering and normalization on the multi-source data. It also performs cross-device, cross-frequency and cross-scenario data alignment based on a unified time axis. The feature extraction module selects multi-omics variables based on the preprocessed multi-source data and extracts the corresponding static multi-omics features.

[0020] The environmental status recognition module is used to identify the patient's current environment, generate corresponding plateau environmental status parameters, and analyze the differences in the patient's environmental response at different altitudes and in different living scenarios.

[0021] Specifically, the system uniformly reads the patient's current location coordinates, real-time altitude, meteorological parameters, oxygen partial pressure monitoring values, activity trajectory points, and historical high-altitude exposure records. Time-attenuation weighted interpolation is used to map various data from different sampling frequencies onto the same time axis. After time alignment, the current altitude is standardized to obtain the corresponding altitude exposure factor. Then, the current oxygen partial pressure change and oxygen partial pressure decrease factor are calculated based on the current total atmospheric pressure. Next, the corresponding meteorological load factor is calculated based on the patient's current ambient temperature, relative humidity, and total atmospheric pressure to obtain the comprehensive environmental pressure on the body. Based on the patient's trajectory speed and dwelling behavior at different times, and simultaneously calculating the spatial migration distance at each time point, the activity-residential exposure factor is calculated. Finally, based on the activity-residential exposure factor, the environmental exposure patterns of the patient under different living scenarios are distinguished, and the altitude exposure factor, oxygen partial pressure decrease factor, and other parameters are correlated. The patient's current exposure intensity to the high-altitude environment was obtained by weighted summation of the exposure load factor and the active residence exposure factor. Then, based on the patient's continuous exposure time, exposure fluctuation, and oxygen partial pressure change trend, the corresponding adaptation index was calculated. Based on the environmental exposure intensity and the adaptation index, the patient's risk value for disadaptation was calculated, and a corresponding risk threshold was set. If the patient's risk value for disadaptation is greater than or equal to the risk threshold, it means that the current exposure has exceeded the patient's tolerance adaptation boundary, and the risk of disadaptation is increased. Conversely, it means that the current environment is still within the patient's relatively tolerable range. At the same time, the average exposure intensity of the patient in the high-altitude pasture scenario and the average exposure intensity in the settled town scenario were collected in real time, and the corresponding difference between the two sets of data was obtained. Then, based on the difference at different times and the average exposure intensity in the settled town scenario, the scene response difference of the patient in different altitudes and residence scenarios was obtained, and the environmental response differences between different residence scenarios were recorded.

[0022] The metabolic atlas construction module constructs a dynamic atlas of metabolic comorbidities in the elderly based on screened multi-omics features, clinical phenotypes, and high-altitude environmental parameters.

[0023] Specifically, multi-omics features, clinical phenotypes, and high-altitude environmental state parameters were extracted from raw data records of omics data, clinical records, and environmental observations. These features, clinical phenotypes, and high-altitude environmental state parameters were then used as entities to construct a map. Synonyms, abbreviations, and different naming conventions across platforms were uniformly mapped to build a standard entity set. After standardization, knowledge nodes were constructed based on these standard entity sets for genes, proteins, metabolites, clinical phenotypes, and high-altitude environmental state parameters. The source attributes, functional attributes, and contextual attributes of each node were integrated into a unified feature vector. Linear encoding was used to map the node features to a unified semantic space. After node construction, the number of pieces of evidence supporting each node across different data sources, cross-validation consistency, and literature or database reliability were statistically analyzed. These results were normalized and weighted according to a unified scale to obtain the initial importance of each node. From the node set, two nodes were randomly selected, and potential associations were simultaneously screened based on multi-omics correlations, clinical co-occurrence patterns, and environmental response patterns. Further, multiple pieces of evidence were used to further analyze the associations. The initial strength of the connection edge between two corresponding nodes is calculated. If the initial strength is higher than a preset threshold, the connection edge is retained; otherwise, it is deleted. The matching score between the node pair corresponding to each connection edge and each relation semantic is calculated. After normalizing the obtained matching scores, the relation semantic with the highest normalized matching score is selected as the relation type of the corresponding connection edge. Then, the initial edge weight of the corresponding connection edge is determined by combining the candidate edge strength and relation type coefficient to complete the construction of the dynamic atlas of metabolic comorbidities in the elderly. When new omics data, clinical records, and environmental observations are added, the entities to be constructed are generated again. The existing nodes and edges in the dynamic atlas of metabolic comorbidities in the elderly are incrementally updated to reflect the latest knowledge of the patient's status evolving over time. At the same time, the change amplitude between two adjacent rounds of the dynamic atlas of metabolic comorbidities in the elderly is counted. If the change amplitude is higher than a preset threshold, the relation reorganization process is triggered to re-evaluate the corresponding connection edge and reduce the weight or delete the edge that lacks new evidence. Then, the activation intensity of the subgraph corresponding to each disease in the dynamic atlas of metabolic comorbidities in the elderly is calculated, and the subgraph of disease with an activation intensity higher than the preset selection value is output.

