Dynamic assessment and adaptive intervention method for digital health literacy of the elderly based on recursive bayesian filtering

Abstract

The application relates to a kind of old people digital health literacy dynamic evaluation and adaptive intervention method based on recursive bayesian filtering, which comprises the following steps: S1, constructing the observation data sequence containing missing mode;S2, using the dynamic state estimation method based on recursive bayesian filtering to dynamically estimate the digital health literacy of the old people, to obtain the optimal state estimation value and estimation covariance matrix at the current time;S3, calculate the uncertainty index at the current time;S4, judge whether it exceeds threshold, if not, then execute step S6, if yes, then trigger the active inquiry mechanism, carry out filtering update;S5, recalculate the uncertainty index, return to step S4 until it does not exceed;S6, predict the risk prediction result at the current time, and combine the uncertainty index, using the context multi-arm tiger machine model to select personalized intervention strategy.Compared with the prior art, the application has the advantages of overcoming the time dispersion defect of static questionnaire method and the like.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and artificial intelligence, and in particular to a method for dynamic assessment and adaptive intervention of digital health literacy of the elderly based on recursive Bayesian filtering. Background Technology

[0002] As the population ages rapidly, the proportion of older adults accessing health information and using digital healthcare services through the mobile internet is constantly increasing. However, due to differences in digital skills, fluctuations in cognitive levels, and irregular behavioral habits, there are significant imbalances and dynamic changes in digital health literacy among older adults.

[0003] Existing methods for assessing and intervening in digital health literacy among older adults have the following main shortcomings: (1) The evaluation method is static and lacks the ability to model state evolution. Existing methods are mostly based on one-off or phased questionnaire surveys (e.g., patent application CN121687464A), using the KAP (Knowledge-Attitude-Behavior) model for static scoring. These methods cannot describe the dynamic process of health literacy changing continuously over time, nor can they capture cognitive and behavioral mutations caused by illness, emotions, etc.

[0004] (2) Inability to effectively handle the problem of sparse and irregular data of the elderly. When older adults use digital health systems, they often fail to log in for extended periods, miss questionnaires, or forget to wear wearable devices, resulting in significant gaps in observational data. Existing methods often employ mean imputation or direct data removal, which disrupts the temporal correlation of the data and leads to substantial biases in the assessment results.

[0005] (3) Lack of quantification and proactive management of the uncertainty of assessment results Existing models only output a definitive score or risk level, failing to inform users or doctors how reliable the outcome is. Blindly trusting assessment results when data is insufficient or of poor quality can lead to misjudgment. Furthermore, the system lacks the ability to proactively consult with users to reduce uncertainty.

[0006] (4) The intervention strategies are fixed and extensive, lacking adaptive learning ability. Existing intervention programs employ fixed intervention protocols (such as patent CN121839150B) or often use hard-coded rules of "high risk - high frequency push, low risk - maintenance". Different elderly people respond very differently to the same intervention, and fixed rules cannot achieve personalized adaptation that makes them "smarter with use".

[0007] (5) Lack of closed-loop feedback and incremental model update mechanism Most systems implement the assessment, prediction, and intervention stages in isolation, without transmitting the intervention results back or updating the model, making it impossible to continuously optimize as data accumulates.

[0008] Therefore, a new approach is urgently needed to address the unique technical challenges in the field of digital health for the elderly, such as data sparsity, sudden changes in state, and large individual differences. Summary of the Invention

[0009] The purpose of this invention is to provide a dynamic assessment and adaptive intervention method for digital health literacy of the elderly based on recursive Bayesian filtering, which solves the problem of time discreteness in static assessment.

[0010] The objective of this invention can be achieved through the following technical solutions: A dynamic assessment and adaptive intervention method for digital health literacy among the elderly based on recursive Bayesian filtering includes the following steps: S1. Obtain multi-source heterogeneous observation data of elderly users and construct an observation data sequence containing missing patterns; S2. Based on the observation data sequence containing the missing pattern, the digital health literacy of the elderly is dynamically estimated using a dynamic state estimation method based on recursive Bayesian filtering to obtain the optimal state estimate and the estimated covariance matrix at the current moment. S3. Based on the estimated covariance matrix at the current time. Calculate the uncertainty index at the current moment; S4. Determine whether the uncertainty index exceeds a preset threshold. If not, proceed to step S6; if yes, trigger the active query mechanism at the current moment and perform a filter update. S5. Recalculate the uncertainty index based on the filter update result, return to step S4, until it does not exceed [the specified value]. ; S6. Based on the optimal state estimate and estimated covariance matrix at the current moment, predict the risk prediction result at the current moment, and combine the uncertainty index to adaptively select personalized intervention strategies using a contextual multi-armed slot machine model.

[0011] Furthermore, the multi-source heterogeneous observational data includes basic demographic data, multi-dimensional scoring data based on the KAP theoretical framework, digital health behavior logs, and health indicator data. The basic demographic data includes age, sex, education level, and history of chronic diseases; The multi-dimensional scoring data includes scoring data for the knowledge dimension, scoring data for the attitude dimension, and scoring data for the behavior dimension, which are represented as follows: , , , In the formula, For the knowledge dimension of the scoring data, For the attitude dimension of the scoring data, For behavioral dimension scoring data, For knowledge dimension The number of questionnaire questions 1 indicates a correct answer, and 0 indicates an incorrect answer. Attitude dimension The number of questionnaire questions For the attitude dimension The original score of the question. For behavioral dimensions The number of questionnaire questions For the behavioral dimension The original score of the question; The digital health behavior log includes APP usage frequency, course learning duration, number of online consultations, and function click paths. The digital health behavior log is represented as follows: , , , In the formula, For digital health behavior logs, For the first The weight of each behavior log item For the first The normalized value of the behavior log. , For the first Minimum and maximum values ​​in the item behavior log, For the first The value of the item behavior log; The health indicators include blood pressure, blood sugar, heart rate, and BMI.

[0012] Furthermore, the steps for constructing the observation data sequence containing the missing pattern include: The multi-source heterogeneous data is sequentially cleaned, normalized, and feature-aligned to obtain a preprocessed observation data sequence. For the preprocessed observation data sequence, a missing selection matrix is ​​introduced to characterize whether there are missing data channels. Represented as: , In the formula, For diagonal matrix functions, For the first Missing data channel indication; The preprocessed observation data sequence and the missing selection matrix are then compared. The data is then spliced ​​together to obtain an observation data sequence that includes the missing patterns.

