Optometric health data monitoring and intelligent analysis management system

Abstract

The present application relates to a kind of visual health data monitoring and intelligent analysis management system, specifically, to the field of visual health management, the scheme generates prospective dynamic risk image by time series hybrid model, provides accurate basis for personalized intervention;Strategy network is based on this image and real-time context, generates and pushes personalized intervention action sequence under meeting medical constraint, realizes individualized instant management;Through the federal aggregation of coordination multiple client nodes, under the premise of protecting data privacy, intervention effect feedback data is used to continuously evolve strategy network;With long cycle value network, long-term effect is evaluated, and meta-optimization adjustment to key parameters such as prediction window, reward function weight and aggregation frequency is triggered;The whole process forms a closed loop from accurate prediction, personalized decision, cooperative evolution to macroscopic tuning, so as to realize the continuous adaptive optimization of eye health management, and significantly improve the accuracy, individualization degree and long-term effectiveness of management.

Description

Technical Field

[0001] This invention relates to the field of visual health management, and more specifically, to a visual health data monitoring and intelligent analysis management system. Background Technology

[0002] In the field of optometric health management, especially for myopia prevention and control in children and adolescents, long-term monitoring of complications of high myopia, and risk follow-up for patients with chronic eye diseases such as glaucoma, it is crucial to build continuous personal eye health records and implement dynamic interventions. The core biological parameters of these populations, such as visual acuity, refractive error, and axial length, will continuously change with growth and development, intensity of eye use, and treatment adherence. Their daily eye use behavior, near-field workload, and outdoor activity time are also in dynamic fluctuation. Therefore, an ideal health management system not only needs to integrate multi-dimensional static data, but also needs to have the ability to deeply analyze users' longitudinal time-series data to gain insight into the evolution trend of eye health status and predict risk trajectories. Based on this, it can generate personalized intervention strategies that can adaptively adjust according to the user's status and feedback, thereby forming a precise, forward-looking, and sustainable management closed loop.

[0003] For example, an eye health management system disclosed in Chinese patent literature (publication number: CN120932883A) uses a combination of data acquisition module, data processing module, data recognition module and data transmission module to detect key indicators of patient vision data, intraocular pressure and refractive error by integrating smart wearable devices and eye examination instruments, replacing the single detection method. At the same time, the integration of smart wearable devices and eye examination instruments also saves staff from the manual recording of test data, thereby improving the accuracy of patient vision test data.

[0004] However, this existing technical solution has significant limitations in achieving truly personalized, long-term dynamic health management. Its data analysis module primarily focuses on identifying and classifying collected data, and the resulting health management plan is essentially a one-time static output based on the user's current and historical status. The system lacks the ability to deeply model and predict the temporal patterns of user eye health indicators, such as the rate of axial length growth and annualized changes in refractive error, and cannot proactively assess the risk level a user might face at a specific future time. More importantly, after the system pushes out the management plan, it fails to establish an effective feedback and optimization mechanism, making it impossible to obtain information about the user's feedback on the suggestions. Furthermore, the system cannot quantify or assess the actual effects of the implemented interventions, such as increasing outdoor activity time or providing vision training. This makes it difficult for the system to dynamically adjust and continuously optimize subsequent management strategies based on intervention effects and user behavior feedback, resulting in a one-way and static health management process. For children in the rapid progression phase of myopia, the system cannot provide early warnings and adjust the intensity of interventions before significant vision deterioration. For patients requiring long-term follow-up, their management plans cannot be finely adjusted as the disease progresses, which may ultimately lead to delayed intervention, reduced user compliance, and significantly diminished long-term health management effectiveness. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing a vision health data monitoring and intelligent analysis management system. Through a data processing module, a strategy generation module, a model evolution module, and a meta-optimization module, it solves the problems mentioned in the background art.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes a data processing module, a strategy generation module, a model evolution module, and a meta-optimization module that are sequentially connected in communication, wherein; Data processing module: Upon receiving historical time-series monitoring data from a user, the module standardizes the data and inputs the processed data into a pre-trained time-series mixture model to output predicted values ​​of eye health indicators and dynamic risk probability distributions for multiple future time points, thereby generating a dynamic risk profile for the user. The strategy generation module: Based on the dynamic risk profile, real-time user behavior data obtained from the user terminal, and environmental context, it constructs the current decision state, calls the strategy network to output a short-term personalized intervention action sequence based on the current decision state and under the condition of meeting preset medical constraints, and pushes the personalized intervention action sequence to the user terminal. Model Evolution Module: When the preset federated aggregation cycle is triggered, it coordinates multiple client nodes to iteratively update the local policy network copy based on the feedback data of the intervention effect after the local user executes the personalized intervention action sequence. The local policy network copy is a local copy of the policy network on each client node. It adopts an adaptive federated aggregation algorithm based on contribution evaluation to perform weighted aggregation of the parameter update amount of the local policy network copy of all client nodes to generate a global parameter update amount. The global parameter update amount is then used to update the policy network. The updated policy network parameters are then distributed back to each client node to synchronously update its local policy network copy. Meta-optimization module: continuously collects new monitoring data and behavioral feedback data after users perform personalized intervention action sequences, evaluates the long-term effect of the currently used strategy network through a long-term value network, and dynamically adjusts at least one of the following when the evaluation result is lower than a preset threshold: the prediction window of the time series hybrid model, the weight of the reward function of the strategy network, or the aggregation frequency of the adaptive federated aggregation algorithm.

[0007] In a preferred embodiment, the specific process of the time-series hybrid model processing the standardized historical time-series monitoring data in the data processing module is as follows: After standardizing the historical time-series monitoring data, the standardized historical time-series monitoring data is simultaneously input into the local trend capture branch, periodic pattern discovery branch, and global dependency modeling branch of the time-series hybrid model for parallel feature extraction. Among them, the local trend capture branch extracts local change pattern features at different time scales from the standardized historical time series monitoring data through causal dilated convolutional layers with different dilation coefficients. The periodic pattern discovery branch calculates the response intensity of the historical time-series monitoring data in different potential periodic dimensions by cross-correlation calculation with a learnable periodic basis function library after standardization of the historical time-series monitoring data, and generates periodic pattern features by weighting them; the global dependency modeling branch calculates the global dependency features between features of all time points in the standardized historical time-series monitoring data by introducing a multi-head self-attention mechanism based on relative position bias with time distance decay. Subsequently, the temporal hybrid model uses an adaptive weight fusion layer to dynamically calculate and assign fusion weights to the local change pattern features, periodic pattern features, and global dependency features at each time step, and outputs the weighted sum as a fused multi-scale feature representation sequence.

[0008] In a preferred embodiment, the specific process of generating the dynamic risk profile is as follows: The fused multi-scale feature representation sequence is input into a state space model layer contained in the temporal hybrid model. The state space model layer models the long-term dynamic dependence of the sequence through its internal structured state equations and outputs the final hidden state of the sequence. Based on the final sequence hidden state, the predicted values ​​of eye health indicators at multiple future time points are calculated through the prediction layer included in the temporal mixture model. Meanwhile, during the inference phase, random discarding is applied multiple times to the local trend capture branch, the periodic pattern discovery branch, the global dependency modeling branch, and the state space model layer to generate multiple sets of eye health indicator prediction values ​​for future time points. Based on the multiple sets of eye health indicator prediction values ​​for future time points, the empirical distribution of the prediction value for each future time point is calculated, and the probability distribution of the eye health risk level for each future time point is output through the risk quantification layer contained in the time series hybrid model. The dynamic risk probability distribution is composed of the probability distribution of eye health risk levels at all future time points; The dynamic risk profile consists of predicted values ​​of eye health indicators at multiple future time points, a dynamic risk probability distribution, and a prediction interval width calculated based on multiple sets of predicted values ​​of eye health indicators at future time points to characterize the uncertainty of the prediction.

[0009] In a preferred embodiment, the specific process of constituting the current decision state and invoking the policy network in the policy generation module is as follows: First, the dynamic risk profile is input into the profile encoder. The profile encoder extracts feature vectors representing future risk trends from the predicted values ​​of eye health indicators at multiple future time points, the dynamic risk probability distribution, and the prediction interval width contained in the dynamic risk profile through an attention mechanism. Simultaneously, the real-time behavior data of the user obtained from the user terminal and the environmental context are input into the context encoder to obtain the real-time context feature vector; Next, the feature vector representing future risk trends extracted by the portrait encoder is concatenated with the real-time context feature vector obtained by the context encoder to form a fused state vector, and the fused state vector is input into a preset personalized state memory unit. The personalized state memory unit updates itself based on the current fused state vector and the user's historical state memory stored within it, and outputs the state memory vector at the current moment. Finally, the fused state vector is concatenated with the current state memory vector to form the current decision state, and this current decision state is input into the policy network.