[0024] The causal inference analysis module, based on the dynamic map of metabolic comorbidities in the elderly, infers the causal relationships between changes in various indicators and identifies the direct and indirect effects of hypoxia exposure.

[0025] Specifically, candidate causal pathways between high-altitude hypoxia exposure, acclimatization status, and metabolic comorbidity outcomes were identified from the latest dynamic map of metabolic comorbidity in the elderly and its disease subplots. High-altitude hypoxia exposure was designated as the exposure node, acclimatization status as the mediator node, and metabolic comorbidity outcomes as the outcome nodes. Insulin resistance, uric acid metabolism disorders, and bone metabolism abnormalities were identified as outcome types for the outcome nodes to be analyzed in each candidate causal pathway. Cumulative scores for each candidate causal pathway were calculated. Upstream nodes connected to both the exposure and outcome nodes were selected from the dynamic map of metabolic comorbidity in the elderly. The spurious associations between the corresponding variables of each upstream node and the weakening of exposure and outcome were assessed using the cumulative scores of each candidate causal pathway. A backdoor adjustment set was established based on the assessment results. After screening the backdoor adjustment set, the total causal effect of high-altitude hypoxia exposure on each outcome was quantitatively estimated to obtain the impact of hypoxia exposure on the three types of metabolic comorbidity outcomes. The overall impact strength of the exposure, i.e., the total effect, is determined by evaluating the effectiveness of any mediator node in each candidate causal path and calculating the effectiveness score of each mediator node. If the effectiveness score is higher than the preset screening value, the corresponding mediator node is included in the front-door mediation set. The front-door criterion is used to calculate the estimated expected value of the corresponding outcome node under a given exposure node and mediator node, i.e., the mediating transmission effect. Based on the total effect and the mediating transmission effect, the direct and indirect effects of hypoxia exposure on each outcome are decomposed. After decomposition, the total effect, direct effect, and indirect effect of the three types of outcomes are statistically analyzed, and the magnitude of the total effect, direct effect, and indirect effect of the three types of outcomes is compared. The mediation ratio is used to determine whether hypoxia exposure mainly works through the direct damage path or the habituation path. The stability and reliability of the inference results are verified using the resampling confidence interval, and the causal relationship between the changes of each indicator is output based on the verification results.

[0026] Example 2, refer to Figure 1 A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening includes a data acquisition and processing module, a unified feature extraction module, an environmental state identification module, a metabolic map construction module, a causal inference and analysis module, a twin modeling prediction module, a nutrition plan generation module, an intervention simulation evaluation module, a safety joint evaluation module, a diet conversion output module, and a dynamic scheduling and updating module.

[0027] The twin modeling prediction module constructs a corresponding virtual patient model based on the patient's static multi-omics characteristics, differences in environmental response, and causal relationships between changes in various indicators. It simulates the patient's physiological evolution trend and predicts metabolic responses under different nutritional intake conditions.