[0013] Furthermore, the execution steps of the dynamic state estimation method based on recursive Bayesian filtering include: definition: Digital health literacy among older adults is set as an unobservable latent state variable. And construct a state-space model, represented as: , in: , , , In the formula, , , These are the actual potential health literacy status variables in the dimensions of knowledge, attitude, and behavior. For the current moment The state transition matrix, For process noise, For process noise covariance, The mean is 0 and the process noise covariance is Gaussian distribution, With a mean of 0 and an observation noise covariance of , Gaussian distribution, For a moment The observation vector, For the observation matrix, To observe the noise, To observe the noise covariance, For the knowledge dimension of the scoring data, For the attitude dimension of the scoring data, For behavioral dimension scoring data, For digital health behavior logs, for Weights related to the knowledge dimension in the middle. for The weights related to the attitude dimension, for Weights related to behavioral dimensions For diagonal matrix functions, These represent the standard deviations of process noise during the evolution of potential states of knowledge, attitudes, and behaviors. These are the standard deviations of observation noise during the observation of knowledge, attitude, and behavior scores, as well as digital health behavior logs. predict: For the current moment According to the previous moment posterior probability distribution of the output , Perform state prediction and error covariance prediction, and output the time. Prior probability distribution , The expression for the state prediction is: , The expression for the error covariance prediction is: , In the formula, For the previous moment The observation vector of the state The optimal estimate, i.e., the estimate from the previous time step. The optimal state estimate, For the previous moment The estimated covariance matrix, For a moment The prior state estimate, For a moment The prior estimate of the covariance matrix; Filter update: According to the time The prior probability distribution is used to calculate the prediction residual. With residual covariance , respectively represented as: , , If time observation vector When some observation channels are missing, a missing selection matrix is ​​introduced. Missing channels are masked, and based on the residual covariance... and prior estimation of covariance matrix Calculate the Kalman gain after processing for missing observations. , is represented as: , In the formula, For the set constant, for 3D identity matrix; Based on the Kalman gain and predicted residuals Update the state to obtain the current time. optimal state estimate , is represented as: , , Based on the Kalman gain Perform covariance update to obtain the current time step. Estimated covariance matrix , is represented as: , , In the formula, It is the identity matrix. Estimating the variance of the knowledge dimension state error. The variance of the state estimation error for the attitude dimension. The variance of the state estimation error for the behavioral dimension. The off-diagonal elements represent the covariance between estimation errors in different dimensions; If all observation channels are valid The Kalman gain calculation formula after missing observation processing degenerates into the standard Kalman gain calculation formula; If the current time There is no valid observation data available for filter updates, i.e., the observation vector. All observation channels are missing, or the selection matrix is ​​missing. If all diagonal elements are 0, then skip the above filtering update operation and directly set the time... Prior state estimate and time Prior estimate of covariance matrix Each as the current moment optimal state estimate and the current moment Estimated covariance matrix ,Right now: , ; Return to the above prediction steps to obtain the optimal state estimate and estimated covariance matrix for the next time step, until the end, completing the dynamic state estimation process for each time step.

[0014] Furthermore, it also includes state mutation detection and adaptive reset steps, specifically including the following: During the dynamic estimation process, based on the predicted residual With residual covariance The normalized sum of squares of the residuals is calculated and expressed as: , In the formula, This is the normalized sum of squares of the residuals; When there are two or more consecutive moments... Exceeding the detection threshold If a state change occurs, the process noise covariance is increased. ,Right now , For adjustment coefficients, Or partially reset the aforementioned time. Estimated covariance matrix ,Right now , The pre-defined covariance reset matrix, This is a preset maximum coefficient. for 3D identity matrix.

[0015] Furthermore, the uncertainty index is expressed as: , In the formula, As an indicator of uncertainty, For a moment The estimated covariance matrix, is the dimension of the state vector.

[0016] Furthermore, the step of performing a filter update includes: Once the proactive inquiry mechanism is triggered, the elderly user's response is mapped to a pseudo-observation vector, represented as: , In the formula, This is a pseudo-observation vector, whose dimension is the same as the observation vector. Consistent For mapping functions, In response; For the pseudo-observation vector Introducing the missing selection matrix The missing selection matrix is ​​used to characterize whether there are missing responses from elderly users. Represented as: , Based on the pseudo-observation vector Calculate the pseudo-observation prediction residuals and pseudo-observation residual covariance , respectively represented as: , , In the formula, For the observation matrix, For a moment The prior state estimate, For a moment The prior estimate of the covariance matrix, The pseudo-observation noise covariance matrix, For the first Noise variance of each pseudo-observation dimension; Based on the pseudo-observation residual covariance and missing selection matrix Calculate the Kalman gain in the pseudo-observation update. , is represented as: , In the formula, Here is the regularization constant. for 3D identity matrix; Kalman gain based on the pseudo-observation update and pseudo-observation prediction residuals Update the state and obtain the time. optimal state estimate , is represented as: , Kalman gain based on the pseudo-observation update Perform covariance update to obtain time step. Estimated covariance matrix This completes one filter update process, where time... Estimated covariance matrix Represented as: .

[0017] Furthermore, the prediction steps for the risk prediction result at the current moment include: Based on the current time The optimal state estimate and the estimated covariance matrix, combined with the time step Observation vectors, construct feature vectors ,in, For a moment The optimal state estimate, For a moment The estimated covariance matrix, For a moment The observation vector; Let the random forest classifier include A decision tree, a set of health management risk level categories for: , For any risk level category ,make Indicates time Predicting as a category in a random forest The number of decision trees, then the categories The voting percentage is expressed as follows: , In the formula, For prediction as a category The proportion of decision trees to all decision trees; The feature vector Input the random forest classifier and output the risk prediction result at the current time, including the health management risk level. and the corresponding confidence level The health management risk level mentioned therein Represented as: , In the formula, The total number of trees in the random forest; The confidence level Represented as: , The prediction step also includes: Determine the confidence level Is it less than the confidence threshold? If yes, then the active query mechanism in step S4 is triggered to re-predict; otherwise, the subsequent steps continue.

[0018] Furthermore, the specific execution steps of the contextual multi-armed slot machine model include: definition: Based on the optimal state estimate at the current moment Uncertainty indicators Health management risk level And history most recently The reward vector for the next action is used to construct the current time step. context vector ,in, Health management risk level one-hot encoding, For the most recent history The reward vector for this action; Set action set Each of the actions A corresponding type of intervention strategy; Set Action After execution, the reward function, constructed by combining behavioral changes, KAP score changes, health indicator improvements, and user feedback within a preset short time window, is expressed as: , in: , , In the formula, For action The corresponding rewards This represents the change in behavioral activity. To increase KAP ratings For the improvement of health indicators, In response to positive user feedback, As negative feedback, For the weights of the reward function, For the first The normalized improvement of a health indicator within a preset short time window The weights for the corresponding health indicators; initialization: For each action Initialize the feature matrix and expected cumulative reward vector ,in For dimension The identity matrix, context vector The dimension; UCB value calculation and action selection: For each action According to the feature matrix and cumulative reward feature vector The estimated values ​​of the LinUCB parameters were obtained. , is represented as: , Based on the context vector and LinUCB parameter estimates Calculate the action UCB value And select the current time. As the optimal intervention action The UCB value and optimal intervention actions They are represented as follows: , , In the formula, To explore the adjustment coefficient; Action execution and observation: Perform the optimal intervention action And observe rewards within a preset short time window. ; Online updates: Based on the aforementioned rewards Online updates are used to determine the optimal intervention action. The corresponding feature matrix and the expected cumulative reward vector are used to obtain the updated feature matrix. and expected cumulative reward vector , respectively represented as: , , In the formula, For assignment / update operations; Time progression: make Return to the UCB value calculation and action selection steps until the end.