[0010] In a preferred embodiment, the specific process of the policy network outputting and pushing personalized intervention action sequences is as follows: The policy network consists of a shared feature layer, multiple parallel action generation subnetworks, a real-time constraint verification and correction layer, and a template engine; The shared feature layer processes the current state of the input decision and extracts high-level features; Each action generation subnetwork corresponds to a future time step. Based on the high-level features extracted from the shared feature layer, it generates intervention action parameters for that time step. The intervention action parameters include a mean vector that defines the center value of the action and a variance vector that defines the range of action exploration. When generating personalized intervention action sequences, the mean vector of each action generation subnetwork is taken and arranged in the order of the corresponding time steps to form the initial action sequence generated by the policy network. The real-time constraint verification and correction layer receives the initial action sequence. Under the premise of satisfying all preset medical inequality constraints and equality constraints, it calculates the final personalized intervention action sequence by solving an optimization problem with the objective of minimizing the weighted Euclidean distance between the initial action sequence and the corrected action sequence. The template engine decodes the final personalized intervention action sequence into natural language instructions, generates the final push message, and sends it to the user's terminal.

[0011] In a preferred embodiment, the model evolution module includes a central server and multiple client nodes; the specific process of coordinating multiple client nodes to iteratively update the local policy network copy based on the intervention effect feedback data after local users execute personalized intervention action sequences is as follows: At the start of each preset federated aggregation cycle, the central server distributes the current policy network parameters as global policy network parameters to each client node. Each client node initializes its local policy network copy with the received global policy network parameters and trains the local policy network copy with the intervention effect feedback dataset stored locally. During training, a loss function that combines the policy optimization objective and the contrastive regularization term is used. The policy optimization objective is to enable the intervention action sequences generated by the local policy network replicas to obtain higher cumulative rewards on the intervention effect feedback dataset; the contrast regularization term is used to encourage the local policy network replicas to learn the overall consistency with the output actions of the current global policy network, while strengthening their ability to distinguish and retain those distinctive actions in the intervention effect feedback dataset that perform well but are different from the average output of the global policy. After completing a preset number of local training cycles, each client node calculates the difference between the parameters of its local policy network replica and the initially received global policy network parameters. This difference is used as the parameter update amount for the client node's local policy network replica. This parameter update amount, along with the performance improvement metrics, dataset size and distribution characteristics, and training stability metrics recorded during local training, are sent to the central server as local training metadata.

[0012] In a preferred embodiment, the specific process of using an adaptive federated aggregation algorithm based on contribution evaluation to weighted aggregate the parameter update amounts of the local policy network replicas of all client nodes is as follows: The central server maintains a historical contribution reputation score for each client node, which is a dynamically updated score based on the quality and consistency of the client node's past updates. After receiving the parameter updates and local training metadata from all client nodes, the central server first constructs a contribution feature vector for each client node that comprehensively reflects the performance improvement indicators, data value, and update quality in its local training metadata. Using a multi-head cross-attention weight generation mechanism, based on the contribution feature vectors of all client nodes and the historical contribution reputation scores of each client node, the central server dynamically calculates the initial weight of each client node in this round of aggregation. The initial weights calculated by the multi-head cross-attention weight generation mechanism are then normalized to ensure that the sum of all weights is one. Finally, before weighted aggregation, Mahalanobis distance anomaly detection is performed on all parameter update amounts based on multivariate statistical analysis. The weights of client nodes corresponding to parameter update amounts that deviate too much from the mainstream update direction are temporarily set to zero, and the remaining weights are renormalized to obtain the corrected weights used for the final weighted aggregation.

[0013] In a preferred embodiment, the specific process of updating the policy network using global parameter update values ​​and subsequent synchronization includes: The central server maintains a small-scale public benchmark verification set containing diverse scenarios; the central server uses corrected weights to perform a weighted summation of the parameter update amounts of each client node that passes anomaly detection to obtain the global parameter update amount; Using a dynamically adjusted global learning rate, the global parameter update amount is added to the current global policy network parameters to complete the policy network update; Afterwards, the central server uses the maintained public benchmark verification set to evaluate the improvement in policy network performance before and after this round of updates, and updates the historical contribution reputation of each client node based on this improvement and the directional consistency between the parameter update amount of each client node and the global parameter update amount. The central server distributes the updated policy network parameters back to all client nodes participating in this round of aggregation. Each client node then replaces its local policy network copy parameters with these updated parameters, achieving synchronous updates.

[0014] In a preferred embodiment, the meta-optimization module includes a trajectory encoder and a long-period value network; the specific process of evaluating the long-term effect of the currently used policy network through the long-period value network is as follows: For each user, new monitoring data and behavioral feedback data are collected after they execute a personalized intervention action sequence within a preset long-term time window. Based on the new monitoring data and behavioral feedback data, a long-term trajectory of the user is constructed. This long-term trajectory includes the current decision state of the user at each time point within the long period, which is composed of the strategy generation module; the actions in the final personalized intervention action sequence generated and pushed by the strategy generation module; the instant reward corresponding to the strategy network calculation in the strategy generation module; and the current decision state corresponding to the dynamic risk profile of the next time point generated by the data processing module based on the new monitoring data. The constructed long-term trajectory of the user is input into the trajectory encoder. The trajectory encoder is based on the attention mechanism. It calculates the importance weight of the joint feature of the current decision state and the immediate reward at each time step in the long-term trajectory, and performs weighted fusion of the encoding of all time steps in the long-term trajectory to output the context-aware representation of the trajectory. The context-aware representation of the trajectory is input into the long-term value network. The long-term value network outputs a high-dimensional value hidden state vector, and maps this value hidden state vector to a scalar through a linear projection layer as the long-term value evaluation value of the user under the current policy network. Calculate the average of the long-term value assessments of all active users as the overall long-term performance assessment of the currently used strategy network. The overall long-term effect assessment value is compared with a preset threshold. If it is lower than the preset threshold, the long-term effect of the currently used strategy network is determined to be lower than expected.

[0015] In a preferred embodiment, the meta-optimization module further includes a historical adjustment record database and a dynamic adjustment controller; the specific process of dynamically adjusting at least one of the prediction window of the time-series hybrid model, the reward function weights of the policy network, or the aggregation frequency of the adaptive federated aggregation algorithm is as follows: After each evaluation, the set of the mean values ​​of the hidden state vectors of all active users, the prediction window of the current time series hybrid model, the reward function weights of the policy network, and the aggregation frequency of the adaptive federated aggregation algorithm are recorded as the current meta-parameter configuration. When a new overall long-term effect evaluation value is subsequently calculated, the change between the overall long-term effect evaluation value and the previously recorded overall long-term effect evaluation value is associated with the corresponding meta-parameter configuration record and stored in the historical adjustment record database. When it is determined that the long-term effect of the currently used strategy network is lower than expected, the dynamic adjustment controller initiates the adjustment process; the dynamic adjustment controller first reads all historical records from the historical adjustment record database; Using all historical records, a Gaussian process regression model is trained. This model takes as input the set of mean values ​​of the unadjusted user value latent state vectors stored in the historical records and the candidate meta-parameter configurations, and predicts the expected and variance of the overall long-term effect evaluation value that may be obtained after adjustment. Next, within the search space that satisfies the preset medical feasibility and meta-parameter configuration, the optimal candidate meta-parameter configuration is searched by maximizing an acquisition function; the acquisition function is the expected value predicted by the Gaussian process regression model minus the current overall long-term effect evaluation value divided by the predicted variance, plus an exploration coefficient multiplied by the predicted variance. The dynamic adjustment controller conducts A / B testing on the optimal candidate meta-parameter configuration obtained from the search in a phased deployment manner. During the test, the causal forest model is used to analyze the difference in long-term value assessment values ​​between the experimental group and the control group, and to assess the heterogeneity of this difference under different user characteristics. If the A / B test results show that the optimal candidate meta-parameter configuration has a statistically significant positive overall effect in the experimental group and no widespread and serious negative effect is found on the specified user group, the dynamic adjustment controller will gradually expand the deployment of the optimal candidate meta-parameter configuration to all users, thus completing the update of the meta-parameter configuration. Meanwhile, the dynamic adjustment controller continuously monitors three metrics: the rate of decline of the overall long-term performance evaluation value within the sliding window, the statistical distance between the mean set of the current user value hidden state vector and the historical normal set, and the amount of communication resources consumed per unit performance improvement during federated learning. When the value of any monitoring metric exceeds its corresponding adaptive calculation threshold, the dynamic adjustment controller also initiates the adjustment process.