[0028] Specifically, the system acquires the patient's multi-omics characteristics, environmental state parameters, and causal inference results, and maps data from different sampling frequencies onto the same time axis to unify the prediction timescale of various data. Then, following the order of "multi-omics characteristics—environmental state—causal results," the time-corresponding data are concatenated into a unified input vector. This input vector is then mapped to a low-dimensional latent space to obtain the initial hidden states of the virtual patient at different times. After obtaining the initial hidden states, a continuous-time dynamic equation for simulating the patient's physiological evolution is constructed, enabling it to output state results at any prediction time to complete the construction of the corresponding virtual patient model. Each candidate nutritional intervention plan is encoded into a structured plan vector, and each plan is then mapped to a continuous-time control signal. The signal is input to the corresponding virtual patient model, which is run under different nutritional conditions. Based on the continuous-time dynamic equation, the virtual patient state is progressively advanced, and the latent state is converted into various interpretable physiological indicators. Short-term, medium-term, and long-term evolutionary trajectories are formed to obtain the dynamic impact of different nutritional conditions on the patient's physiological state. The consistency and stability of the dynamic impact of each nutritional condition on the patient's physiological state are verified. After the verification is passed, based on the output physiological indicators and evolutionary trajectories, the deviation of different plans at each time point and each indicator is compared item by item, and the overall response score under different nutritional conditions is calculated. Then, the candidate nutritional intervention plans are arranged from high to low according to the overall response score, and the physiological indicators and corresponding evolutionary trajectories are output.

[0029] The nutrition plan generation module automatically generates multiple candidate nutrition intervention plans based on the identified simulation results and causal inference results, combined with the patient's current habituation status, metabolic load level and nutritional deficit.

[0030] Specifically, based on the causal inference results, all paths are expanded according to the "exposure node - mediator node - outcome node" link, and the node relationships, directional consistency, and strength of evidence on each link are jointly evaluated. Simultaneously, the overall support for each link is generated, and the link with the highest overall support and directional consistency is selected as the key pathway for the corresponding patient. After the key pathway is determined, the adaptation stability, volatility, and recovery capacity are extracted based on the altitude adaptation status information, and the corresponding adaptation status index is calculated. Then, the three types of load—insulin resistance, uric acid abnormalities, and bone metabolism abnormalities—are weighted and fused to obtain the current status information of the corresponding patients. Based on the current status information of the patients, the intake of protein, essential amino acids, vitamins, minerals, and corresponding functional nutrients is statistically analyzed. The system calculates the gap between the patient's current condition and the pre-set target nutritional needs, forming a nutritional gap vector. It first determines the type of nutritional support that each pathological pathway relies on, and then processes the needs based on the patient's current high status information and adaptation level to form the demand intensity of each nutrient. Based on the demand intensity of each nutrient, it performs a combination search from the pre-set nutrient library and food combination library to generate multiple candidate nutritional intervention plans. It calculates the comprehensive score of each candidate nutritional intervention plan based on the degree of demand matching, feasibility of execution, and plan discrimination. It then sorts the generated multiple candidate nutritional intervention plans from high to low according to the comprehensive score, and finally screens out multiple sets of nutritional intervention plans that are non-overlapping and cover different nutritional focuses based on the plan discrimination, and outputs multiple sets of candidate nutritional intervention plans.

[0031] The intervention simulation evaluation module uses a virtual patient model to reason and search for candidate nutritional intervention programs, simulate the effects of different dosage combinations, and screen feasible nutritional programs.

