Individual aging evaluation and prediction system and method based on multi-dimensional health data and artificial intelligence

Abstract

The invention discloses an individual aging evaluation and prediction system and method based on multi-dimensional health data and artificial intelligence, and belongs to the technical field of intelligent medical treatment, health big data and artificial intelligence, and the method comprises the following steps: collecting the multi-dimensional health data of a user, and processing the multi-dimensional health data to generate a standardized individual health portrait; based on the personal health portrait, utilizing a trained senescence evaluation model to calculate senescence quantitative indexes of the user in a plurality of physiological dimensions; multi-source data are processed by adopting a standardization method based on a Chinese population health norm, and a multi-dimensional aging score is calculated through a multi-task deep learning model trained by a Chinese population queue, so that population specificity and biological rationality of model input and output are ensured; the norm standardization converts the original data of the individual into the deviation degree relative to the healthy age group, and provides a standard with clear clinical significance for the model.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare, health big data and artificial intelligence, and in particular to an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence. Background Technology

[0002] With the accelerating aging of society and the general improvement of public health management awareness, how to effectively delay the physiological aging process and maintain a high level of health has become one of the core challenges in the fields of life sciences and preventive medicine. Traditional health assessments mainly rely on a simple comparison between a single or a few clinical indicators and an individual's calendar age, which makes it difficult to fully reflect the heterogeneous nature of the coordinated degeneration of multiple physiological dimensions in the complex human body system. The limitations of this assessment method lead to biases in the judgment of an individual's aging status and make it impossible to provide accurate guidance for personalized intervention.

[0003] To address these challenges, existing technologies have begun to incorporate artificial intelligence and big data analytics. Some solutions attempt to utilize machine learning models, such as biological age prediction models based on linear regression or simple neural networks, to output a single score for accelerated or decelerated aging by integrating blood biochemical indicators or genomic data. Other solutions focus on using dynamic monitoring data from wearable devices to assess the functional status of specific organ systems. These data-driven methods have, to some extent, improved the objectivity of aging assessments.

[0004] However, these existing technological solutions still have several inherent drawbacks. First, most of these models are black-box or gray-box models, and their predictions lack interpretability. Users can only obtain a conclusion such as "your biological age is two years older than your calendar age," but cannot know whether the cardiovascular, metabolic, or immune systems are causing this accelerated aging. Therefore, it is difficult to translate them into targeted interventions. Second, most existing technologies separate assessment from intervention. They can perform assessments, but the intervention recommendations they provide are often general and static, such as suggesting that all users increase exercise or improve their diet. They lack the ability to dynamically simulate and extrapolate based on individual physiological states. Users cannot predict the potential impact of a specific nutritional supplement, exercise program, or drug intervention on their long-term health trajectory, which seriously affects the accuracy of intervention measures and users' long-term adherence. Finally, many models are built based on cohort data from Western populations, and their norms and algorithms may not be applicable to the unique genetic background, lifestyle, and disease spectrum characteristics of the Chinese population. This further limits their accuracy and reliability in the Chinese market. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence, comprising the following steps:

[0007] S1. Collect users’ multi-dimensional health data, process it, and generate a standardized personal health profile;

[0008] S2. Based on the personal health profile, use a trained aging assessment model to calculate the user's quantitative aging indicators across multiple physiological dimensions.

[0009] S3. Based on the aforementioned multi-dimensional aging quantification indicators, analyze the interaction between various physiological dimensions to generate a comprehensive health age index;

[0010] S4. Based on the personal health profile and aging quantification indicators, calibrate the parameters of the general physiological model and construct the user's personalized digital twin;

[0011] S5. Simulate the effects of different intervention combinations on the personalized digital twin, and generate personalized solutions and prediction reports based on the optimization objectives;

[0012] S6. Based on the effects of the intervention combination, update the aging quantification indicators and the healthy age index, and dynamically calibrate the digital twin.

[0013] In a preferred embodiment of the present invention, in step S1, the standardization refers to processing using a Z-score standardization method based on a health norm database of a specific population; the specific population is the Chinese population, and the norm database is the Chinese adult health norm database.

[0014] In a preferred embodiment of the present invention, in step S2, the aging assessment model is a multi-task deep learning model, which adopts a fully connected neural network structure with a shared bottom layer and independent task heads; the shared bottom layer includes three fully connected layers, with the number of neurons in each layer being 400-600, 200-300, and 100-150, respectively, and the activation function being ReLU.

[0015] In a preferred embodiment of the present invention, in step S4, the calibration of the general physiological model parameters is achieved by solving an optimization problem, the objective function of which is to minimize the Euclidean distance between the aging score vector and the steady-state output vector of the model.

[0016] In a preferred embodiment of the present invention, in step S5, the generation of personalized schemes based on optimization objectives adopts a multi-objective optimization algorithm based on genetic algorithms.

[0017] In a preferred embodiment of the present invention, in step S5, the multi-objective optimization algorithm is a multi-objective optimization algorithm based on genetic algorithm, and the optimization objectives include at least maximizing the magnitude of health age reversal, maximizing the priority matching degree of key physiological dimension improvement, and minimizing the complexity of the intervention plan.

[0018] In a preferred embodiment of the present invention, in step S6, the dynamically calibrated digital twin employs an online learning algorithm based on Kalman filtering.

[0019] In a preferred embodiment of the present invention, it includes:

[0020] The data acquisition and preprocessing module is used to perform step S1;

[0021] The aging score calculation module is used to execute step S2;

[0022] The aging atlas and healthy age generation module is used to perform step S3;

[0023] A digital twin building module is used to perform step S4;

[0024] The intervention plan simulation optimization module is used to execute step S5;

[0025] The dynamic monitoring and update module is used to execute step S6.

[0026] In a preferred embodiment of the present invention, the multi-task deep learning model is a fully connected neural network structure with a shared bottom layer and independent task heads. Its loss function is the weighted sum of prediction errors of each dimension, and the weights are set according to the impact of functional decline in each physiological dimension on the risk of all-cause mortality.

