Dementia risk prediction system based on associated information processing model

Abstract

The invention relates to the technical field of health information, in particular to a dementia risk prediction system based on an associated information processing model, and the system firstly extracts multi-modal hidden variable features through a data processing module, and generates an individualized risk feature matrix through causal reasoning and association rule optimization. And then, the system dynamically generates a personalized intervention path by adopting a reinforcement learning optimization algorithm, and realizes long-term risk prediction through a recurrent neural network. The system outputs personalized prediction reports including dementia risk assessment, trend analysis and intervention suggestions. Experiments show that the accuracy of dementia risk prediction, the dynamic adaptability of an intervention strategy and the efficiency of health management can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of health information technology, and in particular to a dementia risk prediction system based on an association information processing model. Background Art

[0002] Dementia is a common neurodegenerative disease, especially in the elderly population, and its incidence and severity impose a huge burden on individuals, families and society. Early prediction of dementia risk and development of personalized intervention strategies can significantly slow down disease progression, improve patients' quality of life and reduce socioeconomic burden.

[0003] The "risk factor scoring method" and "Cox regression model probability method" used in the existing technology (Chinese invention patent, publication number: CN117373668A, name: A method for establishing a risk prediction model for the onset of senile dementia) have certain application value in traditional dementia risk assessment, but still have the following defects: the risk factor scoring method and the Cox regression model probability method rely on static data analysis, ignore the dynamic association between multimodal data, and are difficult to reflect the risk trend of individual characteristics changing over time; the risk factor scoring method mainly calculates the risk index through a simple weighted accumulation method, while the Cox regression model mainly relies on the covariate influence of a single variable and lacks adaptability to high-order correlations and nonlinear dependencies. Both methods are based on fixed models and cannot dynamically update or optimize prediction strategies. In addition, the sensitivity and specificity in risk assessment are average, making it difficult to support real-time intervention and optimization. Summary of the invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a dementia risk prediction system based on an associative information processing model. The present invention is based on multimodal data fusion, uses a reinforcement learning optimization algorithm to dynamically generate personalized intervention paths, and realizes long-term risk prediction through a recurrent neural network. The system of the present invention generates accurate individualized prediction reports through multi-level data analysis and optimization, which significantly improves the accuracy of dementia risk prediction and the practicality of intervention measures.

[0005] A dementia risk prediction system based on an associative information processing model, comprising: A data processing module is used to collect multimodal data, including genetic data, health data, behavioral data, and environmental data, and process all the data to generate latent variable data; A causal reasoning module is used to construct a causal model based on latent variable data, calculate causal weights and generate an individualized risk feature matrix, wherein the feature matrix is ​​generated by comprehensively analyzing the weighted contributions of different causal relationship variables; based on the feature matrix, a nonlinear regression algorithm is applied to calculate individualized risk index data to reflect the likelihood of dementia for each individual; The intervention optimization module is used to build a reinforcement learning environment based on individualized risk index data, generate dynamically adjusted intervention strategy data through a deep Q-network-based optimization algorithm, and use a reinforcement learning framework to combine real-time updated risk indicators and intervention feedback to optimize the intervention path to reduce the risk of dementia; The prediction and generation module is used to perform long-term risk prediction, short-term trend analysis and individual stratification screening based on dynamically adjusted intervention strategy data, and generate personalized prediction report data, which includes risk assessment and intervention recommendations.

[0006] Preferably, the data processing module includes a missing data filling unit, and the missing data filling unit processes the missing values ​​in the multimodal data by the following steps: For continuous numerical data, mean interpolation is used to fill in missing values ​​based on the mean values ​​in other data records; For categorical variables, the nearest neighbor algorithm-based imputation method is used to select the most common category value in similar data records for imputation; The filled multimodal data generates initial latent variable data.

[0007] Preferably, the data processing module performs time series analysis on the behavior data and the environmental data, extracts trend features and updates latent variable data through the following steps: The behavioral data and environmental data are divided into time windows of fixed length, and the features of the data in each time window are calculated independently; Calculate the mean of the data in each time window to generate the mean parameter in the trend feature; Calculate the rate of change of the mean values ​​in adjacent time windows to generate the rate of change parameter in the trend feature; Applying linear regression analysis to the data in the time window, calculating the trend slope parameter, and generating the slope parameter in the trend feature; The initial latent variable data is updated based on the extracted trend features to obtain updated latent variable data.

[0008] Preferably, the individualized risk signature matrix generated by the causal reasoning module is calculated by the following steps: Determine the causal weight and calculate it using the following formula:

[0009] in, Representation characteristics Features The causal weight of Features and Features The covariance of Features The variance of The eigenvalues ​​in the feature matrix are calculated based on the causal weights, which is achieved through the following formula:

[0010] in, is the eigenvalue of the matrix, Representation characteristics In the characteristic matrix, the eigenvalues The contribution weight of is the latent variable feature, and n is the total number of all feature variables involved in the data calculation process.

[0011] Preferably, the causal reasoning module optimizes the feature matrix in combination with the association rules, which is achieved by the following steps: Use the Apriori algorithm to mine association rules between features from latent variable data; Adjust the feature matrix according to the mined rules and weight the association rules to causal weights; The adjusted feature matrix is ​​regenerated by combining the causal weights and the association weights.

[0012] Preferably, the individualized risk index is calculated by the following formula:

[0013] in, represents the individual risk index, Features The regression coefficient of is the latent variable eigenvalue, is the constant term of regression, and n is the total number of all characteristic variables involved in the data calculation process.

[0014] Preferably, the reinforcement learning environment of the intervention optimization module includes the following parts: The state space consists of individualized risk index data and high-risk factor annotations; Action sets, including dietary modifications, exercise recommendations, and medications; The reward function is calculated by the following formula:

[0015] in, Represents the reward value, is the change in risk index before and after intervention, The economic and time costs of the intervention.

[0016] Preferably, the optimization algorithm of the reinforcement learning adopts a deep Q network, and the training process includes the following steps: Initialize the parameters of the reinforcement learning model and determine the initial state through individualized risk index data; use - Greedy strategies that balance between exploring new interventions and exploiting existing ones; The Q value is updated in multiple rounds of iterations, and the intervention path is continuously optimized through the reward function.

[0017] Preferably, the long-term risk prediction task of the prediction and generation module is implemented through a recurrent neural network, the input features include the individualized risk index and high-risk factor annotations in the intervention strategy data after dynamic adjustment, and the output is a risk curve in the future time.

[0018] Preferably, the personalized prediction report includes the following contents: Summary of dementia risk assessment; long-term trend analysis results, including changes in the risk curve; short-term trend analysis results; individual stratification results and key high-risk factors; specific intervention recommendations for individuals.