[0016] The beneficial effects of this invention are as follows: A forward-looking dynamic risk profile is generated through a temporal hybrid model, providing precise evidence for personalized intervention; based on this profile and real-time context, the strategy network generates and pushes personalized intervention action sequences while meeting medical constraints, achieving individualized real-time management; by coordinating the federated aggregation of multiple client nodes, the strategy network continuously evolves using intervention effect feedback data while protecting data privacy; long-term effects are evaluated using a long-cycle value network, triggering meta-optimization adjustments to key parameters such as prediction window, reward function weight, and aggregation frequency; the entire process forms a closed loop from accurate prediction, personalized decision-making, co-evolution to macro-optimization, thereby achieving continuous adaptive optimization of eye health management and significantly improving the accuracy, personalization, and long-term effectiveness of management. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application. Example

[0021] This embodiment provides, for example Figure 1-2 The vision health data monitoring and intelligent analysis management system shown includes: a data processing module, a strategy generation module, a model evolution module, and a meta-optimization module connected in sequence via communication; wherein; Data processing module: Upon receiving historical time-series monitoring data from a user, the module standardizes the data and inputs the processed data into a pre-trained time-series mixture model to output predicted values ​​of eye health indicators and dynamic risk probability distributions for multiple future time points, thereby generating a dynamic risk profile for the user. The strategy generation module: Based on the dynamic risk profile, real-time user behavior data obtained from the user terminal, and environmental context, it constructs the current decision state, calls the strategy network to output a short-term personalized intervention action sequence based on the current decision state and under the condition of meeting preset medical constraints, and pushes the personalized intervention action sequence to the user terminal. Model Evolution Module: When the preset federated aggregation cycle is triggered, it coordinates multiple client nodes to iteratively update the local policy network copy based on the feedback data of the intervention effect after the local user executes the personalized intervention action sequence. The local policy network copy is a local copy of the policy network on each client node. It adopts an adaptive federated aggregation algorithm based on contribution evaluation to perform weighted aggregation of the parameter update amount of the local policy network copy of all client nodes to generate a global parameter update amount. The global parameter update amount is then used to update the policy network. The updated policy network parameters are then distributed back to each client node to synchronously update its local policy network copy. Meta-optimization module: continuously collects new monitoring data and behavioral feedback data after users perform personalized intervention action sequences, evaluates the long-term effect of the currently used strategy network through a long-term value network, and dynamically adjusts at least one of the following when the evaluation result is lower than a preset threshold: the prediction window of the time series hybrid model, the weight of the reward function of the strategy network, or the aggregation frequency of the adaptive federated aggregation algorithm.

[0022] In this embodiment, it is specifically necessary to explain the following process in the data processing module: the time-series hybrid model processes the standardized historical time-series monitoring data as follows: After standardizing the historical time-series monitoring data, the standardized data is simultaneously input into the local trend capture branch, periodic pattern discovery branch, and global dependency modeling branch of the time-series hybrid model for parallel feature extraction. The historical time-series monitoring data specifically includes spherical power, cylindrical power, axis, uncorrected visual acuity, corrected visual acuity, axial length, corneal curvature, and intraocular pressure values ​​collected from users at different time points, as well as daily near-vision usage time, outdoor activity time, sleep duration, and screen time behavior data extracted from user devices or questionnaires. The standardization process is as follows: for numerical data, Z-Score normalization is used, i.e., subtracting the mean of the indicator calculated from all users' historical data and then dividing by its standard deviation, mapping it to a range of zero mean and unit variance; for categorical data, one-hot encoding is used for vectorization representation. The local trend capture branch extracts local change pattern features at different time scales from standardized historical time-series monitoring data using causal dilated convolutional layers with different dilation coefficients. The local trend capture branch contains three parallel causal dilated convolutional layers with dilation coefficients of 1, 7, and 30, corresponding to daily, weekly, and monthly time scales, respectively. The calculation process for each causal dilated convolutional layer is as follows: a one-dimensional convolutional kernel is defined, which only performs a weighted summation of the input data from the current time step and previous time steps. The dilation coefficient determines the number of time steps skipped by the convolutional kernel when traversing the input sequence. For a convolutional layer with a dilation coefficient of 7, each output feature point is the result of convolution between the current input point and every six previous input points, thus enabling pattern capture across a one-week time window. The periodic pattern discovery branch performs cross-correlation calculations between standardized historical time-series monitoring data and a learnable periodic basis function library to obtain the response strength of the historical time-series monitoring data at different potential periodic dimensions, and then generates periodic pattern features through weighted calculations. The learnable periodic basis function library of the periodic pattern discovery branch consists of a set of learnable vectors with lengths of seven, thirty, ninety, and three hundred and sixty-five, corresponding to weekly, monthly, quarterly, and annual periodic assumptions, respectively. The cross-correlation calculation process is as follows: the input historical time-series monitoring data sequence is subjected to a sliding dot product operation with each periodic basis vector to obtain a response sequence, which reflects the strength of repetition of the input data at the corresponding period length. The response strength is a scalar value obtained by processing the local maxima of the response sequence through a nonlinear activation function. The process of weighted generation of periodic pattern features is as follows: each periodic basis vector is multiplied by its corresponding response strength scalar, all weighted periodic basis vectors are summed, and finally, a fully connected layer is used for feature transformation. The global dependency modeling branch introduces a multi-head self-attention mechanism based on relative positional bias with temporal distance decay to calculate the global dependency features among all time-point features in the standardized historical time-series monitoring data. The calculation process of relative positional bias based on temporal distance decay is as follows: first, calculate the absolute value of the time difference between any two time points in the sequence; then, input the absolute value of the time difference into a negative logarithmic function for calculation; then, multiply the calculation result by a learnable slope parameter; finally, invert the result to obtain a bias value that is inversely proportional to the time distance. The calculation process of the multi-head self-attention mechanism is as follows: First, the input feature sequence is generated into a query vector group, a key vector group, and a value vector group through different linear projection matrices; then, the query vector group and the key vector group are multiplied by a dot product to obtain the original attention score; next, the calculated relative position bias is added to the original attention score; then, the score of each row is calculated using a normalized exponential function to transform the score into a probability distribution form of attention weights; finally, the attention weights are used to perform a weighted summation of the value vector group to obtain the output of the current head; the outputs of all attention heads are concatenated and fused through a linear projection layer to obtain the global dependency features; Subsequently, the temporal hybrid model uses an adaptive weight fusion layer to dynamically calculate and assign fusion weights to the local variation pattern features, periodic pattern features, and global dependency features at each time step, and outputs the weighted sum as the fused multi-scale feature representation sequence. The calculation process of the adaptive weight fusion layer is as follows: First, the local variation pattern features, periodic pattern features, and global dependency features at the same time step are concatenated into vectors; then, the concatenated vector is input into a two-layer feedforward neural network. The first layer of this network uses a linear transformation and a non-linear activation function, while the second layer uses a linear transformation; next, the three scalars output by the second layer are normalized using an exponential function to calculate the sum of the three scalars to one. These three scalars are the dynamic fusion weights corresponding to the three types of features at the current time step; finally, the three types of features are multiplied by their corresponding fusion weights, and then vector addition is performed to obtain the fused multi-scale feature representation at that time step. The specific process for generating a dynamic risk profile is as follows: The fused multi-scale feature representation sequence is input into a state-space model layer within the temporal mixture model. This state-space model layer models the long-term dynamic dependencies of the sequence through its internal structured state equations and outputs the final hidden state of the sequence. The computation process of the state-space model layer is as follows: it maintains a hidden state vector internally. When processing the input at each time step, two core computations are performed. The first computation is updating the hidden state: multiplying the current hidden state vector by a learnable state transition matrix, then multiplying the input features of the current time step by a learnable input projection matrix, adding these two products, and then passing them through a non-linear activation function to obtain the updated hidden state vector. The second computation is generating the output: multiplying the updated hidden state vector by a learnable output projection matrix, and then adding the product of the current input features and another learnable through projection matrix to obtain the output features of the current time step. After processing the entire input sequence, the updated hidden state vector at the last time step is the final hidden state of the sequence. Based on the final sequence hidden state, the predicted values ​​of eye health indicators for multiple future time points are calculated through the prediction layer included in the temporal mixture model. The prediction layer is a fully connected neural network, and its calculation process is as follows: the final sequence hidden state is input into the network, which contains a hidden layer and an output layer; the hidden layer performs a linear transformation on the input and then passes it through a non-linear activation function; the output layer performs a linear transformation on the output of the hidden layer, directly outputting the predicted values ​​of the two core indicators, spherical power and axial length, for the next twelve time points (corresponding to the next twelve months). Simultaneously, during the inference phase, random discarding is applied multiple times to the local trend capture branch, periodic pattern discovery branch, global dependency modeling branch, and state space model layer to generate multiple sets of predicted eye health indicators for future time points. Based on these multiple sets of predicted eye health indicators, the empirical distribution of each predicted value for future time points is calculated, and accordingly, the probability distribution of eye health risk level for each future time point is output through the risk quantification layer included in the temporal mixture model. The application process of random discarding is as follows: during model inference, 20% of the neurons in the above branches and layers are temporarily and randomly set to zero, and then forward propagation is performed to obtain a set of predicted values. Repeat this process fifty times to obtain fifty sets of predicted eye health indicators for future time points; the process of calculating the empirical distribution is as follows: for a specific month in the future, collect the predicted spherical diopter values ​​for that month obtained from fifty inferences, and sort these fifty values ​​from smallest to largest; the risk quantification layer is a multilayer perceptron, and its calculation process is as follows: the final sequence hidden state is concatenated with the encoding vector of the future time point to be evaluated, and input into a two-layer neural network. The output layer of this network uses a normalized exponential function to output three probability values ​​corresponding to "low risk", "medium risk" and "high risk" respectively, and the sum of these three probability values ​​is one; The dynamic risk probability distribution is composed of the probability distribution of eye health risk levels at all future time points; The dynamic risk profile consists of predicted values ​​of eye health indicators at multiple future time points, a dynamic risk probability distribution, and a prediction interval width calculated based on multiple sets of predicted values ​​of eye health indicators at future time points to characterize the uncertainty of the prediction. The calculation process of the prediction interval width is as follows: for each specific month in the future, from the fifty ranked predicted values ​​of spherical diopter, the value ranked forty-eighth is taken as the upper limit of the interval, and the value ranked third is taken as the lower limit of the interval. The difference between the upper limit and the lower limit of the interval is the prediction interval width for that month.