[0019] Furthermore, it also includes incremental learning and closed-loop optimization steps, specifically including the following: Acquire new batches of data within a set interval, including newly added observation vectors in the time series. Missing selection matrix Optimal state estimate Estimating the covariance matrix Uncertainty indicators eigenvectors Health management risk level Confidence level Context vector Optimal intervention action and rewards ; Using the feature vectors in the new data batch The risk labels are used to construct new decision trees in the random forest to replace the old decision trees that take the longest to build or have the lowest performance, while keeping the number of decision trees in the random forest constant. The risk labels include health management risk levels confirmed manually, health management risk levels generated from follow-up results, or health management risk levels confirmed based on preset rules. Using the observation vectors in the new data batch Missing selection matrix Optimal state estimate Estimating the covariance matrix The state-space model parameters in the dynamic state estimation method based on recursive Bayesian filtering are... , , Update; among them, Here is the state transition matrix. For process noise covariance, To observe the noise covariance, update methods include re-estimating it offline using the expectation-maximization algorithm, or updating it online using a sliding window approach. Utilizing the context vector in the new data batch Optimal intervention action and rewards Update the context of each action in the multi-armed slot machine model. Corresponding action expected reward model parameters ,in: , In the formula, For expectations; When using the LinUCB algorithm, the action expected reward model parameters Through the feature matrix and cumulative reward feature vector Represented as: , In the formula, These are the estimated values ​​for the LinUCB parameters; And based on the newly added context vector and rewards Perform the following updates: , , In the formula, and The optimal intervention action is performed respectively. The corresponding feature matrix and cumulative reward feature vector; When updating using Bayesian linear regression, for each action... Maintenance action expected reward model parameters The posterior distribution, and based on the new data batch The posterior distribution is updated to obtain the updated action expectation reward model parameters. Compared with the prior art, the present invention has the following advantages: (1) By constructing an observation data sequence containing missing patterns, this invention avoids the bias caused by directly discarding or interpolating missing data in traditional methods, thereby improving the robustness of the estimation of the real state. At the same time, considering that the digital health literacy of the elderly changes dynamically over time, this invention uses a dynamic state estimation method based on recursive Bayesian filtering to perform state recursion, thereby achieving continuous and real-time state tracking, which can more accurately reflect the "current level" and "trend" of the digital health literacy of the elderly. In addition, it also considers using a contextual multi-armed slot machine model to achieve adaptive selection of personalized intervention strategies, so that it can continuously evolve with the use process.

[0020] (2) The dynamic state estimation method based on recursive Bayesian filtering in this invention can track the evolution of digital health literacy of the elderly in real time and continuously, and output the state estimate and its uncertainty, thus fundamentally overcoming the time discreteness defect of the static questionnaire method.

[0021] (3) By marking missing data for observation data, the state space model naturally supports missing observation data. Even if there is no new questionnaire data for several consecutive days, the filter can still maintain a reasonable estimate based on the state transition model. The missing pattern feature encoding further improves the model's adaptability to irregular behavioral data.

[0022] (4) The present invention significantly improves the reliability of state estimation results in sparse data scenarios by introducing an uncertainty-driven active query mechanism.

[0023] (5) By introducing an automatic detection and adaptive mechanism for state mutations, this invention can promptly identify health literacy mutations caused by diseases, emotions, etc., and automatically reset model parameters to avoid filter divergence.

[0024] (6) This invention uses a contextual multi-armed slot machine model to replace fixed rules, so that the intervention strategy can be continuously optimized according to individual feedback, achieving true "personalized approach".

[0025] (7) The evaluation, prediction, intervention and feedback links of this invention are tightly coupled through incremental update algorithm, which can continuously evolve with the accumulation of usage time and become more and more accurate. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the overall system structure of the present invention; Figure 3 This is a schematic diagram of the recursive Bayesian filter update of the present invention; Figure 4 This is a schematic diagram illustrating the optimal intervention action selection for the contextual multi-armed slot machine model of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0028] Example 1 This embodiment provides a method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, such as... Figure 1 As shown, the method includes the following steps: Step 1: Multi-source heterogeneous data acquisition Collect multi-source data from elderly users, including: (1) Basic demographic data (age, sex, education level, history of chronic diseases); (2) Scoring data of knowledge dimension, attitude dimension and behavior dimension based on the KAP (Knowledge-Attitude-Behavior Model) theoretical framework (collected periodically through a lightweight questionnaire embedded in a mobile terminal). The specific calculation method for the KAP score vector is as follows: The number of questionnaire questions for knowledge dimension K (Knowledge) is Binary scoring (correct = 1, incorrect = 0), after score normalization, we get: , In the formula, For the knowledge dimension of the scoring data, 1 indicates a correct answer, and 0 indicates an incorrect answer; Number of questionnaire items for Attitude Dimension A The Likert scale (1-5), after normalization, yields: , In the formula, For the attitude dimension of the scoring data, For the attitude dimension The original score for the question is 1 = strongly disagree, and 5 = strongly agree. Number of questionnaire items in the behavioral dimension P (Practice) The four-level frequency scores (1-4), after normalization, are: , In the formula, For behavioral dimension scoring data, For the behavioral dimension The original score for the question is 1 = never, 4 = often / always.

[0029] (3) Digital health behavior logs (APP usage frequency, course learning duration, number of online consultations, function click paths, etc.), represented as: , , , In the formula, For digital health behavior logs, For the first The weight of each behavior log item For the first The normalized value of the behavior log. , For the first Minimum and maximum values ​​in the item behavior log, For the first The value of the behavior log, when season Or add a very small positive number to the denominator. To avoid division by zero; (4) Health indicator data (blood pressure, blood sugar, heart rate, BMI, etc. are obtained through wearable devices or manual input).

[0030] Step 2: Data Preprocessing and Missing Pattern Characterization The collected data is cleaned and normalized (Min-Max standardization): Align with features. Specifically, encode missing patterns: Introducing the missing selection matrix , For diagonal matrix functions, For the first The missing data channel indicator is used in the missing selection matrix to mask missing observation channels during subsequent filtering updates. The preprocessed observation data sequence is then compared with the missing selection matrix. The data is then spliced ​​together to obtain an observation data sequence that includes the missing patterns.

[0031] By recording each data channel at time Whether or not the pattern is missing, and using the missing pattern itself as a feature input into subsequent models to reflect the regularity of elderly behavior.

[0032] Step 3: Dynamic state estimation based on recursive Bayesian filtering Treating the digital health literacy of older adults as an unobservable latent state variable (It can be one-dimensional or multi-dimensional, such as the three sub-dimensions of knowledge, attitude, and behavior.) , , , These are the true potential health literacy state variables in the dimensions of knowledge, attitude, and behavior. These variables cannot be directly observed and are obtained only through Bayesian recursive estimation using multi-source observation data. A state-space model is constructed as follows: State transition equation: ,in Here is the state transition matrix. For process noise, For process noise covariance, The distribution follows a Gaussian pattern, describing the natural evolution of health literacy.

[0033] Observation equation: ,in The observation vector is composed of the observed KAP score vector calculated from the data collected in step one and the digital health behavior log. : , Among them, the observation matrix : , Process noise covariance: , In the formula, For the observation matrix, To observe the noise, To observe the noise covariance, for Weights related to the knowledge dimension in the middle. for The weights related to the attitude dimension, for Weights related to behavioral dimensions For diagonal matrix functions, These represent the standard deviations of process noise during the evolution of potential states of knowledge, attitudes, and behaviors. These are the standard deviations of observation noise during the observation of knowledge, attitude, and behavior scores, as well as digital health behavior logs. The mean is 0 and the process noise covariance is Gaussian distribution, With a mean of 0 and an observation noise covariance of , The Gaussian distribution.