[0032] Specifically, based on the generated candidate nutritional intervention plans, the available dosage of each nutrient is divided into multiple discrete levels according to the type of nutrient. Then, the dosage levels of different nutrients are combined using Cartesian methods to generate multiple intake combinations, thus establishing a candidate intervention path space. Each candidate nutritional intervention plan in the candidate intervention path space is decomposed into continuous inputs over multiple time periods and transformed into a counterfactual control sequence that acts over time. Subsequently, the counterfactual control sequence is input into the corresponding virtual patient model to generate state trajectories under different plans. After generating the state trajectories, they are mapped to metabolic indicators and blood pressure fluctuation indicators. The study identifies renal burden and oxidative stress indicators and generates predictive observation vectors for each candidate nutritional intervention. It also calculates the benefit and safety penalty values ​​for each candidate intervention at different times, combining these values ​​to form a comprehensive value. This comprehensive value is then used for enhanced searching and filtering within the candidate intervention path space. If the benefit is greater than or equal to a preset effectiveness threshold and the safety penalty value is less than or equal to a preset safety penalty value, the corresponding candidate nutritional intervention is retained; otherwise, it is discarded. The remaining candidate nutritional interventions are then input into the corresponding virtual patient model according to their comprehensive value, from highest to lowest.

[0033] The safety joint evaluation module performs dual judgment on the simulation results. When a candidate solution meets both the preset effectiveness threshold and safety constraints, it is marked as an output nutrition solution. The diet conversion output module converts the evaluated nutrition solution into a practical and implementable lifestyle solution. The dynamic scheduling and update module combines the patient's behavioral response, intake records, location status, sleep rhythm and interactive feedback to dynamically adjust the reminder time window and push method.

Claims

1. A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening, characterized in that, It includes a unified feature extraction module, an environmental state recognition module, a metabolic map construction module, a causal inference and analysis module, a twin modeling and prediction module, a nutrition plan generation module, and an intervention simulation and evaluation module; The unified feature extraction module filters out multiple omics variables based on preprocessed multi-source data and extracts corresponding static multi-omics features. The environmental state recognition module is used to identify the patient's current environment, generate corresponding plateau environmental state parameters, and analyze the differences in the patient's environmental response under different altitudes and different living scenarios. The metabolic mapping module constructs a dynamic map of metabolic comorbidities in the elderly based on screened multi-omics features, clinical phenotypes, and high-altitude environmental parameters. The causal inference analysis module infers the causal relationship between changes in various indicators based on the dynamic map of metabolic comorbidities in the elderly, and identifies the direct and indirect effects of hypoxia exposure. The twin modeling prediction module constructs a corresponding virtual patient model based on the patient's static multi-omics characteristics, differences in environmental response, and causal relationships between changes in various indicators. It simulates the patient's physiological evolution trend and predicts metabolic responses under different nutritional intake conditions. The nutrition plan generation module automatically generates multiple candidate nutrition intervention plans based on the identified simulation results and causal inference results, combined with the patient's current habituation status, metabolic load level and nutritional deficit. The intervention simulation evaluation module uses a virtual patient model to reason and search for candidate nutritional intervention programs, simulates the effects of different dosage combinations, and screens feasible nutritional programs.

2. The personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 1, characterized in that, It also includes a data acquisition and processing module, a safety joint assessment module, a diet conversion output module, and a dynamic scheduling and update module; The data acquisition and processing module receives the patient's static and dynamic basic data, and performs missing data repair, anomaly removal, unit unification, centralization and normalization processing on the multi-source data. It also performs cross-device, cross-frequency and cross-scenario data alignment based on a unified time axis. The safety joint evaluation module performs dual judgment on the simulation results. When a candidate solution simultaneously meets the preset effectiveness threshold and safety constraints, it is marked as an output nutrition solution. The dietary conversion output module will convert the evaluated nutritional plan into a practical and feasible lifestyle plan. The dynamic scheduling and update module combines the patient's behavioral responses, intake records, location status, sleep rhythm, and interactive feedback to dynamically adjust the reminder time window and push method.