[0019] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention uses multimodal data fusion technology to achieve accurate extraction of latent variable features and dynamic association modeling, overcoming the limitations of traditional methods that rely only on univariate analysis and enhancing the ability to comprehensively characterize factors affecting dementia risk; The present invention uses an intervention optimization algorithm driven by reinforcement learning to achieve a personalized intervention path that can be adjusted in real time, effectively solving the problem of fixed intervention strategies in the prior art and improving dynamic risk management capabilities. The present invention uses the long-term risk prediction technology of recurrent neural networks to achieve accurate prediction of future risk change trends, providing scientific support for planning intervention measures in advance; The present invention realizes the comprehensive optimization of stratified screening and intervention suggestions through a personalized prediction report generation module, and has both high efficiency and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural block diagram of the system of the present invention; Figure 2 This is a schematic diagram of the data flow of the causal reasoning module in the present invention; Figure 3 This is a schematic diagram of the reinforcement learning process of the intervention optimization module in the present invention; Figure 4 A schematic diagram of the process of generating a personalized prediction report in the present invention. DETAILED DESCRIPTION

[0021] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0022] like Figure 1 As shown, a dementia risk prediction system based on an associative information processing model comprises: A data processing module is used to collect multimodal data, including genetic data, health data, behavioral data, and environmental data, and process all the data to generate latent variable data; Genetic data refers to genetic information related to dementia risk extracted from an individual's genome. Genetic variations (such as single nucleotide polymorphisms, SNPs) can reflect an individual's genetic susceptibility to neurodegenerative diseases. The present invention analyzes genetic data to identify gene loci that affect the onset of dementia and their expression relationships, thereby providing a genetic basis for risk assessment.

[0023] Genetic data is usually obtained through high-throughput sequencing technology (such as whole genome sequencing or gene chip technology). Genetic loci such as APOE4, CLU, PICALM, etc. have been shown to be closely related to dementia diseases such as Alzheimer's disease. Genetic data are represented in a coded manner. For example, each locus can be coded as 0 (wild type), 1 (heterozygous) or 2 (mutant). Through the analysis of genetic data, it is possible to identify the contribution of genetic factors to the risk of dementia, especially for individuals with a family history or genetic susceptibility. It has a high predictive value. The introduction of genetic data provides the system with a basic dimension of dementia risk.

[0024] In the embodiment, the genetic data of an individual shows that the APOE4 gene locus is 2 (mutant), which has been proven to be a high risk factor for dementia. In the present invention, the weight of the gene locus is assigned in the latent variable matrix, which directly affects the calculation of the individualized risk index.

[0025] Health data refers to continuous or categorical variables related to an individual's physical health status, which is used to reflect the individual's physiological state and disease risk. Common health data include blood sugar, blood pressure, BMI (body mass index), heart rate, cholesterol level, etc. These indicators are usually closely related to cardiovascular and cerebrovascular health and have an important impact on the onset of dementia.

[0026] Health data can be collected through physical examination reports, medical equipment (such as blood pressure monitors, blood glucose meters) or wearable devices. Health data are usually stored in the form of continuous variables and input into the model after being processed by standardized formulas. For example, hypertension and diabetes are important risk factors for dementia. Long-term increases in blood pressure may cause damage to brain blood vessels, thereby increasing the risk of dementia. Health data can supplement the current health status of an individual that cannot be reflected in genetic data, and provide a controllable dimension for dementia risk assessment (such as improving blood pressure or blood sugar levels through drug intervention).

[0027] In this example, an individual's fasting blood glucose level is 7.5 mmol / L (high), and combined with his BMI value (27, overweight), the system determines that he has a tendency to metabolic syndrome. In the latent variable matrix, these indicators will be combined with behavioral and genetic data to comprehensively assess the individual's risk of dementia.

[0028] Behavioral data reflects an individual's lifestyle and daily behavior habits, which indirectly affect the risk of dementia by affecting physical and brain health. Common behavioral data include exercise volume, eating habits, smoking history, alcohol consumption, etc. Behavioral data can be obtained through individual self-reports, wearable devices (such as pedometers), or smart monitoring devices. Quantitative data (such as walking distance) in behavioral data is stored in the form of time series, and qualitative data (such as dietary preferences) is processed in a vectorized manner.

[0029] Regular exercise can reduce the risk of dementia, and walking frequency and distance are important indicators; long-term smoking and excessive drinking are high-risk behavioral factors for dementia; high-fat, high-sugar diets may increase inflammatory responses and accelerate cognitive decline.

[0030] Behavioral data provide the most modifiable risk factors for dementia, and risk can be significantly reduced through behavioral interventions (such as improving diet and increasing exercise).

[0031] In this example, the data recorded by an individual through a smart bracelet shows that his daily steps are only 2,000 (lower than the recommended value of 8,000). At the same time, his diet record shows that the proportion of high-fat intake is too high. In the model, these behavioral characteristics are marked as high-risk factors, and the individualized risk index is increased.

[0032] Environmental data refers to the impact of external environmental factors on dementia risk, including air pollution, living environment (such as urban or rural) and social support networks. These factors indirectly affect dementia risk by affecting individual health or psychological state over a long period of time. Environmental data can be obtained through public databases, meteorological monitoring stations or questionnaires. Processing methods include: quantitative data (such as PM2.5 concentration) is directly input into the model; qualitative data (such as type of residence) is processed by classification and vectorization.

[0033] Long-term exposure to high PM2.5 environments may increase brain inflammation, thereby accelerating cognitive deterioration. Loneliness and lack of social support are important sources of psychological stress, which may indirectly increase the risk of dementia. Environmental data can capture the long-term impact of an individual's external environment and provide an uncontrollable but critical reference factor for risk assessment.

[0034] For example, an individual lives in an urban area where the PM2.5 concentration is higher than 80 micrograms per cubic meter for a long time, and at the same time, the social support network index is low (such as living alone and having low social frequency). These data are given high weights in the model, which plays an important role in the assessment of the individualized risk index.

[0035] Preferably, the data processing module includes a missing data filling unit, and the missing data filling unit processes the missing values ​​in the multimodal data by the following steps: For continuous numerical data, mean interpolation is used to fill in missing values ​​based on the mean values ​​in other data records; For categorical variables, the nearest neighbor algorithm-based imputation method is used to select the most common category value in similar data records for imputation; The filled multimodal data generates initial latent variable data.

[0036] In the present invention, the missing data filling unit of the data processing module is used to solve the problem of missing values ​​in multimodal data to ensure the integrity and accuracy of subsequent analysis. The sources of multimodal data are complex, and different types of data (such as genetic data, health data, behavioral data, and environmental data) may be missing during the collection process. If these missing values ​​are not filled, the model performance will be reduced or cannot be processed.