[0034] When a data channel is missing, the corresponding observed component is not updated. A recursive Bayesian filter (specifically a Kalman filter or particle filter) is used to estimate the posterior probability distribution of the state online. Output at each time step: the optimal state estimate. And estimate the covariance matrix (This is a statistical characteristic of the posterior distribution), and its diagonal elements represent the uncertainty variance of the estimates for each dimension. Among these, combining... Figure 3 As shown, the execution steps of the dynamic state estimation method of recursive Bayesian filtering include: Prediction steps: For the current moment According to the previous moment posterior probability distribution of the output , Perform state prediction and error covariance prediction, and output the time. Prior probability distribution , The expression for state prediction is: , Error covariance prediction: , In the formula, For the previous moment The observation vector of the state The optimal estimate, i.e., the estimate from the previous time step. The optimal state estimate, For the previous moment The estimated covariance matrix, For a moment The prior state estimate, For a moment The prior estimate of the covariance matrix; Filtering update steps (if observations are valid): According to time The prior probability distribution is used to calculate the prediction residual. With residual covariance , respectively represented as: , ; If time observation vector When some observation channels are missing, a missing selection matrix is ​​introduced. Missing channels are masked, and based on the residual covariance... and prior estimation of covariance matrix Calculate the Kalman gain after processing for missing observations. , is represented as: , In the formula, For the set constant, for 3D identity matrix; Based on Kalman gain and predicted residuals Update the state to obtain the current time. optimal state estimate , is represented as: , ; Based on Kalman gain Perform covariance update to obtain the current time step. Estimated covariance matrix , is represented as: , , In the formula, It is the identity matrix. Estimating the variance of the knowledge dimension state error. The variance of the state estimation error for the attitude dimension. The variance of the state estimation error for the behavioral dimension. The off-diagonal elements represent the covariance between estimation errors in different dimensions; If all observation channels are valid The Kalman gain calculation formula after missing observation processing degenerates into the standard Kalman gain calculation formula; If the current time There is no valid observation data available for filter updates, i.e., the observation vector. All observation channels are missing, or the selection matrix is ​​missing. If all diagonal elements are 0, then skip the above filtering update operation and directly set the time... Prior state estimate and time Prior estimate of covariance matrix Each as the current moment optimal state estimate and the current moment Estimated covariance matrix ,Right now: , ; Return to the above prediction steps to obtain the optimal state estimate and estimated covariance matrix for the next time step, until the end, completing the dynamic state estimation process for each time step.

[0035] Finally, the current time step is output during each filtering process. optimal state estimate Estimating the covariance matrix (Posterior) and missing observation data sequences are used to calculate the current time. Uncertainty indicators and filtering for the next time step.

[0036] Step 4: Active Inquiry Mechanism Based on Uncertainty Measurement Define uncertainty index ,in is the dimension of the state vector (i.e., the root mean square of the marginal standard deviation of each dimension). The trace of the state variance matrix. This indicates the uncertainty in the current state estimate, when Exceeding the preset threshold When the system determines that the current assessment result is too vague, it automatically generates proactive questions, such as: "Have you felt unwell in the past week?" or "Do you remember taking your blood pressure medication every day?" After the user answers, the answers are converted into pseudo-observation data, forcing a filtering update to quickly reduce uncertainty. The steps of this filtering update include: Pseudo-observation data construction: Map user responses to pseudo-observation vectors: , : Mapping function, which transforms the question-and-answer results into a mapping with the observation vector. Numerical values ​​consistent across dimensions. If a user only answers part of the question, the corresponding dimension can be set to missing, and this can be included in the missing selection matrix. (Used to mask missing pseudo-observation dimensions) marked as follows: , False observation noise covariance: , For the first The noise variance of each pseudo-observation dimension represents the uncertainty of the pseudo-observation, and can be set according to the difficulty of the question or the credibility of the user's answer.

[0037] The filter update process (pseudo-observation input) updates the recursive Bayesian filter (Kalman filter) from step three of the pseudo-observation input process once: Based on pseudo-observation vector Calculate the pseudo-observation prediction residuals and pseudo-observation residual covariance , respectively represented as: , , Based on pseudo-observation residual covariance Using the missing selection matrix, calculate the Kalman gain in the pseudo-observation update (only for the effective observation dimension): ,in, This is a regularization constant used to ensure that the matrix is ​​invertible. for 3D identity matrix; Kalman gain based on pseudo-observation updates and pseudo-observation prediction residuals Update the state and obtain the time. optimal state estimate , is represented as: ; Kalman gain based on pseudo-observation updates Perform covariance matrix update to obtain time step. Estimated covariance matrix : ; The uncertainty index has been recalculated after the update: ; if If the threshold is still exceeded, the next round of active inquiry or further supplementary observations can be triggered.

[0038] Pseudo-observation updates can quickly reduce the uncertainty of key dimensions and improve the accuracy of state estimation.

[0039] Step 5: State mutation detection and adaptive reset Calculate the observation residuals and its normalized sum of squares of residuals (Mahaviblanc distance), ,in This refers to the residual covariance. If there are two or more consecutive time points... The theoretical value that deviates significantly from the chi-square distribution (i.e. exceeds the detection threshold) If the state change occurs, it is determined that a state transition has taken place. The response method is to increase the process noise covariance. The possible values ​​of: , For adjustment coefficients, or partial reset times The estimated covariance matrix of the filter ,Right now , This is a pre-defined covariance reset matrix used to increase the uncertainty of the state estimate after detecting a sudden change in state, thus giving subsequent new observation data a higher weight in the filter update. This is a preset maximum coefficient. for The 3D identity matrix enables the filter to quickly track new system dynamics and reduce the impact of sudden changes on state estimation.

[0040] The aforementioned pseudo-observation update achieves unified updates across multiple dimensions. When an active query only yields responses for some dimensions, then... Select matrix Missing dimensions are masked, and the current state estimate is corrected using only the effective pseudo-observation dimensions.

[0041] Steps 4 and 5 above can be performed in parallel after step 3.

[0042] Step 6: Output of Health Management Risk Level and Confidence Level The current state obtained by dynamic estimation and its uncertainty Together with original behavioral characteristics and health indicators, a feature vector is constructed: Input a random forest classifier: , The specific steps include: Let the random forest classifier include A decision tree, a set of health management risk level categories for: , For any risk level category ,make Indicates time Predicting as a category in a random forest The number of decision trees, then the categories The voting percentage is expressed as follows: , In the formula, For prediction as a category The proportion of decision trees in all decision trees, The total number of trees in the random forest; eigenvectors Input a random forest classifier and output the health management risk level. and the corresponding confidence level : , , In the formula, This indicates the degree of consistency between the random forest and the final risk level prediction results. The larger the value, the more reliable the prediction result.

[0043] The prediction step also includes: Determine the confidence level Is it less than the confidence threshold? If yes, then the active inquiry mechanism in step four is triggered to obtain additional information and make a new prediction; otherwise, the subsequent steps continue.