[0027] In a preferred embodiment of the present invention, the physiological system dynamics model is a modular model described by a set of ordinary differential equations, whose state variables cover at least six dimensions, including cardiovascular function, metabolic homeostasis, immune inflammation regulation, neurocognition, liver function, kidney function, musculoskeletal function, and endocrine function.

[0028] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0029] (1) By using a standardized method based on the health norms of the Chinese population to process multi-source data, and by calculating multi-dimensional aging scores through a multi-task deep learning model trained on a Chinese population cohort, the population specificity and biological rationality of the model input and output are ensured. Norm standardization transforms individual raw data into deviations relative to healthy peers, providing the model with a benchmark with clear clinical significance. Models trained on specific populations can accurately learn the aging trajectory of that population. The resulting aging scores truly reflect the user's relative position in the Chinese population spectrum. The scores themselves have intuitive interpretability. Compared with the assessment bias caused by directly applying Western population models, this method fundamentally solves the adaptation problem of the model in the Chinese population, providing doctors and users with clear and reliable quantitative reports on the functional decline of various systems.

[0030] (2) By combining graph attention network and causal inference method to construct personalized directed aging graph, and generating health age index based on it, a leap from static scoring to dynamic causal relationship is realized. The graph network learns the complex nonlinear relationship between dimensions, while the causal discovery algorithm gives the relationship directionality, thereby revealing key aging pathways such as metabolic disorder driving inflammation. The health age index is a weighted fusion of multi-dimensional information, which can accurately identify the root driving factors and secondary accompanying phenomena in the user's aging network. Compared with the existing black box model that only gives a single, general biological age or disordered list of correlations, this method provides an interpretation of the aging mechanism with causal logic, so that the intervention strategy can be targeted and prioritized to regulate key driving nodes, thereby obtaining the maximum health benefits at the lowest cost and realizing precise navigation of intervention.

[0031] (3) By constructing a personalized digital twin based on physiological mechanisms and performing dynamic simulation and closed-loop optimization of intervention programs, this technology deeply integrates assessment, prediction and intervention into a self-evolving system. The digital twin simulates the unique physiological dynamics of an individual by calibrating the parameters of a general physiological model, thereby enabling safe pre-simulation of the long-term effects of different intervention measures in virtual space. Through continuous iteration of the model based on feedback from real-world data, it can provide users with a highly personalized optimal combination of intervention programs that has been verified by simulation and make predictions on future health trajectories. Compared with the static, general suggestions provided by existing technologies and their drawbacks of being disconnected from effect evaluation, this method achieves a qualitative leap from general suggestions to personalized prescriptions, greatly improving users' trust and compliance with intervention programs. At the same time, the system has self-learning capabilities and can adaptively adjust according to changes in user status, making health management a continuously optimized and precise closed-loop process, ultimately ensuring the long-term effective improvement of key physiological indicators. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of a preferred embodiment of the present invention, showing an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence.

[0034] Figure 2 This is a main flowchart of the anti-aging prediction and intervention scheme generation of an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence, according to a preferred embodiment of the present invention.

[0035] Figure 3 This is a flowchart of the personalized digital twin construction and iterative update process of an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence, according to a preferred embodiment of the present invention. Detailed Implementation

[0036] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Application Overview:

[0038] This invention addresses the technical problems of poor interpretability, weak intervention guidance, and lack of personalized simulation capabilities in existing aging assessment models. It constructs a collaborative technology chain integrating multi-source data fusion, multi-dimensional aging atlas generation, and digital twin simulation. Utilizing a multi-task deep network based on a Chinese population cohort, it calculates multi-dimensional aging scores to achieve an interpretable panoramic depiction of aging status. Using this atlas as an initial state, a personalized digital twin is constructed, enabling dynamic simulation and effect preview of multimodal intervention programs in virtual space. This approach deeply couples assessment and intervention, significantly improving the accuracy of healthy age prediction and generating targeted personalized intervention strategies. Ultimately, it effectively improves user compliance and the improvement rate of key physiological indicators.

[0039] The technological breakthrough of this invention lies in addressing the core issues of poor interpretability of aging assessment and weak targeting of intervention programs in personalized health management scenarios by constructing a multidimensional aging map based on the characteristics of the Chinese population and using personalized digital twin dynamic simulation technology. Compared with existing black-box prediction models, this solution combines interpretable aging state quantification with dynamic inference based on physiological mechanisms, which significantly improves the accuracy of intervention strategies and user compliance.

[0040] Example 1:

[0041] Exemplary method:

[0042] like Figure 1 and Figure 2 As shown, an individual aging assessment and prediction system and method based on multi-dimensional health data and artificial intelligence includes the following steps:

[0043] 1. A method for anti-aging prediction based on multi-dimensional health data and artificial intelligence, characterized by comprising the following steps:

[0044] S1. Collect users’ multi-dimensional health data, process it, and generate a standardized personal health profile;

[0045] S2. Based on the personal health profile, use a trained aging assessment model to calculate the user's quantitative aging indicators across multiple physiological dimensions.

[0046] S3. Based on the aforementioned multi-dimensional aging quantification indicators, analyze the interaction between various physiological dimensions to generate a comprehensive health age index;

[0047] S4. Based on the personal health profile and aging quantification indicators, calibrate the parameters of the general physiological model and construct the user's personalized digital twin;

[0048] S5. Simulate the effects of different intervention combinations on the personalized digital twin, and generate personalized solutions and prediction reports based on the optimization objectives;

[0049] S6. Based on the effects of the intervention combination, update the aging quantification indicators and the healthy age index, and dynamically calibrate the digital twin.