[0037] For continuous numerical data (such as blood sugar, blood pressure and other health indicators), mean interpolation is used for filling. Mean interpolation calculates the mean of other records in the data set to fill in missing values, thereby minimizing the deviation introduced by filling. The formula is as follows:

[0038]

[0039] in, To fill the value, is the value of the non-missing record, The total number of all feature variables involved in the data calculation process.

[0040] For categorical variables (such as smoking status, residential area category, etc.), the nearest neighbor algorithm is used to fill in the missing values. The nearest neighbor algorithm calculates the similarity between each record and selects the most common category value in the most similar record to fill in the missing values. The nearest neighbor distance calculation formula is:

[0041]

[0042] in, Representation record and The Euclidean distance between and Record and The characteristic value of is the feature dimension. By comparing the distances of all records, the closest The category values ​​of the records are voted for and filled with the category with the highest frequency.

[0043] After the imputation is completed, all data records are reintegrated to generate a multimodal dataset without missing values. The imputed data are further converted into initial latent variable data to provide input for the subsequent causal inference module.

[0044] Solve the missing problem in multimodal data and provide consistent data input for subsequent analysis. The mean interpolation method of continuous variables avoids too many complex calculations and reduces the impact of data filling on the overall distribution; the nearest neighbor filling method of categorical variables ensures the rationality of the filling value through a voting mechanism. The design of the missing data filling unit improves the system's ability to handle complex data and avoids model failure due to incomplete data.

[0045] Embodiment, in a dementia risk prediction scenario, an individual's genetic data, health data, behavioral data and environmental data are collected. Among them, genetic data: all key gene sites have been measured, and there are no missing values. Health data: blood pressure records are missing, and other records are normal. The system uses the mean interpolation method to use the mean blood pressure of the same group of people (such as 120 / 80mmHg) as the filling value. Behavioral data: exercise frequency data is missing. The system compares the similarity of other records through the nearest neighbor algorithm, and selects the most common exercise frequency to fill in (such as "walking 5 times a week") based on the walking distance and the category of daily activities. Environmental data: some air pollution index records are missing, and the system fills them with the mean of the same area and time period.

[0046] The filled data are integrated to generate a non-missing multimodal data set, which is further used to generate the initial latent variable data. Through latent variable matrix calculation, the key high-risk factors of individuals are identified to provide support for subsequent risk assessment and intervention strategies.

[0047] Preferably, the data processing module performs time series analysis on the behavior data and the environmental data, extracts trend features and updates latent variable data through the following steps: The behavioral data and environmental data are divided into time windows of fixed length, and the features of the data in each time window are calculated independently; Calculate the mean of the data in each time window to generate the mean parameter in the trend feature; Calculate the rate of change of the mean values ​​in adjacent time windows to generate the rate of change parameter in the trend feature; Applying linear regression analysis to the data in the time window, calculating the trend slope parameter, and generating the slope parameter in the trend feature; The initial latent variable data is updated based on the extracted trend features to obtain updated latent variable data.

[0048] In the present invention, behavioral data and environmental data are key dimensions of dynamic change. By extracting trend features through time series analysis, we can effectively capture the change patterns of these data over time, and further update the latent variable data, providing more timely and dynamic input for the subsequent causal reasoning module. Time series analysis uses a fixed-length time window to slice the data, and extracts trend features through statistical and regression methods, including mean, rate of change, and slope, which reflect the dynamic characteristics of behavioral and environmental data from different levels.

[0049] Behavioral data (such as exercise frequency) and environmental data (such as air pollution index) are stored in the form of time series. To ensure the timeliness and consistency of analysis, the system divides the time series data into windows of fixed length, and the data in each time window is processed independently. The window length is set according to the actual application scenario, such as day, week or month, and the specific choice needs to balance timeliness and stability.

[0050] In each time window, the mean parameters of the behavioral data and environmental data are calculated to reflect the average state of the time period. The mean calculation formula is:

[0051]

[0052] in, is the mean, is the observation value in the window, The mean parameter is used to capture long-term stable features, such as the average number of steps per day or air pollution levels over a period of time.

[0053] The rate of change is calculated for the mean values ​​in adjacent time windows to reflect the direction and magnitude of trend changes in the short term. The rate of change calculation formula is:

[0054]

[0055] in, is the rate of change, and are the means of the current window and the next window, respectively. The rate of change parameter can reveal the dynamic characteristics of data fluctuations. For example, a rapid increase in the air pollution index may be associated with short-term risks.

[0056] In each time window, the trend slope is calculated by linear regression method to describe the long-term trend of the data. The slope calculation formula is:

[0057]

[0058] in, is the slope, and They are time and variable value, and is the mean of the time and variable, and n is the total number of all characteristic variables involved in the data calculation process. The slope parameter is used to identify the long-term change direction of the data, such as the gradual decrease in walking distance within a year.

[0059] The trend features (mean, rate of change, and slope) extracted through the above steps are input into the latent variable model as new feature vectors to update the initial latent variable data and generate updated latent variable data. This latent variable data more comprehensively reflects the dynamic characteristics and provides support for subsequent causal reasoning and individualized risk assessment.

[0060] Through time series analysis of behavioral and environmental data, the dynamic changes of data over time can be captured, supplementing information that static data cannot reflect. Trend characteristics (such as slope and rate of change) can reveal the potential relationship between behavioral and environmental data and dementia risk, and improve the strength of association in subsequent model analysis. The dynamic update of latent variable data enables the model to reflect the current risk status, thereby enhancing the system's adaptability to time-sensitive characteristics.

[0061] In an embodiment, in a dementia risk prediction application, behavioral data and environmental data of an individual are collected and processed as follows: The system records the individual's daily steps and the air pollution index of the corresponding area, and divides the data into a time window per week. For example:

[0062] Week 1: number of steps [5000, 5200, 4800, 5500, 5300, 5100, 4900], air pollution index [60, 65, 62, 70, 72, 68, 66]; Week 2: Steps [4800, 4700, 4500, 4600, 4300, 4400, 4200], Air pollution index [75, 78, 80, 85, 88, 86, 82].

[0063] Mean: The mean number of steps in the first week was 5100, and the mean air pollution index was 65.57; the mean number of steps in the second week was 4500, and the mean air pollution index was 82. Change rate: The change rate of step count is -11.76%, and the change rate of air pollution index is +25.07%.

[0064] Trend slope: Calculated by linear regression, the slope of step count is -100 (a decrease of 100 steps per day) and the slope of the air pollution index is +3.71 (an increase of 3.71 points per day).

[0065] The extracted trend features (mean, rate of change, and slope) are input into the latent variable model, and the updated latent variable data contains the dynamic characteristics of the behavior and environmental data. For example:

[0066] Week 1: latent variable features [5100, +0, -100, 65.57, +0, +3.71]; Week 2: Latent variable features [4500, -11.76%, -100, 82, +25.07%, +3.71].