[0023] In this embodiment, it is specifically necessary to explain the process of constructing the current decision state and invoking the policy network in the policy generation module as follows: First, the dynamic risk profile is input into the profile encoder. The profile encoder extracts feature vectors representing future risk trends from the predicted eye health indicators, dynamic risk probability distribution, and prediction interval widths at multiple future time points contained in the dynamic risk profile using an attention mechanism. The profile encoder employs a multi-head attention mechanism for calculation. Specifically, the predicted eye health indicator value, the risk level probability vector at that time point, and the prediction interval width scalar are concatenated to form the "value" vector for that time point. Simultaneously, a learnable "query" vector is generated for the entire dynamic risk profile. Then, the dot product of the "query" vector and the "value" vector at each time point is calculated to obtain the original attention score. Next, the reciprocal of the prediction interval width at each time point is taken as the prior weight and multiplied by the original attention score, so that time points with lower uncertainty receive higher weights. Subsequently, a normalization exponent is calculated on the weighted scores to obtain the final attention weight for each time point. Finally, these weights are used to perform a weighted summation of the "value" vectors at all time points, outputting a fixed-length feature vector, which is the feature vector representing the future risk trend. Simultaneously, real-time user behavior data obtained from the user terminal and the environmental context are input into the context encoder to obtain a real-time context feature vector. The context encoder is a three-layer fully connected neural network. The specific process is as follows: First, the input real-time behavior data and environmental context are concatenated into vectors. The real-time behavior data includes: whether the current activity is near-field eye use, the duration of the current eye use, and the current device screen brightness level. The environmental context includes: the ambient light intensity level and whether the user is in a moving state. Next, the concatenated vector is input into the first fully connected layer, linearly transformed, and then passed through a non-linear activation function. The output is then input into the second fully connected layer, linearly transformed and activated again. Finally, the third fully connected layer (without an activation function) maps the dimensions to a preset feature dimension and outputs the real-time context feature vector. Next, the feature vector representing future risk trends extracted by the profile encoder is concatenated with the real-time context feature vector obtained by the context encoder to form a fused state vector, and the fused state vector is input into a preset personalized state memory unit; the personalized state memory unit is a gated recurrent unit network; the user's historical state memory stored inside it is reflected as the hidden state vector of the previous moment on the network. The personalized state memory unit updates based on the current fused state vector and the user's internally stored historical state memories, and outputs the current state memory vector. The specific calculation process for the update is as follows: First, an "update gate" vector is calculated, which determines how many historical memories are retained. Its value is between zero and one, obtained by concatenating the current fused state vector with the previous hidden state vector and passing it through a fully connected layer and a sigmoid activation function. Second, a "reset gate" vector is calculated, which determines how many historical memories are used to calculate new candidate memories, calculated in the same way as the update gate. Then, a "candidate hidden state" vector is calculated by concatenating the current fused state vector with the previous hidden state vector after being filtered by the reset gate, and passing it through a fully connected layer and a hyperbolic tangent activation function. Finally, the current hidden state vector is calculated, which is the weighted sum of the previous hidden state vector and the candidate hidden state vector. The weights are determined by one minus the update gate vector and the update gate vector, respectively. This final hidden state vector is the current state memory vector. Finally, the fused state vector is concatenated with the current state memory vector to form the current decision state, and this current decision state is input into the policy network. The specific process by which the strategy network outputs and pushes personalized intervention action sequences is as follows: The policy network consists of a shared feature layer, multiple parallel action generation subnetworks, a real-time constraint verification and correction layer, and a template engine; The shared feature layer processes the input decision state and extracts high-level features. The shared feature layer is a fully connected neural network with two hidden layers. Its processing is as follows: the current decision state vector is input into the first hidden layer, linearly transformed, and then passed through a non-linear activation function; the output is input into the second hidden layer, linearly transformed and activated again; finally, a high-level feature vector is output. Each action generation subnetwork corresponds to a future time step. Based on the high-level features extracted from the shared feature layer, it generates intervention action parameters for that time step. The intervention action parameters include a mean vector defining the action center value and a variance vector defining the action exploration range. Each action generation subnetwork is a fully connected neural network with the same structure. Its input is the high-level feature vector output by the shared feature layer. The network contains one hidden layer and one output layer. The hidden layer undergoes linear transformation and activation. The output layer undergoes linear transformation, and the output dimension is twice the dimension of the intervention action parameters. The first half is the mean vector, and the second half is the natural logarithm of the variance vector. The intervention action parameters have five dimensions, corresponding to intervention type, intensity, suggested duration, suggested start time, and additional parameters. When generating personalized intervention action sequences, the mean vector of each action generation subnetwork is taken and arranged in the order of the corresponding time steps to form the initial action sequence generated by the policy network. The real-time constraint verification and correction layer receives the initial action sequence. Under the premise of satisfying all preset medical inequality and equality constraints, it calculates the final personalized intervention action sequence by solving an optimization problem with the objective of minimizing the weighted Euclidean distance between the initial action sequence and the corrected action sequence. The optimization problem is constructed and solved as follows: the optimization variable is the final action sequence, which has the same dimension as the initial action sequence; the objective function is defined as the sum of squared differences between the parameters at each corresponding position between the initial and final action sequences, where each squared term is multiplied by a preset weight coefficient, which is set according to the importance of the parameters, with the intervention type and intensity having higher weights than the start time; inequality constraints include: the total daily intervention time should not exceed thirty minutes; high-intensity conditioning training is prohibited for high-risk users; equality constraints include: at least three outdoor activity suggestions must be included per week. The interior-point method is used to solve this optimization problem. The specific steps are as follows: First, slack variables are introduced to transform the inequality constraints into equality constraints and non-negativity constraints on the slack variables. Then, a Lagrangian function containing the objective function, constraints, and logarithmic barrier function is constructed. Next, the Carlow-Kuhn-Tucker equations corresponding to the Lagrangian function are solved using the Newton-Newton iteration method. In each iteration, the Newton direction is calculated, and a linear search is performed along this direction to determine the step size, and the optimization variables and Lagrangian multipliers are updated. The iteration continues until the gap between the original problem and the dual problem is less than one ten-thousandth, or the maximum number of iterations is reached one hundred. The optimization variables obtained at this time are the final personalized intervention action sequence. The template engine decodes the final personalized intervention action sequence into natural language instructions and, combined with the style preferences output by the user's historical preference model, selects a presentation template for the instructions, generates the final push message, and sends it to the user's terminal. The decoding and generation process is as follows: The template engine pre-stores multiple sets of natural language templates, each corresponding to the same intervention action logic but with different expression styles; the user's historical preference model outputs a style preference score based on the user's past clicks, reading time, and feedback on push messages, with the score ranging from zero to one. A higher score indicates a stronger preference for an encouraging and gamified style, while a lower score indicates a stronger preference for a rigorous and concise style; the template engine determines the basic template group to be called based on the action parameters in the final personalized intervention action sequence, and then selects the template with the closest score in that group based on the style preference score; next, the specific values ​​in the action parameters are filled into the reserved positions in the template; finally, the completed natural language instructions, the corresponding execution timestamp, and an interactive button for user feedback are encapsulated into a complete push message and sent to the user's terminal through the push message service.