[0044] Step 7: Adaptive Intervention for Context-Based Multi-Armed Slots This step primarily involves selecting the optimal action from multiple intervention options based on the current user state (context), and receiving reward feedback, combined with... Figure 4 As shown, the specific steps are as follows: Modeling the personalized intervention problem as a contextual bandit model: Context vector: ,in For the current state, For uncertainty, One-hot coding for health management risk levels Historical reward vector (most recent) (Reward for the next action).

[0045] Action set: Each action This corresponds to a type of intervention strategy (such as: pushing knowledge cards, initiating voice reminders, arranging telephone guidance, adjusting exercise goals, sending encouraging messages, etc.).

[0046] Reward function: Within a short time window after the action is performed (e.g., 3 days), the reward function combines behavioral changes and improvements in health indicators. ,in, , , Changes in behavioral activity (such as increased app usage time); KAP rating improved; Improved health indicators (such as blood pressure and blood sugar); Positive user feedback (clicks, completion of learning tasks); Negative feedback (no response, notifications turned off); Weights can be set based on experiments or experience to balance the contributions of various indicators. For the first The normalized improvement of a health indicator within a preset short time window The weights are assigned to the corresponding health indicators.

[0047] The strategy employs either the LinUCB (Linear Upper Confidence Bound) algorithm or Thompson sampling to dynamically select actions that maximize the expected cumulative reward based on historical reward data. When using the LinUCB algorithm: Each time the system makes a decision, it selects and executes the optimal intervention from the action set based on the current context. User feedback (clicks, learning completion rate, subsequent changes in health indicators) is immediately used to update the reward model for that action.

[0048] LinUCB Action Selection and Reward Model Update: For each action LinUCB parameter initialization (for each action) ): , , For dimension unit vector, context vector Dimensions To initialize the feature matrix, The expected cumulative reward vector; Expected reward estimation for the action: ; Based on context vectors and LinUCB parameter estimates Calculate the action UCB value And select the current time. As the optimal intervention action UCB value and optimal intervention actions They are represented as follows: ; Execute the optimal intervention action Observe rewards within a preset short time window ; Based on rewards Online updates are used to determine the optimal intervention action. The corresponding feature matrix and the expected cumulative reward vector are used to obtain the updated feature matrix. and expected cumulative reward vector , respectively represented as: , In the formula, This is an assignment update operation; in this way, the next action selection will utilize the latest reward information to adaptively optimize the intervention strategy.

[0049] Time recursion: Let Return to the UCB value calculation and action selection steps until the end.

[0050] The LinUCB mentioned above can be replaced by Thompson sampling: for each action parameter Maintain the posterior distribution and sample from the posterior distribution. Select action: The rewards update is the same as LinUCB.

[0051] Step 8: Incremental Learning and Closed-Loop Optimization Each new batch of data (e.g., daily or weekly) includes newly added observation vectors within the time series. Missing selection matrix Optimal state estimate Estimating the covariance matrix Uncertainty indicators eigenvectors Health management risk level Confidence level Context vector Optimal intervention action and rewards Perform the following incremental learning and closed-loop optimization steps: Utilizing feature vectors from new data batches The risk labels are used to construct new decision trees in the random forest to replace the old decision trees that take the longest to build or have the lowest performance, while keeping the number of decision trees in the random forest constant. The risk labels include health management risk levels confirmed manually, health management risk levels generated from follow-up results, or health management risk levels confirmed based on preset rules. Using observation vectors from new data batches Missing selection matrix Optimal state estimate Estimating the covariance matrix The state-space model parameters in the dynamic state estimation method based on recursive Bayesian filtering. , , Update; among them, Here is the state transition matrix. For process noise covariance, To observe the noise covariance, update methods include re-estimating it offline using the expectation-maximization algorithm, or updating it online using a sliding window approach. Utilizing context vectors in new data batches Optimal intervention action and rewards Update the context of each action in the multi-armed slot machine model. Corresponding action expected reward model parameters ,in: , In the formula, For expectations; When using the LinUCB algorithm, the action expectation reward model parameters Through the feature matrix and cumulative reward feature vector Represented as: , In the formula, These are the estimated values ​​for the LinUCB parameters; And based on the newly added context vector and rewards Perform the following updates: , , In the formula, and The optimal intervention action is performed respectively. The corresponding feature matrix and cumulative reward feature vector; When updating using Bayesian linear regression, for each action... Maintenance action expected reward model parameters The posterior distribution, and based on the new data batch Update the posterior distribution to obtain the updated action expectation reward model parameters.

[0052] The entire method forms a closed loop: data collection → dynamic evaluation → prediction → intervention → feedback → model update.

[0053] This embodiment uses elderly people aged 60 and above in a certain community as an example to verify the effectiveness of the above method: (1) Multi-source data acquisition Taking seniors aged 60 and above in a certain community as an example, after registering and installing the APP: Basic demographic data: Age, sex, education level, and history of chronic diseases (hypertension, diabetes, etc.) are collected through registration forms.

[0054] KAP Questionnaire: A lightweight questionnaire (15 questions in total: 5 knowledge questions (correct / incorrect), 5 attitude questions (Likelit scale), and 5 behavior questions (frequency rating scale)) is sent out on the 1st of each month. Raw K, A, and P scores are calculated after user submission. Knowledge Dimension: ， Attitude dimension: ， Behavioral dimension: .

[0055] Digital Behavior Log: Automatically records daily: app launch count, total health information browsing time, online consultation usage, and course learning progress. Uploaded in a package every 24 hours. Behavioral activity integration: , , .

[0056] Health indicators: The band automatically uploads daily average heart rate, blood pressure (if the device supports it), and steps; users can also manually enter blood glucose levels.

[0057] (2) Data preprocessing and missing pattern coding After the server receives the data: Perform Min-Max normalization on continuous numerical variables (such as browsing time and number of steps).

[0058] For questionnaire scores, use the average score (range 0-1).

[0059] Missing pattern encoding: Define a missing indicator for each data channel. 0 indicates missing, and 1 indicates a value. All channels... Missing indicators concatenated into a vector , as additional feature input filters and reward models.

[0060] (3) Dynamic estimation of recursive Bayesian filtering This example uses a linear Kalman filter (assuming that the relationship between state evolution and observation is approximately linear; if the actual relationship is nonlinear, an extended Kalman filter or particle filter can be used instead).

[0061] Model settings: State vector: That is, the true values ​​of the three potential dimensions.

[0062] State transition matrix (Random walk model) , , , (Can be updated via pre-experimentation or EM) Process noise covariance This value is estimated based on the variance of the daily changes in state in the preliminary experimental data, and can be between 0.01 and 0.05. If there is no prior information, it can be initially set to a small value (such as 0.01) and then updated using the EM (Expectation Maximization) algorithm.

[0063] Observation vector The observation matrix is ​​obtained by fusing raw questionnaire scores and behavioral characteristics (e.g., converting activity levels from behavioral logs into a 0-1 behavioral score, and then weighting it with the questionnaire's P-score). Observation noise covariance The initial values ​​were set based on the test-retest experiment. Observation vector: Observation matrix Observation noise covariance .