3. The personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 1, characterized in that, The environmental state recognition module identifies the patient's current environment, generates corresponding plateau environmental state parameters, and analyzes the differences in the patient's environmental response at different altitudes and in different living scenarios. The specific steps are as follows: S1.1: Unify the reading of various raw observation values, including the patient's current location coordinates, real-time altitude, meteorological parameters, oxygen partial pressure monitoring values, activity trajectory points, and historical high-altitude exposure records, and use time-attenuation weighted interpolation to map various data with different sampling frequencies onto the same time axis; S1.2: After completing time alignment, the current altitude is standardized to obtain the corresponding altitude exposure factor. Then, the current oxygen partial pressure change and oxygen partial pressure decrease factor are calculated based on the current total atmospheric pressure. After that, the corresponding meteorological load factor is calculated based on the patient's current ambient temperature, relative humidity and total atmospheric pressure to obtain the comprehensive pressure of the current environment on the body. S1.4: Based on the patient's trajectory speed and dwelling behavior at different times, and the spatial migration distance at each time, calculate the activity-residential exposure factor, and then distinguish the patient's environmental exposure patterns in different living scenarios based on the activity-residential exposure factor. S1.5: The altitude exposure factor, oxygen partial pressure decrease factor, meteorological load factor and activity and residence exposure factor are weighted and summed to obtain the patient's current high-altitude environmental exposure intensity. Then, based on the patient's continuous exposure time, exposure fluctuation and oxygen partial pressure change trend, the corresponding adaptation index is calculated. Based on the environmental exposure intensity and adaptation index, the patient's disadaptation risk value is calculated, and the corresponding risk threshold is set. S1.6: If the patient's acclimatization risk value is greater than or equal to the risk threshold, it indicates that the current exposure has exceeded the patient's tolerance boundary and the acclimatization risk is increased. Conversely, it indicates that the current environment is still within the patient's relatively tolerable range. At the same time, the average exposure intensity of the patient in the high-altitude pasture scenario and the average exposure intensity in the settled town scenario are collected in real time, and the difference between the two sets of data is obtained. Then, based on the difference at different times and the average exposure intensity in the settled town scenario, the scene response difference of the patient in different altitudes and residential scenarios is obtained, and the environmental response differences between different residential scenarios are recorded.

4. The personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 2, characterized in that, The specific steps of the metabolic mapping construction module in constructing a dynamic map of metabolic comorbidities in the elderly are as follows: S2.1: Extract the selected multi-omics features, clinical phenotypes, and plateau environmental state parameters from the original data records of omics data, clinical records, and environmental observations. Use the multi-omics features, clinical phenotypes, and plateau environmental state parameters as entities to be mapped. Map synonyms, abbreviations, and different naming methods across platforms in a unified manner to construct a standard entity set. S2.2: After completing the standardized mapping, knowledge nodes are constructed based on the standard entity set for genes, proteins, metabolites, clinical phenotypes and plateau environment state parameters, respectively. The source attributes, functional attributes and context attributes of each node are integrated into a unified feature vector, and the node features are mapped to a unified semantic space through linear encoding. S2.3: After the nodes are established, the number of evidence supports, cross-validation consistency and literature or database reliability of each node in different data sources are counted respectively. The results are normalized and weighted according to a uniform scale to obtain the initial importance of each node. From the node set, two nodes are randomly selected and potential associations are simultaneously screened from multi-omics correlation, clinical co-occurrence pattern and environmental response pattern. Then, the initial strength of the connection edge between the two corresponding nodes is calculated through multi-evidence fusion. S2.4: When the initial strength is higher than the preset threshold, retain the connection edge; otherwise, delete it. Calculate the matching score between the node pair corresponding to each connection edge and each relation semantic. After normalizing the obtained matching scores, select the relation semantic with the highest normalized matching score as the relation type of the corresponding connection edge. Then, combine the candidate edge strength and relation type coefficient to determine the initial edge weight of the corresponding connection edge to complete the construction of the dynamic map of metabolic comorbidities in the elderly. S2.5: When new omics data, clinical records and environmental observations are integrated, the entities to be built in the graph are generated again, and the existing nodes and edges in the dynamic graph of metabolic comorbidities in the elderly are incrementally updated to reflect the latest knowledge of the patient's status after the evolution over time in real time. At the same time, the change amplitude of the entire dynamic graph of metabolic comorbidities in the elderly between adjacent rounds is statistically analyzed. S2.6: If the change exceeds the preset threshold, the relation reorganization process is triggered to re-evaluate the corresponding connection edges, reduce the weight of edges lacking new evidence, or delete them. Then, the activation intensity of each disease-related subgraph in the dynamic graph of metabolic comorbidities in the elderly is calculated, and the disease subgraph with an activation intensity higher than the preset selection value is output.