[0067] The updated latent variable data showed that the individual's exercise volume decreased significantly and air pollution increased significantly, suggesting that the risk of dementia may increase in the short term. These data will be further input into the causal reasoning module for risk assessment and intervention optimization.

[0068] like Figure 2 As shown, the causal reasoning module is used to build a causal model based on latent variable data, calculate causal weights and generate an individualized risk feature matrix, wherein the feature matrix is ​​generated by comprehensively analyzing the weighted contributions of different causal relationship variables; based on the feature matrix, a nonlinear regression algorithm is applied to calculate individualized risk index data to reflect the possibility of dementia for each individual; The causal reasoning module is the core module of the dementia risk prediction system based on the association information processing model. Its main function is to use latent variable data to build a causal model, calculate causal weights, generate an individualized risk feature matrix, and finally calculate the individualized risk index data through a nonlinear regression algorithm. Causal reasoning analyzes the potential causal relationship between multimodal data, quantifies the impact of each variable on the risk of dementia, and provides a basis for accurate risk assessment and intervention decisions.

[0069] A causal model is a graph structure that describes the causal relationship between variables, where nodes represent latent variable features and edges represent causal relationships. The construction of a causal model is based on latent variable data, using the maximum entropy optimization method combined with conditional independence testing to determine the direction and strength of the causal relationship between variables. For example, genetic variation may further increase the risk of dementia by affecting cognitive function, while long-term lack of exercise may indirectly increase the risk by affecting cardiovascular health.

[0070] The causal weight reflects the strength of the causal contribution of one variable to another variable. Based on the calculated causal weight, the weighted contributions of multiple variables are comprehensively analyzed to generate an individualized risk feature matrix. Each row of the matrix corresponds to a latent variable feature, and each column corresponds to an individualized feature value. Based on the generated feature matrix, the nonlinear regression algorithm is used to calculate the individualized risk index.

[0071] Preferably, the individualized risk signature matrix generated by the causal reasoning module is calculated by the following steps: Determine the causal weight and calculate it using the following formula:

[0072] in, Representation characteristics Features The causal weight of Features and Features The covariance of Features The variance of The eigenvalues ​​in the feature matrix are calculated based on the causal weights, which is achieved through the following formula:

[0073] in, is the eigenvalue of the matrix, Representation characteristics In the characteristic matrix, the eigenvalues The contribution weight of is the latent variable feature, and n is the total number of all feature variables involved in the data calculation process.

[0074] In the present invention, the causal reasoning module calculates the causal weights through latent variable data and generates an individualized risk feature matrix based on the causal weights. The generation of the feature matrix is ​​a quantitative expression of the causal relationship between latent variable features, aiming to identify the weighted contribution of different latent variables to the risk of dementia, thereby providing high-quality input for subsequent risk index calculation.

[0075] Causal weight is an important indicator to quantify the strength of causal relationship between features. By analyzing the statistical relationship between latent variable features, the causal weight between each pair of features is calculated. The specific formula is: The causal weight is quantified by the ratio of covariance to variance right influence rather than simple correlation, thus avoiding the interference of non-causal relationships.

[0076] After calculating the relationship between latent variable features through causal weights, an individualized risk feature matrix is ​​generated. Each row of the feature matrix corresponds to a latent variable feature, and each column corresponds to an individualized eigenvalue. The eigenvalue is calculated using the following formula: .

[0077] During the feature matrix generation process, causal weights can be updated in real time based on new data, and latent variable eigenvalues ​​can also be adjusted based on dynamic changes, so that the system can adapt to the differences between different individuals and dynamic changes over time.

[0078] By calculating causal weights, the causal relationship between latent variable features can be quantified, avoiding the bias that may be introduced by simple correlation analysis. Through weighted calculation of the feature matrix, the generated eigenvalues ​​can more accurately reflect the contribution of individual characteristics to dementia risk, thereby improving the accuracy of model prediction. The real-time update capability of causal weights and feature matrices enables the system to respond quickly to changes in different individuals and time, improving overall performance.

[0079] In the actual application of a dementia risk prediction system, the latent variable data of an individual includes genetic data (such as APOE4 gene locus status), health data (such as blood pressure and BMI), behavioral data (such as daily steps) and environmental data (such as air pollution index). The causal reasoning module performs the following processing on the individual's data:

[0080] The causal weight is calculated by the covariance and variance between latent variable features: The causal weight of the APOE4 gene on the risk of dementia is:

[0081] The causal weight of daily steps on cardiovascular health is:

[0082] The causal weight of air pollution index on dementia risk is:

[0083] Generate an individualized risk feature matrix based on the causal weights and latent variable data. For example, the feature matrix of an individual contains the following values:

[0084] like and but:

[0085] Assume that the air pollution index increases significantly to , the system will recalculate the dementia risk characteristic value in real time:

[0086] The risk is increased. Through the above processing, the system generates an accurate feature matrix based on individual characteristics, providing high-quality input for subsequent risk index calculations.

[0087] Preferably, the causal reasoning module optimizes the feature matrix in combination with the association rules, which is achieved by the following steps: Use the Apriori algorithm to mine association rules between features from latent variable data; Adjust the feature matrix according to the mined rules and weight the association rules to causal weights; The adjusted feature matrix is ​​regenerated by combining the causal weights and the association weights.

[0088] In the present invention, the causal reasoning module not only calculates the causal weights, but also optimizes the feature matrix by combining association rules, thereby further improving the feature matrix's ability to characterize latent variable relationships. Association rules supplement the nonlinear or high-order dependencies that causal weights cannot fully capture by mining the frequent patterns and association relationships between features in latent variable data. The optimized feature matrix combines causal weights with association rule weights to generate more accurate and comprehensive input data, providing more reliable support for subsequent risk assessment.

[0089] Association rule mining uses the Apriori algorithm to identify frequent item sets and association rules from latent variable data. The core of the Apriori algorithm is to discover strong association rules between features by gradually expanding frequent item sets. For example, certain gene variants and specific lifestyle habits (such as smoking or high-fat diet) may jointly have an important impact on the risk of dementia, and this rule may not be directly expressed by a single variable.

[0090] By setting the minimum support threshold, we can filter out the frequent combinations that appear in the latent variable features. For example:

[0091] Gene and behavioral characteristics If it appears in 70% of the data records at the same time, its support is 0.7, which meets the minimum support threshold.

[0092] Generate association rules based on frequent item sets and evaluate the strength of the rules by calculating confidence and lift. For example:

[0093] → The confidence level is 0.8 and the lift is 1.5, indicating that the two are strongly correlated.