[0050] In transforming the anti-aging prediction concept based on multi-dimensional health data and artificial intelligence into a feasible solution, this invention faces a series of profound technical challenges. The key technical gap lies in how to transform multi-source heterogeneous health data into a standardized health profile specific to the Chinese population. This requires the establishment of a high-quality Chinese adult health norm database and the development of a robust preprocessing process that can handle missing and outlier values ​​to overcome the adaptation bias of Western models when applied to the Chinese population.

[0051] The multi-dimensional health data collected in step S1 refers to a set of heterogeneous health indicators from different physiological systems and sources. The sequence of operations for cleaning, transforming and integrating the original health data includes missing value imputation, outlier handling and standardization. The processed data is then transformed into a standardized personal health profile output through an algorithm. The standardized personal health profile is a structured data vector after Z-score standardization, representing the deviation of an individual's health status from that of their healthy peers in China.

[0052] S1. Identification and Collection of Multi-Dimensional Health Data Sources: The goal of this step is to transform heterogeneous raw health data from different sources into a clean, standardized, and directly processable structured data vector, i.e., a personal health profile.

[0053] The data collected in this step refers to a set of health indicators from different physiological systems and sources. Based on extensive epidemiological cohort studies and clinical evidence, indicators that have been proven to be significantly associated with physiological aging and all-cause mortality risk were selected, specifically including:

[0054] Clinical test indicators include C-reactive protein, glycated hemoglobin, total cholesterol / high-density lipoprotein ratio, and white blood cell count. These indicators were chosen because they are easy to obtain through routine physical examinations, cost-effective, and numerous studies have confirmed their predictive value for healthy aging in the Chinese population.

[0055] It should be noted that C-reactive protein is a core marker of systemic chronic inflammation, which is an accelerator of aging, and its level is directly related to the risk of cardiovascular disease, cognitive decline, and other risks. Glycated hemoglobin reflects long-term blood glucose control and metabolic homeostasis, and elevated levels predict the risk of diabetes and the resulting multi-organ damage. The total cholesterol / high-density lipoprotein ratio is a key indicator for assessing cardiovascular disease risk and reflects blood lipid balance better than a single indicator. White blood cell count reflects the overall activity and balance of the immune system, and abnormal counts suggest immune aging or a chronic inflammatory state.

[0056] The genomic and epigenetic data, not whole-genome sequencing, are selectively chosen from sites closely related to the molecular mechanisms of aging to reduce computational complexity and improve model efficiency. These include: DNA methylation clock sites, as methylation levels are currently the most accurate biological marker of aging; telomere length-related single nucleotide polymorphisms, as telomere shortening is a marker of cellular senescence; and gene sites associated with inflammatory senescence. These data provide the molecular basis of aging.

[0057] It should be noted that DNA methylation clock loci are currently the most accurate molecular markers of aging. By analyzing the methylation level of specific loci, epigenetic age can be calculated, and the difference between this and calendar age can strongly predict the decline of physiological function and lifespan. Telomere length-related gene loci are markers of telomere shortening, which is a sign of cellular replicative aging. These gene loci can assess an individual's genetic predisposition to cellular aging. Inflammatory aging-related gene loci identify whether an individual carries a genetic risk that leads to a high inflammatory response, enabling early warning.

[0058] Quantitative features extracted from standard medical images provide objective organ-level data, including: carotid intima-media thickness, extracted from carotid ultrasound images, which is a direct indicator of subclinical atherosclerosis; liver fat fraction, calculated from abdominal ultrasound or MRI images, reflecting the degree of non-alcoholic fatty liver disease; hippocampal volume, obtained through brain MRI image segmentation, which is associated with neurocognitive decline; and bone mineral density T-score, obtained through DXA bone mineral density scanning, which is a core indicator for diagnosing osteoporosis.

[0059] It should be noted that, among these, carotid intima-media thickness is the earliest direct indicator of atherosclerosis, appearing before clinical symptoms and serving as a warning window for cardiovascular health; liver fat fraction is a quantitative diagnostic indicator of non-alcoholic fatty liver disease, which is not only a liver disease but also a manifestation of systemic metabolic disorders; hippocampal volume is a key indicator of the brain's memory center, and its atrophy is an early and sensitive marker of neurocognitive decline; and bone mineral density T-score directly assesses bone strength and is a core basis for diagnosing osteoporosis and predicting fracture risk.

[0060] The missing values ​​in the data above are first handled using the chain equation multiple imputation method. This is a robust statistical method used to handle random or non-random missing values ​​in data. Its principle is to build a predictive model for each variable containing missing values ​​and use other complete variables in the dataset to predict the missing values.

[0061] The specific implementation involves 20 rounds of iterative imputation: In the first round, all missing values ​​are temporarily filled using the mean or median of other variables; then, based on the imputed complete dataset, a predictive model is built for the first missing variable, and this model is used to predict its missing values; next, a model is built for the second missing variable using the updated data, and this process is repeated. This process is repeated 20 times, allowing the imputed values ​​to gradually stabilize. This method preserves the uncertainty of the data and is more scientific than simple deletion or mean imputation.

[0062] Outlier handling employs Tukey's fences criterion. This is a non-parametric outlier identification method based on data distribution, insensitive to skewed distributions, and particularly suitable for medical data. For each continuous variable, the following steps are performed:

[0063] Calculate the first quartile of the variable. and the third and fourth quartiles ,in It is the 25th percentile. It is the 75th percentile.

[0064] Calculate interquartile range .

[0065] Define the normal value range as .

[0066] Anything less than the lower limit or greater than Values ​​that are considered outliers are typically caused by measurement errors or data entry mistakes. Instead of deleting them, they are replaced with appropriate boundary values ​​to preserve the sample size as much as possible.

[0067] It should be explained in detail that generating a standardized personal health profile refers to the process of transforming processed data into a standardized vector with clinical comparability. Its core is Z-score standardization based on the Chinese Adult Health Norm Database (2020 version). This database contains the mean and standard deviation of various indicators of a large number of rigorously screened healthy Chinese people under different age and gender groups.

[0068] A detailed explanation of the standardized formula:

[0069] :

[0070] These are the original measurements.