[0024] In this embodiment, it should be specifically noted that the model evolution module includes a central server and multiple client nodes; the specific process of coordinating multiple client nodes to iteratively update the local policy network copy based on the intervention effect feedback data after local users execute personalized intervention action sequences is as follows: At the start of each preset federated aggregation cycle, the central server distributes the current policy network parameters as global policy network parameters to each client node. Each client node initializes its local policy network copy with the received global policy network parameters and trains the local policy network copy with the intervention effect feedback dataset stored locally. During training, a loss function that combines the policy optimization objective and the contrastive regularization term is used. The loss function is calculated as follows: First, the policy optimization objective term is calculated, which is the near-end policy optimization objective function. Its purpose is to maximize the sum of future expected rewards calculated on the local intervention effect feedback dataset. Specifically, for each interaction record in the dataset, the ratio of the probability of executing the action in the record under the current local policy network replica parameters to the probability of executing the action under the old policy parameters is calculated. This ratio is multiplied by the advantage function estimate corresponding to the record. Then, the product is pruned to ensure that the update magnitude is within a certain range. Finally, the negative value is taken as part of the policy loss. Simultaneously, a value function error term is calculated to evaluate the accuracy of state value estimation; the policy optimization objective term is composed of the weighted sum of the pruned policy loss and the value function error term; then, a contrastive regularization term is calculated, which measures the similarity between the action features output by the current local policy network replica and the action features output by the global policy network, and encourages the local policy to better distinguish those actions in the local data that perform well but differ from the average output of the global policy while maintaining consistency with the overall global policy. The specific calculation process is as follows: For each state in the local dataset, the feature vector of its output action is calculated using both the local policy network replica and the global policy network. The cosine similarity between these two feature vectors is calculated, and a high similarity is desired. Simultaneously, a different action is sampled from the dataset as a negative sample, and its action features are extracted. The cosine similarity between the action features output by the local policy network replica and the action features of this negative sample is calculated, and a low similarity is desired. These two similarities are compared using a normalized exponential function with a temperature coefficient to obtain a loss value. The contrast regularization term is the average of this loss value across all states. Finally, the total loss function is the policy optimization objective term plus a contrast regularization term with a weight coefficient of one-tenth. The policy optimization objective is to enable the intervention action sequences generated by the local policy network replicas to obtain higher cumulative rewards on the intervention effect feedback dataset; The contrast regularization term is used to enhance the ability of local policy network replicas to distinguish and retain distinctive actions in the intervention effect feedback dataset that perform well but differ from the average output of the global policy, while encouraging overall consistency between the learning of local policy network replicas and the output actions of the current global policy network. After completing a preset number of local training cycles, each client node calculates the difference between the parameters of its local policy network replica and the initially received global policy network parameters. This difference is used as the parameter update for the client node's local policy network replica. This parameter update, along with performance improvement metrics, dataset size and distribution characteristics, and training stability metrics recorded during local training, are sent to the central server as local training metadata. The performance improvement metric is the difference in the average reward of the policy on the local dataset before and after training. The dataset size and distribution characteristics include the total number of interaction records in the dataset, and a measure of the uniformity of distribution among different user age groups and different initial vision conditions. The training stability metric is the fitting decay coefficient of the loss function value's descent curve during training. The preset number of local training cycles is five. The specific process of using an adaptive federated aggregation algorithm based on contribution evaluation to weighted aggregate the parameter updates of the local policy network replicas of all client nodes is as follows: The central server maintains a historical contribution reputation score for each client node, which is a dynamically updated score based on the quality and consistency of the client node's past updates. After receiving the parameter updates and local training metadata from all client nodes, the central server first constructs a contribution feature vector for each client node that comprehensively reflects the performance improvement index, data value, and update quality in its local training metadata. The construction process of the contribution feature vector is as follows: First, the performance improvement index is normalized so that it falls between zero and one; for data value, a comprehensive score is calculated, which is the logarithm of the data size divided by ten, plus a measure of data distribution uniformity; for update quality, the L2 norm of the parameter update and the cosine similarity between the parameter update and the average direction of all parameter update are calculated; then, the four scalars—the normalized performance improvement index, the comprehensive score of data value, the L2 norm of the parameter update, and the cosine similarity of the direction—are concatenated into a four-dimensional vector, which is then mapped to a preset thirty-two-dimensional feature space through a three-layer fully connected neural network. The resulting vector is the contribution feature vector. Then, using a multi-head cross-attention weight generation mechanism, based on the contribution feature vectors of all client nodes and the historical contribution reputation of each client node, the initial weight of each client node in this round of aggregation is dynamically calculated. The calculation process of the multi-head cross-attention weight generation mechanism is as follows: Four independent attention heads are set up; for each attention head, a learnable query vector is defined; for each client node, its contribution feature vector is transformed into a key vector through a linear transformation matrix, and then into a value vector through another linear transformation matrix; then, the dot product of the query vector and the key vector of each client node is calculated to obtain an initial attention score; next, a bias term is calculated based on the historical contribution reputation of the client node, specifically by dividing the reputation by five, taking the hyperbolic tangent function, and then multiplying by two, and this bias term is added to the initial attention score; then, the scores of all client nodes are normalized using an exponential function to calculate the preliminary weight score for each client node under that attention head; four sets of preliminary weight scores are calculated for each of the four attention heads; finally, the preliminary weight scores of each client node under the four attention heads are averaged to obtain the final preliminary weight for that node. Next, the initial weights calculated by the multi-head cross-attention weight generation mechanism are normalized to ensure that the sum of all weights is one. Finally, before weighted aggregation, Mahalanobis distance anomaly detection based on multivariate statistical analysis is performed on all parameter updates. The weights of client nodes corresponding to parameter updates that deviate excessively from the mainstream update direction are temporarily set to zero, and the remaining weights are renormalized to obtain corrected weights for the final weighted aggregation. The specific process of Mahalanobis distance anomaly detection is as follows: First, the robust mean vector of all parameter updates is calculated, using the median as an estimate; then, the minimum covariance determinant estimate of these parameter updates is calculated to obtain a covariance matrix; next, for each client node... The parameter update amount is calculated by subtracting the difference vector from the robust mean vector. This difference vector is then multiplied by the inverse of the covariance matrix and the transpose of the difference vector. The square root of the resulting scalar value is the Mahalanobis distance of that node. A threshold is set as the critical value when the chi-square distribution has the degree of freedom equal to the parameter update amount and the confidence level is 99%. If the Mahalanobis distance of a node exceeds this threshold, its parameter update amount is considered abnormal, and its initial weight is set to zero. Otherwise, its initial weight remains unchanged. Finally, all non-zero weights are renormalized so that their sum is one, resulting in the corrected weights for each node. The specific process of updating the policy network and subsequent synchronization using global parameter update values ​​includes: The central server maintains a small-scale public benchmark validation set containing diverse scenarios; the public benchmark validation set contains intervention effect feedback data from simulated environments and a small number of desensitized real users, covering various scenarios with different ages, different initial vision conditions, and different eye habits; The central server uses the corrected weights to perform a weighted summation of the parameter update amounts of each client node that passed the anomaly detection, and obtains the global parameter update amount. A dynamically adjusted global learning rate is used to update the global policy network parameters by adding the global parameter update amount to the current global policy network parameters. The calculation method for the dynamically adjusted global learning rate is as follows: the initial global learning rate is set to 0.1; in each round of aggregation, the sum of all corrected weights is calculated. If the sum is less than 0.5, the global learning rate is multiplied by 0.9 to decay; if the sum is greater than 2, the global learning rate is multiplied by 1.1 to increase; otherwise, the global learning rate remains unchanged; when updating the policy network parameters, the current global policy network parameters are added to the global learning rate multiplied by the global parameter update amount to obtain the updated parameters. Afterwards, the central server uses the maintained public benchmark validation set to evaluate the performance improvement of the policy network before and after this round of updates. Based on this improvement and the consistency of the direction of the parameter update amount of each client node with the global parameter update amount, the historical contribution reputation of each client node is updated. The process of evaluating the performance improvement is as follows: the policy network parameters before and after the update are evaluated on the public benchmark validation set, the average reward of each on the validation set is calculated, and the average reward before the update is subtracted from the average reward after the update to obtain the performance improvement. The formula for updating historical contribution reputation is: the new reputation equals the old reputation multiplied by 0.9, plus 0.1 multiplied by a reputation increment; The calculation process for the reputation increment is as follows: multiply the corrected weight of the node by the performance improvement, divide by the larger value between the L2 norm and 1 of the node's parameter update, and then add 0.05 multiplied by the cosine similarity between the node's parameter update and the global parameter update. Finally, this reputation increment is cropped to between -1 and 1. The preset threshold for the performance improvement is 0.01. If it is lower than this value, the meta-optimization module may trigger an adjustment. The central server distributes the updated policy network parameters back to all client nodes participating in this round of aggregation. Each client node then replaces its local policy network copy parameters with these updated parameters, achieving synchronous updates.