[0094] The mined association rules are introduced into the feature matrix by assigning weights. The association rule weights reflect the strength of the rules, and combined with the causal weights, an optimized feature matrix is ​​generated:

[0095]

[0096] in, Represents the optimized weight, which integrates the causal weight and the association rule weight; represents the causal weight and represents the latent variable right causal contribution of Represents the weight of the association rule, which is calculated by confidence and lift; Represents the adjustment coefficient of the association weight, which is used to balance the contribution of causal weight and association weight.

[0097] The causal weights and association rule weights are combined to generate an optimized feature matrix, in which each column represents the contribution of a latent variable feature to the overall risk. Based on causal reasoning, the optimized matrix further enhances the model's ability to express high-order dependencies.

[0098] By combining causal weights and association rule weights, the feature matrix more comprehensively reflects the complex relationship between latent variable features and improves the model's ability to identify nonlinear dependencies. The optimized feature matrix enables the model to capture the impact of more latent features on dementia risk, thereby generating a more accurate risk index. The introduction of association rules effectively compensates for the shortcomings of causal reasoning for high-order interaction patterns, and is particularly suitable for analysis scenarios of multimodal data.

[0099] In the embodiment, in a certain dementia risk prediction system, the specific application of the causal reasoning module combined with the association rule to optimize the feature matrix is ​​as follows: Mining frequent item sets and association rules between features from latent variable data, examples are as follows: Gene , behavioral characteristics , Environmental characteristics It appears in 60% of the data records, with a support of 0.6; → , the confidence level is 0.8 and the lift is 1.3; → , the confidence level is 0.7 and the lift is 1.2.

[0100] The association rule weights are introduced into the feature matrix, and the optimized weights are calculated using the formula. For example:

[0101] Raw causal weights: ; .

[0102] Association rule weight: ; .

[0103] Optimized weights:

[0104]

[0105] After integrating the causal weights and the association weights, the column eigenvalues ​​of the feature matrix are updated:

[0106] Assumptions and ,but:

[0107] The optimized feature matrix reflects more high-order correlations, making the model more accurate in predicting the individual's dementia risk.

[0108] Preferably, the individualized risk index is calculated by the following formula:

[0109] in, represents the individual risk index, Features The regression coefficient of is the latent variable eigenvalue, is the constant term of regression, and n is the total number of all characteristic variables involved in the data calculation process.

[0110] In the present invention, the individualized risk index is a comprehensive assessment of the likelihood of dementia for each individual, and its calculation is based on latent variable characteristics and corresponding regression coefficients. Through a nonlinear regression model, the individualized risk index combines the influence of each latent variable characteristic and quantifies it into a specific risk value to reflect the likelihood of an individual suffering from the disease.

[0111] Regression coefficient and constant term It is obtained through training with historical data and can minimize the error between the predicted risk index and the actual observed value.

[0112] Each latent variable eigenvalue The contribution to the individualized risk index is given by its regression coefficient For example, high-weight features (such as the APOE4 gene) have a more significant impact on the risk index, while low-weight features (such as dietary preferences) contribute less. Through this weighting method, the model is able to highlight the impact of key features while retaining the combined effects of other minor factors.

[0113] The training process of the model includes the following steps: using historical multimodal data as input, latent variable features The feature matrix generated by the previous module; set the actual risk label (such as clinical diagnosis results or medical scores) as output; use nonlinear regression optimization algorithms (such as gradient descent) to adjust the regression coefficients and constant term , in order to minimize the error between the model prediction results and the actual risk labels.

[0114] The individualized risk index can be recalculated based on the hidden variable features updated in real time to adapt to the dynamically changing individual characteristics. For example, an increase in the individual's exercise frequency may reduce the risk index, and the system will automatically update the index value based on the new data.

[0115] In an embodiment, in a dementia risk prediction scenario, the latent variable characteristics of an individual include the following: Feature data input: APOE4 gene locus status ( ):1.2; Daily steps ( ):0.8; BMI ( ):0.6.

[0116] Regression coefficients: The regression coefficient of APOE4 gene ( ):0.9; The regression coefficient of daily steps ( ): -0.5 (negative values ​​indicate protective effects); The regression coefficient of BMI ( ):0.7.

[0117] Constant term: The regression constant term ( ):0.3.

[0118] Risk index calculation:

[0119] Substitute specific values:

[0120] The calculation results show that the individual's risk index is 1.4, which is at a medium risk level.

[0121] Assume that an individual increases his step count from 0.8 to 1.2 by increasing his exercise. The system will recalculate the risk index based on the new feature value:

[0122]

[0123] After the update, the risk index dropped from 1.4 to 1.2, showing the significant effect of improved exercise on risk reduction.

[0124] like Figure 3 As shown, the intervention optimization module is used to build a reinforcement learning environment based on individualized risk index data, generate dynamically adjusted intervention strategy data through an optimization algorithm based on a deep Q network, and the optimization algorithm adopts a reinforcement learning framework, combined with real-time updated risk indicators and intervention feedback, to optimize the intervention path to reduce the risk of dementia; In the present invention, the intervention optimization module is used to generate dynamically adjusted intervention strategy data through a reinforcement learning model. The module takes individualized risk index data as input, combines real-time updated risk indicators and intervention feedback, optimizes the intervention path, and thus reduces the individual's risk of dementia. The core implementation is a reinforcement learning algorithm based on a deep Q network (DQN), which selects the optimal intervention strategy through simulation and feedback mechanisms.

[0125] The intervention optimization module transforms the dementia risk prediction problem into a reinforcement learning problem, using a reinforcement learning environment to describe individual states, possible intervention measures, and their effects.

[0126] DQN is a reinforcement learning algorithm based on Q-learning, combined with a deep neural network, for dealing with optimization problems in continuous state space. In the intervention optimization module, the core task of DQN is to learn the optimal intervention strategy, which is achieved through the following steps:

[0127] Initialize the weights of the deep neural network and randomly select actions for initial exploration; use Greedy strategy, a balance between exploring new actions and exploiting existing strategies; Use the Q-value function to update the strategy; Through multiple rounds of iterative network training, the Q value gradually converges and generates the optimal intervention strategy.

[0128] The module updates the individual's status based on real-time data, recalculates the risk index and optimizes the intervention path. For example, if an individual's increased exercise frequency causes the risk index to decrease, the module will reduce the priority of further increasing exercise and instead recommend other intervention measures, such as improving diet.

[0129] Preferably, the reinforcement learning environment of the intervention optimization module includes the following parts: The state space consists of individualized risk index data and high-risk factor annotations; Action sets, including dietary modifications, exercise recommendations, and medications; The reward function is calculated by the following formula:

[0130] in, Represents the reward value, is the change in risk index before and after intervention, The economic and time costs of the intervention.