[0071] And the mean of this indicator in the Chinese Adult Health Norms Database, representing the specific age and gender group to which the user belongs.

[0072] Standard deviation of corresponding groups in the norm database

[0073] It should be noted that, for example, if a user's CRP is 3.2 mg / L, and their age is 50-54 years, the mean CRP in the Chinese Adult Health Norm Database group is 1.5 mg / L, and the standard deviation of CRP in the corresponding group in the norm database is 0.8 mg / L. The calculation formula is as follows:

[0074] :

[0075] this The Z-score indicates that the user's CRP level is 2.125 standard deviations higher than the average level of healthy individuals of the same age and sex. This is a clinically significant deviation that directly quantifies the degree of abnormality in their inflammatory state. Compared to using the raw value, the Z-score makes indicators of different dimensions and ranges comparable.

[0076] For time series signals, time-domain and frequency-domain features need to be extracted using a sliding window. For categorical variables, one-hot encoding is performed. Finally, the Z-scores of all continuous variables, the time-series features, and the encoded values ​​of the categorical variables are concatenated into a fixed-length string. Dimensional Personal Health Data Vector .

[0077] It should be noted that the construction of health profiles relies on comparing individual data with Chinese health norms. Using only static data from a single test cannot reflect dynamic physiological rhythms and adaptability. Furthermore, the Chinese population may possess specific lifestyle and physiological rhythms, and this dynamic information is crucial for aging assessment.

[0078] Furthermore, temporal features are quantitative indicators extracted from time-series data, such as heart rate and activity levels continuously monitored by wearable devices. For example, calculating the mean and standard deviation of heart rate every 5 minutes, or discovering periodic patterns in activity through frequency domain analysis. These features transform continuous dynamic images into quantifiable indicators, enabling health profiles to include dynamic functional information from systems such as cardiovascular regulation and neural activity. This is indispensable for accurately assessing the decline in physiological function in the Chinese population under real-life scenarios, avoiding the biases that may arise from relying solely on static snapshots.

[0079] Furthermore, directly inputting these text categories into the model would render the model unable to understand their meaning. Therefore, encoding techniques must be used to convert qualitative descriptions into numerical forms that the model can process. One-hot encoding creates an independent binary vector for each category, thus representing each option fairly and unambiguously. Only then can this categorical information, reflecting the unique lifestyle and cultural background of the Chinese population, be seamlessly integrated into the health profile and participate in the calculation along with Z-score-standardized continuous variables such as blood indicators, ensuring that the final assessment results truly match the user's background characteristics.

[0080] Step S2 involves using a trained aging assessment model for inference. The core of this process lies in the model's ability to predict aging trajectories learned from historical data. It employs a multi-task deep learning model trained on Chinese population cohort data. Through supervised learning, it fits aging assessment models with aging score labels across various physiological dimensions, calculates the mathematical process of model inference, and inputs the health profile. Mapped to an aging score vector, the resulting quantitative aging index is a standardized score value.

[0081] S2. Preprocessing and Standardized Vector Generation of Multi-Source Heterogeneous Data: The core of this step is to use a specially trained model to generate health profiles. It is mapped to a set of quantitative scores that reflect the degree of aging of different physiological systems.

[0082] This model is a multi-task deep learning model. Its training means that the model parameters have been optimized and determined through supervised learning using longitudinal follow-up data from a large-scale Chinese population cohort.

[0083] The model structure employs a fully connected neural network architecture with a shared bottom layer and independent task heads. The shared bottom layer consists of three fully connected layers with 512, 256, and 128 neurons per layer, respectively, using the ReLU activation function. Its function is to learn the high-level feature representations common to all physiological dimensions. Each independent task head is a single-layer fully connected network responsible for mapping the features output from the shared bottom layer to the aging score of the corresponding physiological dimension. There are a total of eight task heads, corresponding to eight preset physiological dimensions.

[0084] The supervisory signal for building the training label model is not subjectively set, but rather based on objective risk calculations using cohort data to ensure the biological significance of the scores. Specifically, for each individual in the cohort... Age Real-time health data to track their future Did any major health events related to this physiological dimension occur during the year? For each physiological dimension... Select a set of core clinical indicators Then, using the Cox proportional hazards model, with age... and indicator set As a covariate, it fits the risk of health events occurring in this dimension. Individual In dimensions Aging rating tags It is defined as the number of additional risk years for an individual relative to healthy people of the same age.

[0085] The specific calculation is as follows: = This individual in dimension The equivalent age The equivalent age can be obtained by back-calculating the individual's calculated risk score into a benchmark Cox model with age as the only covariate.

[0086] :

[0087] N: Number of training samples.

[0088] ;individual In dimensions The real label on it.

[0089] The model is for individuals In dimensions The predicted score.

[0090] ;No. The weights for each dimension are determined based on the independent contribution of functional decline in each physiological dimension to the risk of all-cause mortality. The data are derived from a meta-analysis of large-scale epidemiological studies.

[0091] Calculate quantitative indicators of aging across multiple physiological dimensions for users: After model training is complete, for new users, generate their health profile vectors. The input model shares a common underlying layer for feature extraction, and then each task head performs parallel computation, ultimately outputting an 8-dimensional aging score vector. Each Both are Z-score values, representing the degree of aging of the user relative to healthy peers in the corresponding physiological dimension. Positive values ​​indicate accelerated aging, while negative values ​​indicate delayed aging.

[0092] Step S3 involves analyzing the interaction relationships between physiological dimensions using graph attention networks and causal inference algorithms. This analysis establishes a directed causal influence network between these dimensions and generates a weighted formula to calculate the health age index. The output behavior yields a single numerical index, representing the user's health age index based on the equivalent age of multi-dimensional physiological functions.

[0093] S3. Generation of Personal Multidimensional Aging Map and Healthy Age Index: This step aims to improve the interpretability of the model, moving from static scoring to dynamic causal relationships.