[0064] Recursive process (for each time unit, such as one day): predict: , ; Update (if observation data is available for the day): Calculate the Kalman gain. , ,renew ,renew If there is no observed data, the update step is skipped, and the predicted value is used directly as the estimated value. , Output: Estimated standard deviation .

[0065] (4) Active inquiry trigger Define uncertainty index (Three-dimensional mean standard deviation). If (Threshold adjustable) The system pushes a short question to the user's app: "Have you felt any changes in your health knowledge in the past week? Please rate it on a scale of 1-5." After the user answers, the answer is considered an observation (measurement noise is 0.1). The measurement noise of 0.1 corresponds to the typical error range between the subjective answer and the actual state, and is set to the observation noise. Approximately twice the number of diagonal elements, to reflect that the relative reliability of proactively requested information is slightly lower than that of objective questionnaire data. Perform an additional filter update, then zero out the trigger.

[0066] (5) Mutation detection and reset Calculate the residual vector (Only if observations are available). Calculate the squared Mahalanobis distance. In the absence of mutations, Take the threshold. (corresponding to p=0.999), if twice consecutively A mutation has been detected. The system executes: Multiply by 10, and then press daily. Exponential decay until the original value is restored, and Reset to the initial diagonal matrix .

[0067] (6) Risk prediction and confidence level The last 7 days Mean, trend slope and the last day A pre-trained random forest model (300 trees, maximum depth 10) is input with 15 dimensions of features, including diagonal elements and missing rate. The model outputs a risk category (low / medium / high). The prediction confidence is defined as the proportion of trees in the random forest that predict the majority class. If this proportion is below 0.6, an active query is triggered (e.g., "Please upload your blood pressure readings for the last three days"), and the prediction is re-evaluated after obtaining new data. The above parameters can be adjusted according to the actual data scale; this embodiment does not constitute a limitation on the parameter range.

[0068] (7) Adaptive intervention for context-based multi-armed slot machines Action set: 6 actions in total.

[0069] a1: Push a short health knowledge video (1 minute); a2: Send a notification ("It's time to check your blood pressure today"); a3: Push positive reinforcement ("You did a great job, keep it up"); a4: Initiate an in-app voice call (a 5-minute interview will be conducted by AI customer service); a5: Send a short interactive quiz (3 questions); a6: No intervention (control); Context feature vector ; Action set: (Health videos, behavioral reminders, attitude incentives, voice calls, interactive quizzes, no intervention).

[0070] Algorithm: LinUCB (Linear Upper Confidence Bound) Parameter initialization: , ; Expected reward estimate: ; UCB value calculation: ; Select action: ; Reward function definition: Within 3 days after the action is executed, , The sum of the rewards is the actual observed value: ; Online parameter update: Observation reward Then, update the LinUCB parameters: , ; Use the latest action selection next time. The UCB value is automatically adjusted to form a closed-loop adaptive optimization.

[0071] Implementation details: The system will automatically make a decision for each active user at 8:00 AM every day and push the corresponding optimal intervention action. Feedback data was collected for three consecutive days after the intervention was implemented to calculate the reward. Reward information is updated online. With confidence intervals, the system learns over a long period the most effective intervention combination for each user.

[0072] (8) Incremental learning and closed-loop optimization Daily updates: Newly acquired observation data (questionnaires, behavior logs, health indicators) are used to update the state estimates of the Kalman filter without requiring retraining.

[0073] Weekly Updates: Using the data accumulated over the past week (context, intervention actions, rewards), batch ridge regression is used to update the LinUCB parameters for each action.

[0074] Monthly Update: Retraining the random forest risk prediction model (using all user data from the past three months, full retraining), while simultaneously re-estimating the filtering model using the EM algorithm. , , Parameters. Note that when performing full monthly training, a different time slice should be used than that used for weekly incremental updates (e.g., weekly updates use the most recent week, while monthly retraining uses the past three months excluding the most recent week) to avoid data leakage.

[0075] Based on the above method, this embodiment also provides the following complete user case, taking an elderly person as an example: Background: Mr. Li, 72 years old, with mild cognitive impairment and a history of hypertension, takes antihypertensive medication daily. Week 1 of using this system.

[0076] Days 1-3: The system obtains initial data through questionnaires and wristbands, and performs filtering and estimation. The uncertainty is low. The risk prediction is "medium risk" (behavioral score is low). LinUCB selects action a2 (behavioral reminder) and sends a message: "You haven't measured your blood pressure today." After seeing the reminder, Grandpa Li measures and uploads his blood pressure, receiving a reward of 0.5.

[0077] Days 4-6: Grandpa Li forgot to charge his bracelet, resulting in no data for three consecutive days, and he did not open the app (questionnaire missing). Filtered prediction: state estimation (slowly decaying random walk), uncertainty. The uncertainty rose from 0.08 to 0.22, exceeding the 0.15 threshold. The system triggered an active inquiry: "How have you been feeling these past few days? Is your blood pressure normal?" Grandpa Li answered "normal." The system treated the answer as a false observation, and after updating, the uncertainty decreased back to 0.12.

[0078] Day 7: Grandpa Li recharged, data recovered. After filter update, residual... Slightly larger but no mutation triggered. The system detected that the behavior score was still low, the action was A5 (interactive quiz), and pushed three questions about hypertension. Grandpa Li answered two questions correctly, the system received a positive reward, and updated. .

[0079] After one month of continuous operation, the system gradually learned that the most effective intervention for Grandpa Li was the combination of A2 (behavioral reminders) and A5 (interactive quizzes), while A1 (videos) was downgraded due to Grandpa Li's poor eyesight and infrequent clicks. Grandpa Li's... The risk level was gradually increased from 0.6 to 0.75, and the risk level was reduced to "low risk".

[0080] Extensions and variations: Filtering algorithm variations: For nonlinear relationships, unscented Kalman filtering (UKF) or particle filtering can be used. The latter is particularly suitable for highly non-Gaussian noise scenarios and can be implemented with a small number of particles (e.g., 200).

[0081] Multi-user federated learning: In scenarios with high privacy protection requirements, LinUCB model updates can be deployed on the user's end, uploading only the anonymized gradient information, and the server aggregates to generate a global baseline model.

[0082] Integration with electronic medical records: Diagnostic information and medication records are obtained through hospital interfaces and used as additional observation variables input into the filter to improve estimation accuracy.

[0083] Example 2 This embodiment provides a dynamic assessment and adaptive intervention system for digital health literacy among the elderly based on recursive Bayesian filtering, such as... Figure 2 As shown, the system includes a multi-source data acquisition layer (questionnaire, APP logs, wearable devices), a data preprocessing and missing pattern encoding module, a recursive Bayesian filtering dynamic evaluation module (including a state estimator, uncertainty calculator, and mutation detector), an active inquiry management module, a random forest risk prediction module, a contextual multi-armed slot machine intervention strategy module, and an incremental learning and feedback update module. The multi-source data acquisition layer is used to execute step one of Embodiment 1 above; the data preprocessing and missing pattern encoding module is used to execute step two of Embodiment 1 above; the recursive Bayesian filtering dynamic evaluation module is used to execute steps three to five of Embodiment 1 above, wherein the state estimator is used to execute step three of Embodiment 1 above, the uncertainty calculator is used to execute the uncertainty index calculation in step four of Embodiment 1 above, and the mutation detector is used to execute step five of Embodiment 1 above; the active inquiry management module is used to execute the active inquiry and first-time filtering update step in step four of Embodiment 1 above; the random forest risk prediction module is used to execute step six of Embodiment 1 above; the contextual multi-armed slot machine intervention strategy module is used to execute step seven of Embodiment 1 above; and the incremental learning and feedback update module is used to execute step eight of Embodiment 1 above.