[0131] In the present invention, the reinforcement learning environment of the intervention optimization module is used to build an adaptable decision-making framework, which generates dynamically adjusted intervention strategy data through the organic combination of state space, action set and reward function. The design goal of the module is to maximize the effect of intervention measures on reducing the risk of dementia while controlling economic and time costs.

[0132] State Space It is the core component of the reinforcement learning environment and is used to describe the individual's current health status and risk level. In this invention, the state space consists of individualized risk index and high-risk factor annotation:

[0133]

[0134] in, Represents the individualized risk index at the current moment; Represents the characteristic value of high-risk factors (such as BMI, daily steps, gene mutation status, etc.). The design of the state space can fully reflect the current health status of the individual and provide information support for the selection of intervention strategies.

[0135] The action set A defines the intervention measures that the system can take, including the following: Dietary adjustments: such as reducing fat and sugar intake and increasing antioxidant intake; Exercise suggestions: such as setting daily walking goals and increasing low- to medium-intensity aerobic exercise; Drug treatment: such as the use of anti-inflammatory drugs or smoking cessation aids.

[0136] Each intervention measure in the action set is highly correlated with health risk factors and can indirectly reduce the dementia risk index by adjusting the individual's high-risk factor characteristic values.

[0137] Reward Function The calculation formula used to quantify the effect of each intervention is:

[0138] The design of the reward function can achieve multi-objective optimization, obtain positive rewards by reducing the risk index, and avoid negative benefits caused by high-cost interventions.

[0139] By accurately describing the health status of individuals through state space, the reinforcement learning environment can generate personalized intervention strategies, thus avoiding a "one-size-fits-all" intervention approach. The real-time updated state space and reward function enable the system to respond quickly to changes in the health of individuals and generate real-time adjusted intervention strategies. Through the design of the reward function, the system can find the best balance between risk reduction and cost control, thereby improving the efficiency and feasibility of actual intervention. Each intervention measure in the action set directly targets high-risk factors. The reinforcement learning module optimizes the intervention path through multiple rounds of iterations, which can significantly reduce the individual's risk of dementia.

[0140] In the embodiment, the initial state of an individual is: Risk Index : 2.0 (high risk); High-risk factor characteristic values: BMI is 30 (obesity), daily step count is 1500 (low activity), and smoking status is current smoking.

[0141] Reinforcement learning environment initialization, state space initialization is ; The action set includes the following interventions: : Dietary modification (reducing calorie intake); : Increase exercise (daily walking goal 8,000 steps); : Quit smoking (use auxiliary medicine).

[0142] First intervention: System selection action (Add movement), the adjusted state becomes:

[0143] The risk index dropped to , the reward function calculates the reward value: Among them, the economic cost Indicates time investment.

[0144] Second Intervention: In the new state, the system selects the action (Dietary adjustment) to further optimize the status:

[0145] The risk index dropped to , the reward value is:

[0146] The system generates the optimal strategy path after multiple rounds of iterations based on the real-time updated status and reward function: Step 1: Increase exercise (daily walking goal of 8,000 steps); Step 2: Diet adjustment (increase dietary fiber intake); Step 3: Quit smoking (use auxiliary drugs). After the above interventions, the individual risk index dropped from 2.0 to 1.0, significantly reducing the risk of dementia.

[0147] Preferably, the optimization algorithm of the reinforcement learning adopts a deep Q network, and the training process includes the following steps: Initialize the parameters of the reinforcement learning model and determine the initial state through individualized risk index data; use Greedy strategies, which balance between exploring new interventions and exploiting existing ones; The Q value is updated in multiple rounds of iterations, and the intervention path is continuously optimized through the reward function.

[0148] In this invention, the reinforcement learning optimization algorithm generates dynamically adjusted personalized intervention strategies through a deep Q network (DQN), aiming to optimize the intervention path through multiple rounds of training iterations and significantly reduce the individual's risk of dementia. The algorithm uses the high adaptability of the reinforcement learning model to achieve a balance between exploring the effects of different intervention measures and optimizing existing strategies, thereby reducing the individual risk index while controlling the intervention cost.

[0149] The reinforcement learning model uses individualized risk index data and high-risk factor annotations as initial states :

[0150] in, is the initial value of the individualized risk index, It is the characteristic value of high-risk factors related to risk (such as BMI, number of steps per day, etc.). These states describe the health status of individuals and provide initial conditions for the optimization process.

[0151] To achieve a balance between exploring new interventions (expanding the strategy space) and leveraging existing strategies (improving performance), DQN adopts Greedy strategy.

[0152] The exploration phase is based on probability Randomly select intervention actions , try new strategies; Utilization stage based on probability Select the current best action .

[0153] The decision formula is: This strategy ensures that the model explores more potential valid paths in the early stages and focuses on optimizing the existing best path in the later stages.

[0154] After each round of intervention, the Q value is updated based on the immediate feedback. The Q value is updated using the following formula:

[0155]

[0156] in, Indicates in status Select Action the expected returns to be obtained; Represents the instant reward, which is calculated by the following formula , is the change in risk index, The economic and time costs of intervention. represents the discount factor, which represents the weight of future rewards; Represents the learning rate, which is used to control the step size of Q value update.

[0157] During the training process of reinforcement learning, the system will simulate the combination paths of different intervention strategies many times and adjust the Q value through reward feedback. As the number of training rounds increases, the Q value gradually converges and eventually generates the optimal intervention strategy path, which can significantly reduce the individual risk index at the lowest cost.

[0158] Through real-time status updates, DQN can automatically adjust intervention strategies based on individual health changes, making decisions more flexible and accurate. The reward function comprehensively considers risk reduction and intervention costs to ensure that the optimization path can significantly reduce risks while minimizing resource consumption. The introduction of the greedy strategy enables the model to explore more possible strategy paths in the early stage, while ensuring that it focuses on the optimal path in the later stage, improving the overall optimization efficiency. Through multiple rounds of training and iterative optimization, the generated intervention path can significantly reduce the individual risk index, thereby improving the scientificity and practicality of health management.

[0159] In an embodiment, in a dementia risk prediction application, the initial state of an individual is: Individualized risk index : 2.0 (high risk); High-risk factor characteristics: BMI is 30 (obesity), daily step count is 1500 (low activity), and smoking status is current smoking.

[0160] Reinforcement learning optimization process: Initial state .

[0161] Action Collection: : Adjust your diet; : Increase exercise; : Quit smoking.

[0162] Round 1 Intervention: use Greedy strategy to choose action (Increase exercise).

[0163] After adjustment, the number of steps per day increased to 8,000, and the risk index decreased to .

[0164] Instant Rewards: , where intervention cost Indicates the time investment cost.

[0165] Round 2 Intervention: New Status . Select Action (Dietary adjustments).