[0094] The analysis here is not a simple calculation of correlation coefficients, but rather a construction of a directed, weighted graph of aging interactions using a constrained causal discovery method guided by graph attention networks (GAT).

[0095] The principle is to first use GAT to learn the prior knowledge of the correlation strength between dimensions, and then use it as a constraint to integrate into the learning of the directed acyclic graph (DAG) structure based on the NOTEARS framework, so as to obtain a personalized network that conforms to both the statistical characteristics of the data and the causal logic.

[0096] The graph construction was scored across eight dimensions. For graph nodes, GAT learns the association strength by calculating the attention coefficients between nodes, generating an association prior matrix. Then The weighted adjacency matrix is ​​added as a penalty weight term to NOTEARS's optimization objective function and solved under DAG constraints. . non-zero elements , Size and sign represent the dimension To dimension The direction and intensity of causal influence.

[0097] Health Age Index Calculation The formula for calculation using the weighted formula is as follows:

[0098] :

[0099] in, For the user's calendar age, For the first Aging score across multiple dimensions, For the first The weights of each dimension are not arbitrarily set, but are derived from the independent contribution of each dimension score to mortality risk in the multivariate Cox proportional hazards model, and are normalized.

[0100] This makes It has a clear biological interpretation: its value reflects the user's equivalent age based on current multidimensional physiological functions.

[0101] Understandably, for example, a person who is 50 years old on a calendar date has an H=53 years old, which means that their overall physiological function is equivalent to the average level of a 53-year-old.

[0102] In step S4, the parameters of the general physiological model are adjusted using a calibration optimization algorithm. This makes its output match the user's actual aging score vector. The matching process is described by a set of ordinary differential equations. It simulates the dynamic interaction of multiple physiological dimensions to form a general physiological model. The process of constructing an instantiated digital twin transforms the calibrated model into a virtual representative of the user. The calibrated physiological model instance simulates the user's specific physiological state, such as the homeostatic response and the intervention effect, into a personalized digital twin.

[0103] S4. Personalized Digital Twin Construction and Calibration: This step introduces a parameterized general physiological system dynamics model (GPSDM) and uses an optimization algorithm to personalize its parameters in order to generate a digital twin for the user.

[0104] In one specific embodiment of the present invention, the model can be simplified to a compartment model described by a system of ordinary differential equations, with each physiological dimension corresponding to a state variable. This represents the dimension over time. Its functional level. Its dynamic equations are described as follows:

[0105] :

[0106] in:

[0107] Representative dimension The natural decline of function This represents the decay rate.

[0108] Representing other dimensions Dimension Synergistic or antagonistic effects, These are the influence coefficients. This coefficient matrix can be compared with the aging interaction map obtained in step S3. Related.

[0109] This represents the promoting effect of positive external factors.

[0110] This represents the damaging effects of negative external factors.

[0111] The goal of calibration is to find a set of personalized parameters. This ensures that the model's output in steady state is... , and the user's current true aging score vector As close as possible. This is achieved by solving the following optimization problem:

[0112] :

[0113] in, This is a regularization term used to prevent overfitting. It can be solved using quasi-Newton methods or gradient descent.

[0114] Parameters are adjusted iteratively using the gradient descent method. This makes the steady-state score of the model output... As close to reality as possible This is equivalent to finding the model parameters that best simulate the unique physiological dynamics of each user.

[0115] calibrated model parameters By substituting into a general physiological model, a personalized digital twin of the user is obtained. This model instance can simulate the user's steady-state response under specific physiological conditions and their dynamic response to various interventions.

[0116] In step S5, the dynamic deduction process of the intervention combination effect is simulated on the digital twin. A multimodal measure set, such as dietary adjustments combined with exercise, is used. The effect of this intervention combination is quantified through model parameter adjustments, and the algorithm is guided by a multi-objective function, such as maximizing... Simultaneously, based on the optimization objective, the process of generating automated personalized solutions and predictive reports minimizes complexity.

[0117] S5. Simulation and optimization of multimodal intervention schemes: This step involves conducting virtual clinical trials on a digital twin.

[0118] The intervention knowledge base contains hundreds of quantified interventions, each corresponding to a specific adjustment of one or more parameters in the GPSDM.

[0119] The system starts with the currently calibrated digital twin, applies one or more interventions, and then runs GPSDM to simulate the physiological trajectory over a future period. The optimization goal is to find the optimal combination of interventions. To maximize the reversal of healthy age Simultaneously satisfying other constraints, this is formulated as a multi-objective optimization problem, which can be solved using evolutionary algorithms such as NSGA-II. The final output is a set of Pareto optimal solutions and a detailed simulation prediction report.

[0120] In step S6, the intervention feedback data is used as the basis for updating, emphasizing data-driven adaptability. The process of updating and revising the aging score and twin parameters is carried out. The dynamic calibration of the twin parameters is adjusted in real time through online learning algorithms to achieve continuous optimization and iterative learning to continuously improve the prediction and intervention effects.

[0121] S6. Dynamic Monitoring and Iterative Model Updates: This step enables the system's self-evolution. After a user performs a recommended intervention, the system periodically collects new data, repeats steps S1-S3, and obtains new observation scores. At the same time, the intervention plan from the previous cycle will be implemented. Input a digital twin, run a simulation of the same duration, and obtain the predicted score. Calculate the prediction error This error is used to update the digital twin parameters through an online learning algorithm. in This is Kalman gain. This process enables the digital twin to adaptively adjust to the user's real physiological changes, becoming more accurate with use.

[0122] Example 2:

[0123] Based on Example 1, this embodiment refines and strengthens the specific method of map construction in the core step S3, aiming to improve the accuracy, stability and interpretability of identifying key driving paths of user aging.