[0025] In this embodiment, it should be specifically noted that the meta-optimization module includes a trajectory encoder and a long-period value network; the specific process of evaluating the long-term effect of the currently used policy network through the long-period value network is as follows: For each user, new monitoring data and behavioral feedback data generated after they execute a personalized intervention action sequence within a preset long-term time window are collected. Based on the new monitoring data and behavioral feedback data, a long-term trajectory of the user is constructed. This long-term trajectory includes the current decision state of the user at each time point within the long period, which is composed of the strategy generation module; the actions in the final personalized intervention action sequence generated and pushed by the strategy generation module; the instantaneous reward corresponding to the strategy network calculation in the strategy generation module; and the current decision state corresponding to the dynamic risk profile of the next time point generated by the data processing module based on the new monitoring data. The preset long-term time window is twelve months. When constructing the long-term trajectory, if a certain data point is missing, linear interpolation or forward imputation is used to complete it. The constructed long-term trajectory of the user is input into the trajectory encoder. The trajectory encoder, based on an attention mechanism, calculates the importance weights of the joint features of the current decision state and immediate reward at each time step in the long-term trajectory, and then weights and fuses the encodings of all time steps in the long-term trajectory to output a context-aware representation of the trajectory. The trajectory encoder is based on a multi-head attention mechanism. Its specific calculation process is as follows: First, the data at each time step in the long-term trajectory is processed. The input of each time step is formed by concatenating the current decision state vector and the immediate reward scalar, and then mapping it to a unified feature space through a linear transformation layer to obtain the initial encoding for that time step. Next, a learnable global query vector is generated for the entire trajectory. Then, this global query vector is calculated and its relationship to the initial value of each time step in the trajectory is calculated. The dot product between the encodings yields the original attention score for each time step. Then, a learnable bias vector based on the time step position is introduced and added to the original attention score to encode temporal order information. Next, a normalized exponential function is applied to the summed score, transforming it into a probability distribution. The probability value corresponding to each time step in this distribution is the importance weight for that time step. Finally, the initial encodings for all time steps are multiplied by their corresponding importance weights and summed to obtain the context-aware representation of the trajectory. This process is executed in parallel in four independent "attention heads," each using different parameters to generate four different context-aware representations. These four representations are then concatenated and fused through a linear projection layer to output the final context-aware representation of the trajectory. The context-aware representation of the trajectory is input into a long-term value network. The long-term value network outputs a high-dimensional value hidden state vector, which is then mapped to a scalar through a linear projection layer. This scalar serves as the long-term value assessment value for the user under the current policy network. The long-term value network is a fully connected neural network with three hidden layers. The calculation process is as follows: the context-aware representation of the trajectory is input into the first hidden layer, undergoes a linear transformation, and is then passed through a non-linear activation function. The output is input into the second hidden layer, undergoes another linear transformation and activation, and then the output is input into the third hidden layer. A linear transformation is performed, but without passing through an activation function, directly outputting a high-dimensional value hidden state vector with a preset dimension of sixteen. Next, this value hidden state vector is input into a linear projection layer. This layer has no bias term and only calculates the value by the dot product of a weight vector and the input vector, outputting a scalar, which is the long-term value assessment value for the user. Calculate the average of the long-term value assessment values ​​of all active users as the overall long-term performance assessment value of the currently used strategy network; the overall long-term performance assessment value is calculated by adding the long-term value assessment values ​​of all users currently served by the system and whose data has been updated in the past month, and then dividing by the total number of these users. The overall long-term effect assessment value is compared with a preset threshold. If it is lower than the preset threshold, the long-term effect of the currently used strategy network is determined to be lower than expected. The preset threshold is set to 0.7. The meta-optimization module also includes a historical adjustment record database and a dynamic adjustment controller; the specific process of dynamically adjusting at least one of the following: the prediction window of the time-series hybrid model, the reward function weights of the policy network, or the aggregation frequency of the adaptive federated aggregation algorithm is as follows: After each evaluation, the set of mean values ​​of all active users' hidden value vectors, along with the prediction window of the current time-series hybrid model, the reward function weights of the policy network, and the aggregation frequency of the adaptive federated aggregation algorithm, are recorded as the current meta-parameter configuration. The calculation process of the set of mean values ​​of all active users' hidden value vectors is as follows: calculate the average value of the sixteen-dimensional hidden value vectors output by the long-period value network of all active users in each dimension to obtain a sixteen-dimensional mean vector, which is the set of mean values. Once a new overall long-term effect assessment value is calculated, the change between this overall long-term effect assessment value and the previously recorded overall long-term effect assessment value is associated with the corresponding meta-parameter configuration record and stored in the historical adjustment record database. The change is the new overall long-term effect assessment value minus the previously recorded overall long-term effect assessment value. The historical adjustment record database is stored using a time-series database, and each record contains a timestamp, a vector of pre-adjustment mean sets, meta-parameter configuration, and the change in the overall long-term effect assessment value. When it is determined that the long-term effect of the currently used strategy network is lower than expected, the dynamic adjustment controller initiates the adjustment process; The dynamic adjustment controller first reads all historical records from the historical adjustment record database; Using all historical records, a Gaussian process regression model is trained. This model takes the set of means of the unadjusted user value latent state vectors stored in the historical records and candidate meta-parameter configurations as inputs, and predicts the expected and variance of the overall long-term effect evaluation value after adjustment. The training process of the Gaussian process regression model is as follows: First, a kernel function is defined. For the value latent state mean set, a radial basis function kernel is used. For the meta-parameter configuration, a Marton kernel is used for the prediction window and aggregation frequency, and a linear kernel is used for the reward function weight vector. Then, the outputs of these two kernel functions are multiplied to obtain the final composite kernel function. Then, based on this composite kernel function and all input-output pairs in the historical records, the hyperparameters of the kernel function are optimized by maximizing the marginal likelihood function to complete the model training. After training, for any new input combination, the model can predict the expected and variance of the corresponding change in the overall long-term effect evaluation value. Next, within the search space satisfying the preset medical feasibility and meta-parameter configuration, the optimal candidate new configuration of meta-parameters is searched by maximizing an acquisition function. The acquisition function is the expected value predicted by the Gaussian process regression model minus the current overall long-term effect assessment value, divided by the predicted variance, plus an exploration coefficient multiplied by the predicted variance. The search process adopts a Bayesian optimization framework. The specific steps are as follows: First, define the search space; the prediction window ranges from three months to twelve months as an integer; the reward function weights are two non-negative real numbers between zero and one that sum to one; the aggregation frequency ranges from... The time interval is an integer between twelve and seventy-two hours. Then, one hundred candidate meta-parameter configurations are randomly sampled within the search space. For each candidate configuration, it is combined with the current set of mean values ​​of the hidden values ​​and input into a trained Gaussian process regression model to obtain the expected value and variance of the prediction. Next, the acquisition function value of each candidate configuration is calculated by first calculating the difference between the expected value of the prediction and the current overall long-term effect evaluation value, then dividing by the prediction variance, and then adding 0.1 (exploration coefficient) multiplied by the prediction variance. Finally, the candidate configuration with the largest acquisition function value is selected as the optimal new candidate meta-parameter configuration. The dynamic adjustment controller conducts A / B testing by progressively deploying the optimal candidate meta-parameter configuration obtained from the search in a traffic-sharing manner. During the test, a causal forest model is used to analyze the difference in long-term value assessment values ​​between the experimental group and the control group, and to evaluate the heterogeneity of this difference under different user characteristics. The specific process of the A / B test is as follows: 10% of users are randomly selected as the experimental group, and their service strategy is switched to use the optimal candidate meta-parameter configuration; the remaining 90% of users are the control group, and their original meta-parameter configuration is maintained; the test lasts for two weeks. The construction process of the causal forest model is as follows: using whether the user is in the experimental group as the treatment variable, the user's long-term value assessment value at the end of the test period as the outcome variable, and the user's age, baseline vision, historical eye habits, and other characteristics as covariates, two hundred unpruned decision trees are constructed; each tree is trained on random subsamples and random feature subsets, and the individual treatment effect of each user is estimated by calculating the difference in the mean of the results in the leaf nodes between the treatment group and the control group; by analyzing the distribution of individual treatment effects of all users, the overall effect and heterogeneity of the new configuration are evaluated. If the A / B test results indicate that the optimal candidate meta-parameter configuration produced a statistically significant positive overall effect in the experimental group, and no widespread and severe negative effects were found on the specified user group, then the dynamic adjustment controller will gradually expand the deployment of the optimal candidate meta-parameter configuration to all users, completing the update of the meta-parameter configuration. The criteria for "statistically significant positive overall effect" are: the p-value of the two-tailed T-test of the difference in long-term value assessment values ​​between the experimental group and the control group is less than 0.05, and the mean of the difference is greater than 0.01. The criteria for "widespread and severe negative effect" are: there exists a subgroup defined by a certain user characteristic, whose mean individual treatment effect is less than -0.05, and the proportion of the number of people in this subgroup to the total number of people in the experimental group exceeds 5%. If the positive effect criteria are met but the negative effect criteria are not met, then the gradual deployment will be initiated, expanding the deployment scope by 20% each week until all users are covered. Meanwhile, the dynamic adjustment controller continuously monitors three metrics: the rate of decline of the overall long-term performance evaluation value within the sliding window, the statistical distance between the mean set of the current user value hidden state vector and the historical normal set, and the amount of communication resources consumed per unit performance improvement during federated learning. When the value of any monitoring metric exceeds its corresponding adaptive calculation threshold, the dynamic adjustment controller also initiates the adjustment process. The calculation and threshold settings for monitoring indicators are as follows: Rate of decline: Calculate the slope of the linear regression of the overall long-term effect assessment value over the past four weeks. If the slope is less than -0.01, it is triggered. Statistical Distance: Calculate the Mahalanobis distance between the current sixteen-dimensional value hidden state mean vector and the mean vector of the set of mean vectors over the past thirty days. The Mahalanobis distance is calculated as follows: First, calculate the covariance matrix of all mean vectors over the past thirty days. Then, calculate the difference between the current vector and the historical mean vectors. Multiply the difference vector by the inverse of the covariance matrix, and then multiply by the transpose of the difference vector. The square root of the result is the Mahalanobis distance. If the distance is greater than 3.0, it is triggered. Unit communication resource consumption: Calculate the total number of bytes transmitted in the most recent federated aggregation cycle and divide it by the improvement in the overall long-term effect assessment value (set to a very small number if there is no improvement); if the calculation result is greater than 0.01 effect value per 100 megabytes, then trigger; The preset threshold for the overall long-term effect evaluation value is 0.7. This embodiment provides a specific method for using a vision health data monitoring and intelligent analysis management system. The specific method is as follows: First, the data processing module is responsible for cutting-edge risk prediction. The system receives and standardizes the user's eye health indicators (such as spherical power and axial length) and behavioral data (such as screen time) from previous periods. These data are input into an advanced time-series hybrid model. This model extracts time-series features from different dimensions through the parallel work of local trend capture branches, periodic pattern discovery branches, and global dependency modeling branches. The fused features capture long-term dependencies through a state-space model layer. Finally, it not only outputs the predicted values ​​of core indicators for multiple future time points (such as the next 12 months), but also generates multiple sets of predictions through random discarding technology to quantify uncertainty. This generates a comprehensive dynamic risk profile that includes predicted values, dynamic risk probability distribution, and prediction interval width, providing a forward-looking and quantitative basis for intervention decisions. Subsequently, the strategy generation module makes real-time personalized decisions based on this profile. The module integrates the dynamic risk profile with the user terminal's real-time behavior and environmental context. Through the collaborative processing of the profile encoder, context encoder, and personalized state memory unit, it constructs a current decision state that includes historical memory and real-time state. The strategy network takes this state as input, and its internal shared feature layer and multiple action generation sub-networks jointly generate a preliminary sequence of future short-term intervention actions. To ensure absolute safety, the real-time constraint verification and correction layer quickly optimizes this sequence to ensure that it fully meets preset medical constraints such as "the total daily intervention time does not exceed thirty minutes," forming the final personalized intervention action sequence. Finally, the template engine decodes it into easy-to-understand natural language instructions (such as "It is recommended to conduct 20 minutes of distant viewing training tomorrow afternoon") and pushes it to the user terminal for execution. Subsequently, the model evolution module drives the continuous optimization of the policy network while protecting privacy. At preset intervals, the central server distributes the latest policy network parameters to client nodes such as medical institutions. Each node uses the feedback data on the intervention effect after local users perform interventions to train its local policy network copy. The innovative contrastive regularization term in training encourages local innovation while maintaining consistency with global knowledge. After training, each node uploads the parameter update and local training metadata. The server adopts an adaptive federated aggregation algorithm based on contribution evaluation, uses a multi-head cross-attention weight generation mechanism to dynamically evaluate the contribution of each node, and combines Mahalanobis distance anomaly detection to filter abnormal updates. The weighted aggregation generates global parameter update to update the global policy network, and synchronizes the new parameters back to all nodes, thereby achieving secure knowledge sharing and collaborative evolution across institutions. Finally, the meta-optimization module acts as the system's "super brain" for macro-level monitoring and tuning. This module continuously collects long-term data, constructs long-term trajectories for each user, and evaluates the long-term effectiveness of the current policy network through a trajectory encoder and a long-term value network. When the overall long-term performance evaluation value is lower than a preset threshold, or when abnormal indicators such as performance degradation or data distribution drift are detected, the module initiates an adjustment process. Its dynamic adjustment controller, based on a historical adjustment record database, uses a Gaussian process regression model and Bayesian optimization to intelligently search for new configurations of meta-parameters such as the prediction window of the time-series hybrid model, the reward function weights of the policy network, or the aggregation frequency of the adaptive federated aggregation algorithm, within the scope of medical feasibility. After the new configuration is confirmed to be effective through A / B testing and causal forest model evaluation, it is gradually deployed to the entire system.