5. A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 3, characterized in that, The specific steps of the causal inference analysis module inferring the causal relationship between changes in various indicators and identifying the direct and indirect effects of hypoxia exposure are as follows: S3.1: From the latest dynamic map of metabolic comorbidities in the elderly and its disease sub-maps, candidate causal pathways between high altitude hypoxia exposure, acclimatization status and metabolic comorbidity outcomes are located. In each candidate causal pathway, high altitude hypoxia exposure is the exposure node, acclimatization status is the mediator node, and metabolic comorbidity outcome is the outcome node. The three outcomes of insulin resistance, uric acid metabolism disorder and bone metabolism disorder are used as the outcome types of the outcome nodes to be analyzed in each candidate causal pathway, and the cumulative score of each candidate causal pathway is calculated. S3.2: Select upstream nodes that are connected to both the exposure node and the outcome node from the dynamic map of metabolic comorbidities in the elderly, and evaluate the pseudo-association between the corresponding variables of each upstream node and the weakening exposure and the outcome by the cumulative score of each candidate causal path. Based on the evaluation results, establish a backdoor adjustment set. After screening the backdoor adjustment set, quantitatively estimate the total causal effect of high altitude hypoxia exposure on each outcome to obtain the overall impact strength of hypoxia exposure on the three types of metabolic comorbidities, i.e., the total effect. S3.3: Validate any mediator node in each candidate causal path and calculate the validity score of each mediator node. If the validity score is higher than the preset screening value, the corresponding mediator node is included in the front door mediator set. The front door criterion is used to calculate the estimated expectation of the corresponding outcome node under the given exposure node and mediator node, i.e. the mediator transmission effect. S3.4: Based on the total effect and the mediating transmission effect, the direct and indirect effects of hypoxia exposure on each outcome are decomposed. After decomposition, the total effect, direct effect and indirect effect of the three types of outcomes are statistically analyzed, and the magnitude of the total effect, direct effect and indirect effect of the three types of outcomes are compared. The mediating proportion is used to determine whether hypoxia exposure mainly works through the direct damage pathway or mainly through the habituation pathway. The stability and reliability of the inference results are verified by using resampling confidence intervals, and the causal relationship between the changes of each indicator is output based on the verification results.

6. A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 4, characterized in that, The specific steps of the twin modeling prediction module in constructing a corresponding virtual patient model, simulating the patient's physiological evolution trend, and predicting metabolic responses under different nutritional intake conditions are as follows: S4.1: Obtain the patient's multi-omics features, environmental state parameters, and causal inference results, and map data with different sampling frequencies onto the same time axis to unify the prediction time scale of various types of data. Then, in the order of "multi-omics features - environmental state - causal results", the various types of data after time correspondence are concatenated into a unified input vector. Then, the input vector is mapped to a low-dimensional latent space to obtain the initial hidden state of the virtual patient at different times. S4.2: After obtaining the initial hidden state, construct a continuous-time dynamic equation for simulating the physiological evolution of the patient, and enable it to output the state result at any prediction time to complete the construction of the corresponding virtual patient model. Encode each candidate nutritional intervention plan into a structured plan vector, then map each plan into a continuous-time control signal, and input each continuous-time control signal into the corresponding virtual patient model. S4.3: Run the virtual patient model under different nutritional conditions, based on the continuous-time dynamic equation, gradually advance the virtual patient state, and then convert the hidden state into various interpretable physiological indicators and form short-term, medium-term and long-term evolution trajectories to obtain the dynamic impact of different nutritional conditions on the patient's physiological state. S4.4: Verify the consistency and stability of the dynamic effects of various nutritional conditions on the patient's physiological state. After the verification is passed, based on the output physiological indicators and evolution trajectories, compare the degree of deviation of different plans at each time point and each indicator, calculate the overall response score under different nutritional conditions, and then arrange each candidate nutritional intervention plan in descending order of overall response score, while outputting various physiological indicators and corresponding evolution trajectories.