[0084] In addition, this system adopts a cloud-edge collaborative architecture: On the user side: Seniors interact with the system through a health management APP on their smartphones (supports iOS / Android). The APP includes a questionnaire module, a message push module, and an automatic data collection SDK (log upload).

[0085] Wearable devices: Compatible with mainstream smart bracelets / watches (such as Xiaomi Mi Band and Apple Watch), and can synchronize data such as heart rate, blood pressure, and steps with the APP via Bluetooth.

[0086] Cloud server: Deploys data receiving API, database (MySQL + Redis), and algorithm computing cluster (Python + NumPy + Scikit-learn + TensorFlow), responsible for performing dynamic filtering, prediction, and intervention decisions.

[0087] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] 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 implemented 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. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0089] 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 processor, 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 and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] 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 method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, characterized in that, Includes the following steps: S1. Obtain multi-source heterogeneous observation data of elderly users and construct an observation data sequence containing missing patterns; S2. Based on the observation data sequence containing the missing pattern, the digital health literacy of the elderly is dynamically estimated using a dynamic state estimation method based on recursive Bayesian filtering to obtain the optimal state estimate and the estimated covariance matrix at the current moment. S3. Based on the estimated covariance matrix at the current time. Calculate the uncertainty index at the current moment; S4. Determine whether the uncertainty index exceeds a preset threshold. If not, proceed to step S6; if yes, trigger the active query mechanism at the current moment and perform a filter update. S5. Recalculate the uncertainty index based on the filter update result, return to step S4, until it does not exceed [the specified value]. ; S6. Based on the optimal state estimate and estimated covariance matrix at the current moment, predict the risk prediction result at the current moment, and combine the uncertainty index to adaptively select personalized intervention strategies using a contextual multi-armed slot machine model.

2. The method for dynamic assessment and adaptive intervention of digital health literacy of the elderly based on recursive Bayesian filtering according to claim 1, characterized in that, The multi-source heterogeneous observational data includes basic demographic data, multi-dimensional scoring data based on the KAP theoretical framework, digital health behavior logs, and health indicator data. The basic demographic data includes age, sex, education level, and history of chronic diseases; The multi-dimensional scoring data includes scoring data for the knowledge dimension, scoring data for the attitude dimension, and scoring data for the behavior dimension, which are represented as follows: ， ， ， In the formula, For the knowledge dimension of the scoring data, For the attitude dimension of the scoring data, For the behavioral dimension of the scoring data, For knowledge dimension The number of questionnaire questions 1 indicates a correct answer, and 0 indicates an incorrect answer. Attitude dimension The number of questionnaire questions For the attitude dimension The original score of the question. For behavioral dimensions The number of questionnaire questions For the behavioral dimension The original score of the question; The digital health behavior log includes APP usage frequency, course learning duration, number of online consultations, and function click paths. The digital health behavior log is represented as follows: ， ， ， In the formula, For digital health behavior logs, For the first The weight of each behavior log item For the first The normalized value of the behavior log. , For the first Minimum and maximum values ​​in the item behavior log, For the first The value of the item behavior log; The health indicators include blood pressure, blood sugar, heart rate, and BMI.

3. The method for dynamic assessment and adaptive intervention of digital health literacy of the elderly based on recursive Bayesian filtering according to claim 1, characterized in that, The steps for constructing the observation data sequence containing the missing pattern include: The multi-source heterogeneous data is sequentially cleaned, normalized, and feature-aligned to obtain a preprocessed observation data sequence. For the preprocessed observation data sequence, a missing selection matrix is ​​introduced to characterize whether there are missing data channels. Represented as: ， In the formula, For diagonal matrix functions, For the first Missing data channel indication; The preprocessed observation data sequence and the missing selection matrix are then compared. The data is then spliced ​​together to obtain an observation data sequence that includes the missing patterns.

4. The method for dynamic assessment and adaptive intervention of digital health literacy of the elderly based on recursive Bayesian filtering according to claim 1, characterized in that, The execution steps of the dynamic state estimation method based on recursive Bayesian filtering include: definition: Digital health literacy among older adults is set as an unobservable latent state variable. And construct a state-space model, represented as: ， in: ， ， ， In the formula, , , These are the actual potential health literacy status variables in the dimensions of knowledge, attitude, and behavior. For the current moment The state transition matrix, For process noise, For process noise covariance, The mean is 0 and the process noise covariance is Gaussian distribution, With a mean of 0 and an observation noise covariance of , Gaussian distribution, For a moment The observation vector, For the observation matrix, To observe the noise, To observe the noise covariance, For the knowledge dimension of the scoring data, For the attitude dimension of the scoring data, For the behavioral dimension of the scoring data, For digital health behavior logs, for Weights related to the knowledge dimension in the middle. for The weights related to the attitude dimension, for Weights related to behavioral dimensions For diagonal matrix functions, These represent the standard deviations of process noise during the evolution of potential states of knowledge, attitudes, and behaviors. These are the standard deviations of observation noise during the observation of knowledge, attitude, and behavior scores, as well as digital health behavior logs. predict: For the current moment According to the previous moment posterior probability distribution of the output , Perform state prediction and error covariance prediction, and output the time. Prior probability distribution , The expression for the state prediction is: ， The expression for the error covariance prediction is: ， In the formula, For the previous moment The observation vector of the state The optimal estimate, i.e., the estimate from the previous time step. The optimal state estimate, For the previous moment The estimated covariance matrix, For a moment The prior state estimate, For a moment The prior estimate of the covariance matrix; Filter update: According to the time The prior probability distribution is used to calculate the prediction residual. With residual covariance , respectively represented as: ， ， If time observation vector When some observation channels are missing, a missing selection matrix is ​​introduced. Missing channels are masked, and based on the residual covariance... and prior estimation of covariance matrix Calculate the Kalman gain after processing for missing observations. , represented as: ， In the formula, For the set constant, for 3D identity matrix; Based on the Kalman gain and predicted residuals Update the state to obtain the current time. optimal state estimate , represented as: ， ， Based on the Kalman gain Perform covariance update to obtain the current time step. Estimated covariance matrix , represented as: ， ， In the formula, It is the identity matrix. The variance of the estimation error for the knowledge dimension state. The variance of the state estimation error for the attitude dimension. The variance of the state estimation error for the behavioral dimension. The off-diagonal elements represent the covariance between estimation errors in different dimensions; If all observation channels are valid The Kalman gain calculation formula after missing observation processing degenerates into the standard Kalman gain calculation formula; If the current time There is no valid observation data available for filter updates, i.e., the observation vector. All observation channels are missing, or the selection matrix is ​​missing. If all diagonal elements are 0, then skip the above filtering update operation and directly set the time... Prior state estimate and time Prior estimate of covariance matrix Each as the current moment optimal state estimate and the current moment Estimated covariance matrix ,Right now: , ; Return to the above prediction steps to obtain the optimal state estimate and estimated covariance matrix for the next time step, until the end, completing the dynamic state estimation process for each time step.