[0026] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0032] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A vision health data monitoring and intelligent analysis management system, characterized in that, Specifically, it includes: The data processing module, strategy generation module, model evolution module, and meta-optimization module are sequentially connected via communication. Data processing module: Upon receiving historical time-series monitoring data from a user, the module standardizes the data and inputs the processed data into a pre-trained time-series mixture model to output predicted values ​​of eye health indicators and dynamic risk probability distributions for multiple future time points, thereby generating a dynamic risk profile for the user. The strategy generation module: Based on the dynamic risk profile, real-time user behavior data obtained from the user terminal, and environmental context, it constructs the current decision state, calls the strategy network to output a short-term personalized intervention action sequence based on the current decision state and under the condition of meeting preset medical constraints, and pushes the personalized intervention action sequence to the user terminal. Model Evolution Module: When the preset federated aggregation cycle is triggered, it coordinates multiple client nodes to iteratively update the local policy network copy based on the feedback data of the intervention effect after the local user executes the personalized intervention action sequence. The local policy network copy is a local copy of the policy network on each client node. It adopts an adaptive federated aggregation algorithm based on contribution evaluation to perform weighted aggregation of the parameter update amount of the local policy network copy of all client nodes to generate a global parameter update amount. The global parameter update amount is then used to update the policy network. The updated policy network parameters are then distributed back to each client node to synchronously update its local policy network copy. Meta-optimization module: continuously collects new monitoring data and behavioral feedback data after users perform personalized intervention action sequences, evaluates the long-term effect of the currently used strategy network through a long-term value network, and dynamically adjusts at least one of the following when the evaluation result is lower than a preset threshold: the prediction window of the time series hybrid model, the weight of the reward function of the strategy network, or the aggregation frequency of the adaptive federated aggregation algorithm.

2. The vision health data monitoring and intelligent analysis management system according to claim 1, characterized in that: In the data processing module, the specific process of the time-series hybrid model processing the standardized historical time-series monitoring data is as follows: After standardizing the historical time-series monitoring data, the standardized historical time-series monitoring data is simultaneously input into the local trend capture branch, periodic pattern discovery branch, and global dependency modeling branch of the time-series hybrid model for parallel feature extraction. Among them, the local trend capture branch extracts local change pattern features at different time scales from the standardized historical time series monitoring data through causal dilated convolutional layers with different dilation coefficients. The periodic pattern discovery branch calculates the response intensity of the historical time-series monitoring data in different potential periodic dimensions by cross-correlation calculation with a learnable periodic basis function library after standardization of the historical time-series monitoring data, and generates periodic pattern features by weighting them; the global dependency modeling branch calculates the global dependency features between features of all time points in the standardized historical time-series monitoring data by introducing a multi-head self-attention mechanism based on relative position bias with time distance decay. Subsequently, the temporal hybrid model uses an adaptive weight fusion layer to dynamically calculate and assign fusion weights to the local change pattern features, periodic pattern features, and global dependency features at each time step, and outputs the weighted sum as a fused multi-scale feature representation sequence.

3. The vision health data monitoring and intelligent analysis management system according to claim 2, characterized in that: The specific process for generating the dynamic risk profile is as follows: The fused multi-scale feature representation sequence is input into a state space model layer contained in the temporal hybrid model. The state space model layer models the long-term dynamic dependence of the sequence through its internal structured state equations and outputs the final hidden state of the sequence. Based on the final sequence hidden state, the predicted values ​​of eye health indicators at multiple future time points are calculated through the prediction layer included in the temporal mixture model. Meanwhile, during the inference phase, random discarding is applied multiple times to the local trend capture branch, the periodic pattern discovery branch, the global dependency modeling branch, and the state space model layer to generate multiple sets of eye health indicator prediction values ​​for future time points. Based on the multiple sets of eye health indicator prediction values ​​for future time points, the empirical distribution of the prediction value for each future time point is calculated, and the probability distribution of the eye health risk level for each future time point is output through the risk quantification layer contained in the time series hybrid model. The dynamic risk probability distribution is composed of the probability distribution of eye health risk levels at all future time points; The dynamic risk profile consists of predicted values ​​of eye health indicators at multiple future time points, a dynamic risk probability distribution, and a prediction interval width calculated based on multiple sets of predicted values ​​of eye health indicators at future time points to characterize the uncertainty of the prediction.

4. The vision health data monitoring and intelligent analysis management system according to claim 3, characterized in that: In the policy generation module, the specific process of constructing the current decision state and invoking the policy network is as follows: First, the dynamic risk profile is input into the profile encoder. The profile encoder extracts feature vectors representing future risk trends from the predicted values ​​of eye health indicators at multiple future time points, the dynamic risk probability distribution, and the prediction interval width contained in the dynamic risk profile through an attention mechanism. Simultaneously, the real-time behavior data of the user obtained from the user terminal and the environmental context are input into the context encoder to obtain the real-time context feature vector; Next, the feature vector representing future risk trends extracted by the portrait encoder is concatenated with the real-time context feature vector obtained by the context encoder to form a fused state vector, and the fused state vector is input into a preset personalized state memory unit. The personalized state memory unit updates itself based on the current fused state vector and the user's historical state memory stored within it, and outputs the state memory vector at the current moment. Finally, the fused state vector is concatenated with the current state memory vector to form the current decision state, and this current decision state is input into the policy network.