[0124] This optimized embodiment employs a constrained causal discovery method guided by graph attention networks. Its core principle lies in not directly using the attention weights learned by GAT as the final dimensional association network, but rather treating them as prior knowledge and combining them with a causal structure learning algorithm based on a scoring function. By optimizing under the hard constraint that the graph structure must satisfy directed acyclicity, and utilizing GAT priors to guide the search, a more robust and personalized aging causal network that conforms to both statistical data characteristics and causal logic is ultimately obtained.

[0125] The implementation of this refinement includes two key stages: association prior learning and causal structure optimization.

[0126] Association prior learning based on multi-head GAT:

[0127] Input: The 8-dimensional aging score vector obtained in step S2 Each dimension is treated as a graph node.

[0128] Graph Construction: Constructing a fully connected undirected graph Each node For each physiological dimension, the initial features of a node constitute its standardized aging score. .

[0129] Multi-head attention calculation: using Each attention head is used to capture correlation patterns in different subspaces. For the first... Size, node and Attention coefficient between The calculation is as follows:

[0130] :

[0131] in:

[0132] It is the first Each node has a unique learnable weight matrix used to perform linear transformations on node features. It is the transformed feature dimension.

[0133] This indicates a vector concatenation operation.

[0134] It is the first A unique attention vector for each size.

[0135] LeakyReLU is a non-linear activation function that allows tiny negative values ​​to pass through.

[0136] Subsequently, for the nodes Calculated from all neighbors Normalization is performed using the softmax function to obtain standardized attention weights. :

[0137] :

[0138] Quantified in the head From the perspective of nodes For nodes Importance

[0139] Node feature update and prior matrix generation: Each node New features It is obtained by weighted sum of the features of its neighboring nodes:

[0140] :

[0141] Will The output features of each node are concatenated to obtain the node. The final expression Subsequently, a preliminary prior matrix of association strength is calculated. Its elements Define as a node and Cosine similarity between the final feature representations:

[0142] :

[0143] It is a symmetric matrix that reflects the strength of data-driven inter-dimensional correlations.

[0144] Causal structure optimization based on the NOTEARS framework:

[0145] Optimization goal: to As prior information, it is incorporated into a NOTEARS-based DAG learning framework. The goal is to find a weighted adjacency matrix. This matrix corresponds to a directed acyclic graph and aims to fit the data as well as possible while remaining consistent with prior knowledge. The optimization problem is formalized as follows:

[0146] :

[0147] The constraints are:

[0148] :

[0149] It is by A matrix formed by stacking the score vectors of each sample.

[0150] It is the Frobenius norm, and the data fitting term encourages... The data can be well explained by linear structure equations. .

[0151] It is an L1 norm, and the sparsity constraint term forces Many elements in the graph are zero, resulting in a concise cause-effect graph.

[0152] The key innovation in representing the Hadamard product is the prior guiding term: it represents the prior matrix. As a penalty weight. If It's very large, so the corresponding position The L1 penalty weight is large, so the optimizer is more inclined not to apply it. Setting the edge to zero means retaining it; conversely, setting it to zero makes it easier to sparsify. This effectively introduces the association information learned by GAT as a constraint into the causal discovery process.

[0153] It is a directed acyclic graph constraint. This is a continuous, differentiable algebraic constraint that ensures The graph it represents does not contain cycles, which is a basic requirement for causal relationships.

[0154] Parameter determination: hyperparameters and Controlling sparsity and the strength of prior influences. The optimal value is determined by combining grid search with model fit on the validation set. The optimization problem can be solved using a gradient-based algorithm. The final result... This is a directed, weighted matrix of causal influence strength between dimensions, whose non-zero elements... Size and sign represent the dimension To dimension The direction and intensity of causal influence.

[0155] The improvements and solutions: This refined approach significantly enhances the quality of aging profile construction and addresses two core issues that may exist in the basic method of Example 1:

[0156] Methods based on independence, such as LINGAM, may be unstable when the sample size is limited or the noise level is high. GAT, on the other hand, learns correlation rather than causality. This method combines GAT's powerful association pattern learning capabilities with NOTEARS's strict directed acyclic graph constraint. GAT provides prior knowledge... This effectively narrows the search space for causal structures, guiding the optimizer to search for DAGs that better match the strong correlation patterns in the data, thus obtaining more stable and reasonable causal structures with limited data. This is especially true when dealing with highly correlated physiological dimensions, enabling more reliable identification of causal directions.

[0157] So that the final result is It is a clear and concise weighted directed adjacency matrix that explicitly indicates which dimension influences which dimension and the extent of that influence. This is more interpretable than a simple undirected graph or a complex attention coefficient matrix. Furthermore, the system can more accurately identify key driving nodes and vulnerable nodes in the aging network, thus providing more precise targets for the intervention optimization in step S5.

[0158] Example 3:

[0159] Exemplary system:

[0160] like Figure 3 As shown, the system materializes the six core steps of the aforementioned methodology into six collaborative functional modules. Through standardized application programming interfaces and data buses, they exchange data and control commands, forming a closed-loop health management platform that integrates data acquisition, intelligent analysis, simulation, and dynamic optimization.

[0161] Data Acquisition and Preprocessing Module: This module serves as the system's data entry point, possessing the capability to access and process multi-source heterogeneous data. Its hardware infrastructure includes secure data interfaces for interfacing with hospital laboratory information systems, image archiving, and communication systems, used to acquire clinical testing indicators and medical image features.

[0162] It integrates a data parsing unit for high-throughput gene sequencers to process genomic and epigenetic data; and provides wireless communication protocol stacks such as Bluetooth and Wi-Fi to connect to and receive dynamic physiological signals generated by various medical-grade wearable devices.

[0163] In addition, this module provides a user interface for entering lifestyle and psychological state questionnaire data. At the software level, the module has a powerful preprocessing pipeline built-in for data cleaning, outlier handling, and multiple chain equation imputation.

[0164] Step S1 involves using the Z-score normalization algorithm based on the "Chinese Adult Health Norm Database" to transform all continuous raw data into deviations relative to healthy individuals of the same age and sex in China. This process ultimately generates a high-dimensional, structured personal health data vector with clear biological significance, laying a solid foundation for subsequent analysis.