[0166] After adjusting for BMI down to 28, the risk index was further reduced to .

[0167] Instant Rewards:

[0168] Round 3 Intervention: New Status . Select Action (Quit smoking).

[0169] After adjustment, the smoking status becomes 0 (successful smoking cessation), and the risk index is reduced to .

[0170] Instant Rewards:

[0171] After multiple rounds of training, the model generated the optimal strategy path: Step 1: Increase exercise; Step 2: Adjust diet; Step 3: Quit smoking. The final result: The risk index dropped from 2.0 to 1.0, a 50% reduction in risk.

[0172] Preferably, the long-term risk prediction task of the prediction and generation module is implemented through a recurrent neural network, the input features include the individualized risk index and high-risk factor annotations in the intervention strategy data after dynamic adjustment, and the output is a risk curve in the future time.

[0173] In this invention, the long-term risk prediction task of the prediction and generation module is implemented through a recurrent neural network (RNN), whose core goal is to predict the individualized risk change trend in the future based on the dynamically adjusted intervention strategy data. With its powerful time series modeling capabilities, RNN captures the dynamic correlation between individualized risk index and high-risk factor characteristics, providing scientific predictions for long-term health management.

[0174] The input features of the long-term risk prediction task include the individualized risk index and high-risk factor annotations in the dynamically adjusted intervention strategy data, in the form of a time series:

[0175] in, Indicates the current time Individualized risk index; Represents high-risk factor annotations (such as BMI, daily steps, smoking status, etc.). The input features capture the dynamic changes of individual health status and its intervention measures over time, providing the necessary information for future risk prediction.

[0176] RNN processes time series data with its recursive network structure. Record historical information and predict future risk curves. The update formula for the hidden state is:

[0177]

[0178] in, Indicates the current time The hidden state of Represents the weight matrix from hidden state to hidden state; Represents the weight matrix input to the hidden state; Indicates the current input features (including and ); represents the bias term; Represents an activation function (such as tanh or ReLU). The output prediction value is the future risk curve:

[0179]

[0180] The model aims to minimize the error between the predicted risk curve and the true value, and uses the mean square error (MSE) as the loss function:

[0181] RNN optimizes parameters through back-propagation and gradient descent methods , and , in order to improve the prediction accuracy.

[0182] RNN can capture the dynamic patterns of individual risk index and high-risk factor annotations changing over time, generate risk curves in the future, and provide forward-looking support for health management. The module combines the individual's current health status and intervention history to generate long-term risk prediction results for specific individuals, avoiding a "one-size-fits-all" prediction model. By predicting future risk changes, the module provides a quantitative basis for formulating early intervention strategies to reduce long-term risks. RNN's time-dependent modeling capabilities enable the prediction results to reflect the trend of risk changes and provide comprehensive information support for decision-making.

[0183] In the example, the initial state of an individual is as follows: Current Risk Index =2.0; high-risk factor characteristic values: BMI is 30, daily steps are 2000, and smoking status is current smoking.

[0184] Long-term risk prediction process: The input sequence includes past individualized risk indexes and high-risk factor annotations:

[0185] in, =2.0 indicates the initial risk index; =1.8 indicates the updated risk index after intervention.

[0186] Through the recursive calculation of RNN, the temporal dependency between input features is captured and the future risk curve is predicted: , assuming the output is: , indicating that with continuous intervention, the risk index will gradually decrease.

[0187] Use historical data to verify model performance by optimizing the loss function:

[0188] Adjusting network weights , and bias , making the prediction results more accurate.

[0189] like Figure 4 As shown, the prediction and generation module is used to perform long-term risk prediction, short-term trend analysis and individual stratified screening based on the dynamically adjusted intervention strategy data, and generate personalized prediction report data, wherein the report includes risk assessment and intervention recommendations.

[0190] Preferably, the personalized prediction report includes the following contents: Summary of dementia risk assessment; long-term trend analysis results, including changes in the risk curve; short-term trend analysis results; individual stratification results and key high-risk factors; specific intervention recommendations for individuals.

[0191] In the present invention, the prediction and generation module completes long-term risk prediction, short-term trend analysis and individual stratification screening by comprehensively analyzing the dynamically adjusted intervention strategy data, and generates a personalized prediction report. This module can not only comprehensively evaluate the current health status of an individual, but also predict future risk change trends, providing a scientific basis for the formulation of intervention measures. The core goal of the module design is to achieve personalized, dynamic and forward-looking health management.

[0192] Based on the individualized risk index and high-risk factor characteristics in the dynamically adjusted intervention strategy data, the risk change trend in the future is predicted. Input features: individual current risk index and dynamic intervention history data, such as daily step changes, BMI changes, etc. Output results: future risk curve, showing the change trend of individual risk in a specific time range in the future (such as half a year, a year). Provide data support for planning intervention measures in advance. For example, the risk curve shows that the risk may increase in the next three months, and a stricter health management plan can be formulated in advance.

[0193] Short-term trend analysis focuses on the health changes of individuals after recent interventions, and aims to evaluate the intervention effect by analyzing the magnitude and direction of recent risk index changes. Input features: recent risk index change records, high-risk factor feature change records (such as changes in daily steps and dietary intake in the past month). Output results: short-term trend analysis results, including risk change rate and intervention effect evaluation. For example, if an individual has increased his exercise volume recently and the risk change rate has decreased, the system can feedback the positive effect of exercise intervention.

[0194] According to the individualized risk index and high-risk factor characteristics, individuals are screened in layers (such as low-risk, medium-risk, and high-risk groups). Input characteristics: risk index and specific high-risk factor annotations. Output results: stratified screening results and a list of key high-risk factors. Provide guidance for the allocation of medical resources and intervention priorities. For example, high-risk individuals may need more medical resource support, while low-risk individuals can maintain existing intervention measures.

[0195] Based on the results of long-term risk prediction, short-term trend analysis and individual stratified screening, a personalized prediction report is generated. The report content includes: a summary of dementia risk assessment; long-term trend analysis results, including changes in the risk curve; short-term trend analysis results; individual stratification results and a list of key high-risk factors; and specific intervention recommendations for individuals (such as adjusting diet, increasing exercise, drug treatment, etc.).

[0196] The module provides scientific risk assessment and intervention effect assessment tools through long-term prediction and short-term trend analysis, making health management more accurate and efficient. Personalized prediction reports can provide tailored health management plans for each individual, enhancing the pertinence of intervention measures. Through individual stratified screening, the module can prioritize medical resources according to risk levels and provide more support for high-risk individuals. Short-term trend analysis can reflect the intervention effect in real time and provide a basis for adjusting intervention strategies.