5. The vision health data monitoring and intelligent analysis management system according to claim 4, characterized in that: The specific process by which the strategy network outputs and pushes personalized intervention action sequences is as follows: The policy network consists of a shared feature layer, multiple parallel action generation subnetworks, a real-time constraint verification and correction layer, and a template engine; The shared feature layer processes the current state of the input decision and extracts high-level features; Each action generation subnetwork corresponds to a future time step. Based on the high-level features extracted from the shared feature layer, it generates intervention action parameters for that time step. The intervention action parameters include a mean vector that defines the center value of the action and a variance vector that defines the range of action exploration. When generating personalized intervention action sequences, the mean vector of each action generation subnetwork is taken and arranged in the order of the corresponding time steps to form the initial action sequence generated by the policy network. The real-time constraint verification and correction layer receives the initial action sequence. Under the premise of satisfying all preset medical inequality constraints and equality constraints, it calculates the final personalized intervention action sequence by solving an optimization problem with the objective of minimizing the weighted Euclidean distance between the initial action sequence and the corrected action sequence. The template engine decodes the final personalized intervention action sequence into natural language instructions, generates the final push message, and sends it to the user's terminal.

6. The vision health data monitoring and intelligent analysis management system according to claim 5, characterized in that: The model evolution module includes a central server and multiple client nodes; the specific process of coordinating multiple client nodes to iteratively update the local policy network copy based on the intervention effect feedback data after local users execute personalized intervention action sequences is as follows: At the start of each preset federated aggregation cycle, the central server distributes the current policy network parameters as global policy network parameters to each client node. Each client node initializes its local policy network copy with the received global policy network parameters and trains the local policy network copy with the intervention effect feedback dataset stored locally. During training, a loss function that combines the policy optimization objective and the contrastive regularization term is used. The policy optimization objective is to enable the intervention action sequences generated by the local policy network replicas to obtain higher cumulative rewards on the intervention effect feedback dataset; the contrast regularization term is used to encourage the local policy network replicas to learn the overall consistency with the output actions of the current global policy network, while strengthening their ability to distinguish and retain those distinctive actions in the intervention effect feedback dataset that perform well but are different from the average output of the global policy. After completing a preset number of local training cycles, each client node calculates the difference between the parameters of its local policy network replica and the initially received global policy network parameters. This difference is used as the parameter update amount for the client node's local policy network replica. This parameter update amount, along with the performance improvement metrics, dataset size and distribution characteristics, and training stability metrics recorded during local training, are sent to the central server as local training metadata.

7. The vision health data monitoring and intelligent analysis management system according to claim 6, characterized in that: The specific process of using the adaptive federated aggregation algorithm based on contribution evaluation to weighted aggregate the parameter update amounts of the local policy network replicas of all client nodes is as follows: The central server maintains a historical contribution reputation score for each client node, which is a dynamically updated score based on the quality and consistency of the client node's past updates. After receiving the parameter updates and local training metadata from all client nodes, the central server first constructs a contribution feature vector for each client node that comprehensively reflects the performance improvement indicators, data value, and update quality in its local training metadata. Using a multi-head cross-attention weight generation mechanism, based on the contribution feature vectors of all client nodes and the historical contribution reputation scores of each client node, the central server dynamically calculates the initial weight of each client node in this round of aggregation. The initial weights calculated by the multi-head cross-attention weight generation mechanism are then normalized to ensure that the sum of all weights is one. Finally, before weighted aggregation, Mahalanobis distance anomaly detection is performed on all parameter update amounts based on multivariate statistical analysis. The weights of client nodes corresponding to parameter update amounts that deviate too much from the mainstream update direction are temporarily set to zero, and the remaining weights are renormalized to obtain the corrected weights used for the final weighted aggregation.

8. The vision health data monitoring and intelligent analysis management system according to claim 7, characterized in that: The specific process of updating the policy network and subsequent synchronization using global parameter update values ​​includes: The central server maintains a small-scale public benchmark verification set containing diverse scenarios; the central server uses corrected weights to perform a weighted summation of the parameter update amounts of each client node that passes anomaly detection to obtain the global parameter update amount; Using a dynamically adjusted global learning rate, the global parameter update amount is added to the current global policy network parameters to complete the policy network update; Afterwards, the central server uses the maintained public benchmark verification set to evaluate the improvement in policy network performance before and after this round of updates, and updates the historical contribution reputation of each client node based on this improvement and the directional consistency between the parameter update amount of each client node and the global parameter update amount. The central server distributes the updated policy network parameters back to all client nodes participating in this round of aggregation. Each client node then replaces its local policy network copy parameters with these updated parameters, achieving synchronous updates.

9. The vision health data monitoring and intelligent analysis management system according to claim 8, characterized in that: The meta-optimization module includes a trajectory encoder and a long-period value network; the specific process of evaluating the long-term effect of the currently used policy network through the long-period value network is as follows: For each user, new monitoring data and behavioral feedback data are collected after they execute a personalized intervention action sequence within a preset long-term time window. Based on the new monitoring data and behavioral feedback data, a long-term trajectory of the user is constructed. This long-term trajectory includes the current decision state of the user at each time point within the long period, which is composed of the strategy generation module; the actions in the final personalized intervention action sequence generated and pushed by the strategy generation module; the instant reward corresponding to the strategy network calculation in the strategy generation module; and the current decision state corresponding to the dynamic risk profile of the next time point generated by the data processing module based on the new monitoring data. The constructed long-term trajectory of the user is input into the trajectory encoder. The trajectory encoder is based on the attention mechanism. It calculates the importance weight of the joint feature of the current decision state and the immediate reward at each time step in the long-term trajectory, and performs weighted fusion of the encoding of all time steps in the long-term trajectory to output the context-aware representation of the trajectory. The context-aware representation of the trajectory is input into the long-term value network. The long-term value network outputs a high-dimensional value hidden state vector, and maps this value hidden state vector to a scalar through a linear projection layer as the long-term value evaluation value of the user under the current policy network. Calculate the average of the long-term value assessments of all active users as the overall long-term performance assessment of the currently used strategy network. The overall long-term effect assessment value is compared with a preset threshold. If it is lower than the preset threshold, the long-term effect of the currently used strategy network is determined to be lower than expected.

10. The vision health data monitoring and intelligent analysis management system according to claim 9, characterized in that: The meta-optimization module further includes a historical adjustment record database and a dynamic adjustment controller; the specific process of dynamically adjusting at least one of the following: the prediction window of the time-series hybrid model, the reward function weights of the policy network, or the aggregation frequency of the adaptive federated aggregation algorithm is as follows: After each evaluation, the set of the mean values ​​of the hidden state vectors of all active users, the prediction window of the current time series hybrid model, the reward function weights of the policy network, and the aggregation frequency of the adaptive federated aggregation algorithm are recorded as the current meta-parameter configuration. When a new overall long-term effect evaluation value is subsequently calculated, the change between the overall long-term effect evaluation value and the previously recorded overall long-term effect evaluation value is associated with the corresponding meta-parameter configuration record and stored in the historical adjustment record database. When it is determined that the long-term effect of the currently used strategy network is lower than expected, the dynamic adjustment controller initiates the adjustment process; the dynamic adjustment controller first reads all historical records from the historical adjustment record database; Using all historical records, a Gaussian process regression model is trained. This model takes as input the set of mean values ​​of the unadjusted user value latent state vectors stored in the historical records and the candidate meta-parameter configurations, and predicts the expected and variance of the overall long-term effect evaluation value that may be obtained after adjustment. Next, within the search space that satisfies the preset medical feasibility and meta-parameter configuration, the optimal candidate meta-parameter configuration is searched by maximizing an acquisition function; the acquisition function is the expected value predicted by the Gaussian process regression model minus the current overall long-term effect evaluation value divided by the predicted variance, plus an exploration coefficient multiplied by the predicted variance. The dynamic adjustment controller conducts A / B testing on the optimal candidate meta-parameter configuration obtained from the search in a phased deployment manner. During the test, the causal forest model is used to analyze the difference in long-term value assessment values ​​between the experimental group and the control group, and to assess the heterogeneity of this difference under different user characteristics. If the A / B test results show that the optimal candidate meta-parameter configuration has a statistically significant positive overall effect in the experimental group and no widespread and serious negative effect is found on the specified user group, the dynamic adjustment controller will gradually expand the deployment of the optimal candidate meta-parameter configuration to all users, thus completing the update of the meta-parameter configuration. Meanwhile, the dynamic adjustment controller continuously monitors three metrics: the rate of decline of the overall long-term performance evaluation value within the sliding window, the statistical distance between the mean set of the current user value hidden state vector and the historical normal set, and the amount of communication resources consumed per unit performance improvement during federated learning. When the value of any monitoring metric exceeds its corresponding adaptive calculation threshold, the dynamic adjustment controller also initiates the adjustment process.

Images