5. A method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, as described in claim 4, is characterized in that... It also includes state mutation detection and adaptive reset steps, specifically including the following: During the dynamic estimation process, based on the predicted residual With residual covariance The normalized sum of squares of the residuals is calculated and expressed as: ， In the formula, This is the normalized sum of squares of the residuals; When there are two or more consecutive moments... Exceeding the detection threshold If a state change occurs, the process noise covariance is increased. ,Right now , For adjustment coefficients, Or partially reset the aforementioned time. Estimated covariance matrix ,Right now , The pre-defined covariance reset matrix, This is a preset maximum coefficient. for 3D identity matrix.

6. The method for dynamic assessment and adaptive intervention of digital health literacy of the elderly based on recursive Bayesian filtering according to claim 1, characterized in that, The uncertainty index is expressed as follows: ， In the formula, As an indicator of uncertainty, For a moment The estimated covariance matrix, is the dimension of the state vector.

7. A method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, as described in claim 1, is characterized in that... The steps for performing a filter update include: Once the proactive inquiry mechanism is triggered, the elderly user's response is mapped to a pseudo-observation vector, represented as: ， In the formula, This is a pseudo-observation vector, whose dimension is the same as the observation vector. Consistent For mapping functions, In response; For the pseudo-observation vector Introducing the missing selection matrix The missing selection matrix is ​​used to characterize whether there are missing responses from elderly users. Represented as: ， Based on the pseudo-observation vector Calculate the pseudo-observation prediction residuals and pseudo-observation residual covariance , respectively represented as: ， ， In the formula, For the observation matrix, For a moment The prior state estimate, For a moment The prior estimate of the covariance matrix, The pseudo-observation noise covariance matrix, For the first Noise variance of each pseudo-observation dimension; Based on the pseudo-observation residual covariance and missing selection matrix Calculate the Kalman gain in the pseudo-observation update. , represented as: ， In the formula, Here is the regularization constant. for 3D identity matrix; Kalman gain based on the pseudo-observation update and pseudo-observation prediction residuals Update the state and obtain the time. optimal state estimate , represented as: ， Kalman gain based on the pseudo-observation update Perform covariance update to obtain time step. Estimated covariance matrix This completes one filter update process, where time... Estimated covariance matrix Represented as: 。 8. A method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, as described in claim 1, is characterized in that... The prediction steps for the risk prediction result at the current moment include: Based on the current time The optimal state estimate and the estimated covariance matrix, combined with the time step Observation vectors, construct feature vectors ,in, For a moment The optimal state estimate, For a moment The estimated covariance matrix, For a moment The observation vector; Let the random forest classifier include A decision tree, a set of health management risk level categories for: ， For any risk level category ,make Indicates time Predicting as a category in a random forest The number of decision trees, then the categories The voting percentage is expressed as follows: ， In the formula, For prediction as a category The proportion of decision trees to all decision trees; The feature vector Input the random forest classifier and output the risk prediction result at the current time, including the health management risk level. and the corresponding confidence level The health management risk level mentioned therein Represented as: ， In the formula, The total number of trees in the random forest; The confidence level Represented as: ， The prediction step also includes: Determine the confidence level Is it less than the confidence threshold? If yes, then the active query mechanism in step S4 is triggered to re-predict; otherwise, the subsequent steps continue.

9. A method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, as described in claim 1, is characterized in that... The specific execution steps of the contextual multi-armed slot machine model include: definition: Based on the optimal state estimate at the current moment Uncertainty indicators Health management risk level And history most recently The reward vector for the next action is used to construct the current time step. context vector ,in, Health management risk level one-hot encoding, For the most recent history The reward vector for this action; Set action set Each of the actions A corresponding type of intervention strategy; Set Action After execution, the reward function, constructed by combining behavioral changes, KAP score changes, health indicator improvements, and user feedback within a preset short time window, is expressed as: ， in: ， ， In the formula, For action The corresponding rewards This represents the change in behavioral activity. To increase KAP ratings For the improvement of health indicators, In response to positive user feedback, As negative feedback, For the weights of the reward function, For the first The normalized improvement of a health indicator within a preset short time window The weights for the corresponding health indicators; initialization: For each action Initialize the feature matrix and expected cumulative reward vector ,in For dimension The identity matrix, context vector The dimension; UCB value calculation and action selection: For each action According to the feature matrix and cumulative reward feature vector The estimated values ​​of the LinUCB parameters were obtained. , represented as: ， Based on the context vector and LinUCB parameter estimates Calculate the action UCB value and select the current time. As the optimal intervention action The UCB value and optimal intervention actions They are represented as follows: ， ， In the formula, To explore the adjustment coefficient; Action execution and observation: Perform the optimal intervention action And observe rewards within a preset short time window. ; Online updates: Based on the aforementioned rewards Online updates are used to determine the optimal intervention action. The corresponding feature matrix and expected cumulative reward vector are used to obtain the updated feature matrix. and expected cumulative reward vector , respectively represented as: ， ， In the formula, For assignment / update operations; Time progression: make Return to the UCB value calculation and action selection steps until the end.

10. A method for dynamic assessment and adaptive intervention of digital health literacy among the elderly based on recursive Bayesian filtering, as described in claim 1, is characterized in that... It also includes incremental learning and closed-loop optimization steps, specifically including the following: Acquire new batches of data within a set interval, including newly added observation vectors in the time series. Missing selection matrix Optimal state estimate Estimating the covariance matrix Uncertainty indicators eigenvectors Health management risk level Confidence level Context vector Optimal intervention action and rewards ; Using the feature vectors in the new data batch The risk labels are used to construct new decision trees in the random forest to replace the old decision trees that take the longest to build or have the lowest performance, while keeping the number of decision trees in the random forest constant. The risk labels include health management risk levels confirmed manually, health management risk levels generated from follow-up results, or health management risk levels confirmed based on preset rules. Using the observation vectors in the new data batch Missing selection matrix Optimal state estimate Estimating the covariance matrix The state-space model parameters in the dynamic state estimation method based on recursive Bayesian filtering are... , , Update; among them, Here is the state transition matrix. For process noise covariance, To observe the noise covariance, update methods include re-estimating it offline using the expectation-maximization algorithm, or updating it online using a sliding window approach. Utilizing the context vector in the new data batch Optimal intervention action and rewards Update the context of each action in the multi-armed slot machine model. Corresponding action expected reward model parameters ,in: ， In the formula, For expectations; When using the LinUCB algorithm, the action expected reward model parameters Through the feature matrix and cumulative reward feature vector Represented as: ， In the formula, These are the estimated values ​​for the LinUCB parameters; And based on the newly added context vector and rewards Perform the following updates: ， ， In the formula, and The optimal intervention action is performed respectively. The corresponding feature matrix and cumulative reward feature vector; When updating using Bayesian linear regression, for each action... Maintenance action expected reward model parameters The posterior distribution, and based on the new data batch Update the posterior distribution to obtain the updated action expectation reward model parameters.

Images