[0197] In the embodiment, after an individual receives the dementia risk prediction system service, the prediction and generation module of the system generates the following results: Input data, dynamically adjusted intervention strategy data: current individualized risk index ; BMI was 28 (high); daily step count was 4000 (low activity); current smoking status. Intervention history data: step count increased from 2000 to 4000, BMI decreased from 30 to 28 in the past month.

[0198] Long-term risk prediction: The module predicts the risk curve for the next 6 months based on the input characteristics: after 1 month, the risk index will drop to 1.6; after 3 months, the risk index will stabilize at 1.4; after 6 months, the risk index will further drop to 1.2. The prediction results show that under the existing intervention measures, the risk is expected to gradually decrease.

[0199] Short-term trend analysis shows that the risk index has dropped from 2.0 to 1.8 in the past month, a decrease of 10%. The increase in the number of steps contributed the most to the risk reduction, reaching 50%, the contribution of dietary adjustment was 30%, and the contribution of BMI reduction was 20%. The analysis results show that exercise intervention is the most effective measure at present.

[0200] Individual stratified screening: Based on the current risk index and high-risk factor characteristics, individuals are stratified into the "medium risk" group. Key high-risk factors include BMI, number of steps, and smoking status. The screening results recommend further reducing BMI and quitting smoking.

[0201] Personalized forecast report generation. The report generated by the system includes the following: Risk Assessment Summary: The individual is currently at medium risk with a risk index of 1.8.

[0202] Long-term trend analysis: The risk index will gradually decrease to 1.2 in the next 6 months, indicating that existing intervention measures are relatively effective.

[0203] Short-term trend analysis: The risk index decreased by 10% in the past month, mainly due to exercise intervention.

[0204] Stratification results and key high-risk factors: medium-risk group, high-risk factors are BMI, number of steps and smoking status.

[0205] Intervention recommendations: Increase the number of steps per day to 6,000, further improve diet (reduce high-fat intake), and use medication to assist in quitting smoking.

[0206] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A dementia risk prediction system based on an associative information processing model, characterized in that: include: A data processing module is used to collect multimodal data, including genetic data, health data, behavioral data, and environmental data, and process all the data to generate latent variable data; A causal reasoning module is used to construct a causal model based on latent variable data, calculate causal weights and generate an individualized risk feature matrix, wherein the feature matrix is ​​generated by comprehensively analyzing the weighted contributions of different causal relationship variables; based on the feature matrix, a nonlinear regression algorithm is applied to calculate individualized risk index data to reflect the likelihood of dementia for each individual; The intervention optimization module is used to build a reinforcement learning environment based on individualized risk index data, generate dynamically adjusted intervention strategy data through a deep Q-network-based optimization algorithm, and use a reinforcement learning framework to combine real-time updated risk indicators and intervention feedback to optimize the intervention path to reduce the risk of dementia; The prediction and generation module is used to perform long-term risk prediction, short-term trend analysis and individual stratification screening based on dynamically adjusted intervention strategy data, and generate personalized prediction report data. The personalized prediction report includes risk assessment and intervention recommendations.

2. The dementia risk prediction system based on the association information processing model according to claim 1, characterized in that: The data processing module includes a missing data filling unit, which processes the missing values ​​in the multimodal data by the following steps: For continuous numerical data, mean interpolation is used to fill in missing values ​​based on the mean values ​​in other data records; For categorical variables, the nearest neighbor algorithm-based imputation method is used to select the most common category value in similar data records for imputation; The filled multimodal data generates initial latent variable data.

3. The dementia risk prediction system based on the association information processing model according to claim 2, characterized in that: The data processing module performs time series analysis on the behavior data and the environmental data, extracts trend features and updates the latent variable data through the following steps: The behavioral data and environmental data are divided into time windows of fixed length, and the features of the data in each time window are calculated independently; Calculate the mean of the data in each time window to generate the mean parameter in the trend feature; Calculate the rate of change of the mean values ​​in adjacent time windows to generate the rate of change parameter in the trend feature; Applying linear regression analysis to the data in the time window, calculating the trend slope parameter, and generating the slope parameter in the trend feature; The initial latent variable data is updated based on the extracted trend features to obtain updated latent variable data.

4. The dementia risk prediction system based on the association information processing model according to claim 1, characterized in that: The individualized risk signature matrix generated by the causal reasoning module is calculated by the following steps: Determine the causal weight and calculate it using the following formula: in, Representation characteristics Features The causal weight of Features and Features The covariance of Features The variance of The eigenvalues ​​in the feature matrix are calculated based on the causal weights, which is achieved through the following formula: in, is the eigenvalue of the matrix, Representation characteristics In the characteristic matrix, the eigenvalues The contribution weight of is the latent variable feature, and n is the total number of all feature variables involved in the data calculation process.

5. The dementia risk prediction system based on the association information processing model according to claim 4 is characterized in that: The causal reasoning module optimizes the feature matrix in combination with the association rules, and is implemented by the following steps: Use the Apriori algorithm to mine association rules between features from latent variable data; Adjust the feature matrix according to the mined rules and weight the association rules to causal weights; The adjusted feature matrix is ​​regenerated by combining the causal weights and the association weights.

6. The dementia risk prediction system based on the association information processing model according to claim 1, characterized in that: The individualized risk index is calculated by the following formula: in, represents the individual risk index, Features The regression coefficient of is the latent variable eigenvalue, is the constant term of regression, and n is the total number of all characteristic variables involved in the data calculation process.

7. The dementia risk prediction system based on the association information processing model according to claim 1, characterized in that: The reinforcement learning environment of the intervention optimization module consists of the following parts: The state space consists of individualized risk index data and high-risk factor annotations; Action sets, including dietary modifications, exercise recommendations, and medications; The reward function is calculated by the following formula: in, Represents the reward value, is the change in risk index before and after intervention, The economic and time costs of the intervention.

8. The dementia risk prediction system based on the association information processing model according to claim 7, characterized in that: The optimization algorithm of the reinforcement learning adopts a deep Q network, and the training process includes the following steps: Initialize the parameters of the reinforcement learning model and determine the initial state through individualized risk index data; use - Greedy strategies that balance between exploring new interventions and exploiting existing ones; The Q value is updated in multiple rounds of iterations, and the intervention path is continuously optimized through the reward function.

9. The dementia risk prediction system based on the association information processing model according to claim 1, characterized in that: The long-term risk prediction task of the prediction and generation module is implemented through a recurrent neural network. The input features include the individualized risk index and high-risk factor annotations in the intervention strategy data after dynamic adjustment, and the output is a risk curve in the future.

10. The dementia risk prediction system based on the association information processing model according to claim 9, characterized in that: The personalized prediction report includes the following contents: Dementia Risk Assessment Summary; Long-term trend analysis results, including changes in the risk curve; short-term trend analysis results; individual stratification results and key high-risk factors; and specific intervention recommendations for individuals.

Images