[0165] Aging Score Calculation Module: This module is one of the core computing engines of the system. It carries a multi-task deep learning model trained on large-scale Chinese population cohort data. The model adopts a fully connected neural network structure with a shared underlying layer and independent task heads.

[0166] The shared bottom layer consists of multiple hidden layers, used to learn high-level feature representations common to all physiological dimensions; each independent task head is a shallow network specifically responsible for predicting a standardized aging score for a specific physiological dimension.

[0167] The loss function of this module is designed as a weighted sum of the prediction errors of each dimension. The weights are not set arbitrarily, but are set according to the independent contribution of functional decline in each physiological dimension to the risk of all-cause mortality, thereby ensuring the clinical significance and predictive power of the scoring results.

[0168] This module executes step S2, which receives the preprocessed health data vector and calculates in parallel the user's aging score, which is precisely quantified across multiple preset physiological dimensions.

[0169] Aging Atlas and Healthy Age Generation Module: This module aims to improve the interpretability of the model, achieving a leap from scoring to mechanism. It executes step S3, taking the multi-dimensional score output by the aging score calculation module as input.

[0170] In terms of software implementation, this module integrates an advanced library of graph learning and causal discovery algorithms. It utilizes graph attention networks to learn the complex nonlinear correlation strengths between nodes in various physiological dimensions, generating a prior correlation matrix. Then, a constrained causal discovery method guided by graph attention networks is employed, integrating the prior matrix as a constraint into a directed acyclic graph structure optimization based on the NOTEARS framework. Ultimately, a weighted, directed, personalized aging interaction map is obtained.

[0171] This map clearly reveals key causal pathways such as metabolic disorders driving inflammation. This module integrates scores from various dimensions and their weights on health risks, and calculates a single, comprehensive health age index through a weighted formula, providing users with an intuitive overview of their aging status.

[0172] Digital Twin Building Module: This module is the cornerstone of the system's dynamic simulation. It executes step S4, responsible for creating a personalized digital twin of each user's physiological system. At the core of this module is a parameterized general physiological system dynamic model, mathematically described by a set of ordinary differential equations. This modular model's state variables encompass at least six dimensions: cardiovascular function, metabolic homeostasis, immune-inflammatory regulation, neurocognition, liver function, kidney function, musculoskeletal function, and endocrine function.

[0173] The module incorporates a highly efficient optimization algorithm that uses the user's personal health data vector and multi-dimensional aging score as calibration targets. By solving a nonlinear minimization problem, it performs personalized calibration on hundreds of parameters in the general model, enabling the theoretical model to simulate the user's unique physiological dynamics and homeostasis with high fidelity, thus becoming the user's digital representative in virtual space.

[0174] Intervention Program Simulation and Optimization Module: This module is the brain behind the system's generation of personalized health prescriptions. It executes step S5, running on the personalized digital twin generated by the digital twin construction module. This module connects to a predefined, continuously updated knowledge base of intervention measures, where each intervention corresponds to adjustment rules for specific parameters of the digital twin.

[0175] The module employs a multi-objective optimization algorithm based on genetic algorithms to simulate the application of different intervention combinations on a digital twin and runs long-term simulations to evaluate their optimization effects on multiple objectives. By searching for the Pareto optimal solution set, the module can output a detailed simulation prediction report, providing users with a variety of highly personalized optimal intervention combinations validated in virtual clinical trials and their expected effects.

[0176] Dynamic Monitoring and Update Module: This module ensures the system functions as a learning organism, achieving closed-loop optimization. It executes step S6, continuously monitoring new data after the user implements the intervention plan. Through periodic repetition of data collection and aging score calculation processes, it obtains the user's new actual observation score. The intervention plan implemented within that time period is input into the digital twin to obtain the predicted score at the same time point. The deviation vector between the predicted and actual observation values ​​is calculated, and an online learning algorithm based on Kalman filtering is used to iteratively update the model parameters of the digital twin based on this deviation.

[0177] This process allows the user's digital twin to continuously approximate their actual physiological changes, making the system's predictions and recommendations increasingly accurate.

[0178] Understandably, users submit their multi-dimensional health data through the interface of the data collection and preprocessing module.

[0179] The system then automatically initiates the analysis process: after preprocessing, the data flows sequentially through the aging score calculation module, the aging atlas and healthy age generation module to generate an assessment report, and the digital twin construction module to create an initial digital twin for the user; the intervention plan simulation and optimization module simulates and optimizes based on this, generates a personalized intervention plan and submits it to the user.

[0180] Once the user begins implementing the plan, the dynamic monitoring and update module is activated, initiating a continuous learning and optimization cycle for the system.

[0181] Through close collaboration among its modules, the entire system deeply integrates assessment, prediction, and intervention, ultimately achieving personalized, dynamic, and precise anti-aging management.

[0182] Example 4: A 52-year-old male user, calendar age 52. After the system collects his data, it undergoes preprocessing to obtain... Output of the multidimensional aging score calculation module: , , , The remaining dimensions are close to 0. GAI and LINGAIM analyses show... The dominant pathway. The Healthy Age Index is calculated as follows:

[0183]

[0184] Recommended scheme for simulation optimization module after digital twin calibration:

[0185] 1. Mediterranean diet; 2. 150 minutes of brisk walking per week; 3. 250mg of NMN daily.

[0186] Simulation predicts one year later =54.2 years old, =4.5 years old.

[0187] The user underwent a follow-up examination 6 months later, and the actual... =56.1 years old, residuals are used for EKF to update Li Sheng's body parameters.

[0188] Comparative Example 1: For the same user, a traditional single biological age model output a biological age of 55 years, but the reason for this could not be explained, nor could any actionable intervention suggestions be provided. The user attempted high-intensity interval training on their own, but discontinued it after 3 months due to a knee injury. Metabolic indicators showed no improvement.