7. A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 1, characterized in that, The specific steps by which the nutrition protocol generation module automatically generates multiple candidate nutrition intervention protocols are as follows: S5.1: Based on the causal inference results, all paths are expanded according to the "exposure node - mediator node - outcome node" link, and the node relationships, directional consistency and evidence strength on each link are jointly evaluated. At the same time, the comprehensive support of each link is generated, and the link with the highest comprehensive support and directional consistency is selected as the key pathway for the corresponding patient. S5.2: After the key pathways are identified, the acclimatization stability, volatility and recovery capacity are extracted based on the corresponding status information of high altitude adaptation, and the corresponding acclimatization status index is calculated. Then, the three types of loads, namely insulin resistance, abnormal uric acid and abnormal bone metabolism, are weighted and fused to obtain the corresponding current status information of the patients. S5.3: Based on the patient's current status information, the intake of protein, essential amino acids, vitamins, minerals and corresponding functional nutrients are statistically analyzed. The gap between these intakes and the preset target nutritional needs is calculated item by item to form a nutritional gap vector. First, the type of nutritional support that each pathological pathway relies on is determined. Then, the needs are processed in combination with the patient's current high status information and habituation level to form the demand intensity of each nutrient. S5.4: Based on the intensity of demand for each nutrient, perform a combination search from the preset nutrient library and food combination library to generate multiple candidate nutritional intervention plans. Calculate the comprehensive score of each candidate nutritional intervention plan based on demand matching degree, feasibility, and plan differentiation. Sort the generated multiple candidate nutritional intervention plans in descending order of comprehensive score. Then, combine the plan differentiation to screen out multiple sets of nutritional intervention plans that do not overlap with each other and cover different nutritional focuses, and output multiple sets of candidate nutritional intervention plans.

8. A personalized nutritional intervention system for metabolic comorbidities in the elderly based on multi-omics screening according to claim 6, characterized in that, The intervention simulation evaluation module simulates the effects of different dosage combinations, and the specific steps for screening feasible nutritional regimens are as follows: S6.1: Based on the generated candidate nutritional intervention plans, the available dose of each nutrient is divided into multiple discrete levels according to the type of nutrient. Then, the dose levels of different nutrients are combined in a Cartesian combination to generate multiple intake combinations in order to establish a candidate intervention path space. S6.2: Decompose each group of candidate nutritional intervention schemes in the candidate intervention path space into continuous inputs over multiple time periods, and transform them into counterfactual control sequences that act over time. Then, input the counterfactual control sequences into the corresponding virtual patient model to generate state trajectories under different schemes. S6.3: After generating the state trajectory, map it to metabolic indicators, blood pressure fluctuation indicators, kidney burden indicators and oxidative stress indicators, and generate the prediction observation vector of each candidate nutritional intervention program. At the same time, calculate the benefit and safety penalty value of each candidate nutritional intervention program at different times. S6.4: Combine the benefits and safety penalty values ​​of each candidate nutritional intervention plan at different times to form its corresponding comprehensive plan value, and perform enhanced search and screening in the candidate intervention path space. If the benefit is greater than or equal to the preset effectiveness threshold and the safety penalty value is less than or equal to the preset safety penalty value, the corresponding candidate nutritional intervention plan is retained; otherwise, it is removed. The remaining candidate nutritional intervention plans are then input into the corresponding virtual patient model in descending order of comprehensive plan value.

Images