[0189] Comparative Example 2: For the same user, using a multidimensional model that does not consider norms for the Chinese population, the output is...

[0190] The study underestimated metabolic issues, and the recommended plan only included general dietary advice. Simulation predictions... =1.2 years old, no significant improvement was observed after actual implementation.

[0191] Experimental Example 1:

[0192] To verify the actual efficacy of the anti-aging prediction system and method proposed in this invention, we designed and implemented a 12-month longitudinal comparative study. This experiment aims to specifically evaluate the advantages of the system of this invention compared with existing mainstream methods in three key dimensions: accuracy of healthy age prediction, long-term user compliance, and improvement of core physiological indicators.

[0193] We selected 100 Chinese male participants with similar baseline characteristics and randomly divided them into three groups for intervention:

[0194] The results are shown in the table below:

[0195] index Example 5 Comparative Example 1 Comparative Example 2 Correlation coefficient between predicted healthy age and actual observed values ​​one year later 0.87 0.42 0.58 Intervention adherence rate (6 months) 78% 35% 52% The proportion of key dimension (metabolism) improvements >0.5 SD 68% 22% 39% Net reversal of healthy age (1 year) 3.2 ± 1.1 years 0.4 ± 0.9 years 1.1 ± 1.0 years

[0196] Data shows that the present invention is significantly superior to the comparative scheme in terms of prediction accuracy, intervention effectiveness, and user compliance.

[0197] In terms of prediction accuracy, the correlation coefficient between the predicted and actual values ​​of the system of this invention is as high as 0.87, which is much higher than 0.42 and 0.58 of Comparative Example 1 and Comparative Example 2, respectively. This proves that the multi-task deep learning model based on the Chinese population cohort training, combined with the personalized aging interaction map constructed by causal inference, can more accurately characterize and predict the aging trajectory of individuals.

[0198] Example 5 is more than twice that of Example 1. The interpretable intervention path and digital twin simulation preview function provided by this invention enable subjects to understand clearly, thereby greatly improving their willingness and stickiness to implement the intervention.

[0199] In the experimental group, over 80% of the subjects experienced a significant reversal of metabolic function, far exceeding the performance of Comparative Example 5 and Comparative Example 2. This demonstrates the effectiveness of the closed-loop system of this invention, from assessment to personalized intervention and dynamic optimization. The system not only identifies problems but also develops efficient and personalized intervention plans by identifying key causal drivers, thereby achieving substantial health improvements.

[0200] This experimental example, through a rigorous controlled design, demonstrates that the system of this invention is not a simple improvement on existing technologies, but rather a new paradigm of anti-aging management that achieves significant advantages in prediction accuracy, user engagement, and actual health outcomes by integrating population-specific data, interpretable AI, and dynamic simulation optimization.

[0201] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for individual aging assessment prediction based on multi-dimensional health data and artificial intelligence, characterized in that, The method comprises the following steps: S1, collecting multi-dimensional health data of a user, and generating a standardized personal health portrait after processing; S2, based on the personal health portrait, using a trained aging assessment model to calculate the user's aging quantitative indicators in multiple physiological dimensions; S3, based on the multi-dimensional aging quantitative indicators, analyzing the interaction relationship between each physiological dimension, and generating a comprehensive health age index; S4, based on the personal health portrait and the aging quantitative indicators, calibrating the parameters of the general physiological model, and constructing the personalized digital twin of the user; S5, simulating the effects of different intervention combinations on the personalized digital twin, and generating personalized schemes and prediction reports based on optimization objectives; S6, updating the aging quantitative indicators and the health age index according to the intervention combination effects, and dynamically calibrating the digital twin. 2.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S1, the standardization refers to the Z-score standardization method based on the health norm database of a specific population; the specific population is the Chinese population, and the norm database is the Chinese adult health norm database. 3.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S2, the aging assessment model is a multi-task deep learning model, which adopts a fully connected neural network structure with a shared bottom layer and independent task heads; the shared bottom layer includes three fully connected layers, each layer has 400-600, 200-300, and 100-150 neurons respectively, and the activation function is ReLU. 4.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S3, the analysis of the interaction relationship between each physiological dimension adopts a constraint causal discovery method guided by an attention network. 5.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S4, the calibration of the general physiological model parameters is realized by solving an optimization problem, and the objective function is the minimization of the Euclidean distance between the aging score vector and the model steady-state output vector. 6.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S5, the generation of personalized schemes based on optimization objectives adopts a multi-objective optimization algorithm based on a genetic algorithm. 7.The method of individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to claim 1, characterized in that: In step S6, the dynamic calibration of the digital twin adopts an online learning algorithm based on Kalman filtering.

8. A system for individual aging assessment prediction based on multi-dimensional health data and artificial intelligence, based on the method for individual aging assessment prediction based on multi-dimensional health data and artificial intelligence according to any one of claims 1 to 7, characterized in that, It comprises: a data acquisition and preprocessing module for executing step S1; an aging score calculation module for executing step S2; an aging atlas and health age generation module for executing step S3; a digital twin construction module for executing step S4; an intervention scheme simulation optimization module for executing step S5; a dynamic monitoring and updating module for executing step S6. 9.The system for individual aging assessment and prediction based on multi-dimensional health data and artificial intelligence according to claim 8, characterized in that: The multi-task deep learning model adopts a fully connected neural network structure with a shared bottom layer and independent task heads, and its loss function is the weighted sum of the prediction errors of each dimension, and the weights are set according to the influence of the functional decline of each physiological dimension on the all-cause mortality risk. 10.The system for individual aging assessment and prediction based on multi-dimensional health data and artificial intelligence according to claim 8, characterized in that: The physiological system dynamics model is a modular model described by ordinary differential equations, and its state variables cover at least six dimensions of cardiovascular function, metabolic homeostasis, immune inflammation regulation, neural cognition, liver function, kidney function, muscle and skeletal function, and endocrine function.

Images