Self-adaptive decision path planning method and system based on multi-level risk

Abstract

The invention discloses a self-adaptive decision path planning method and system based on multistage risks, and the method comprises the steps: carrying out the dynamic evaluation and grading early warning of the disability risk of a target object through a gray clustering algorithm and long and short-term memory network coupled early warning model; based on a risk assessment result and a disability prevention and control knowledge base, generating an initial decision path by solving a multi-objective optimization function integrating expected effectiveness, performability and burden degree; during path execution, collecting effect deviation and execution deviation, and dynamically adjusting the weight of the multi-objective optimization function by using a fuzzy weight controller to reconstruct a decision path; and based on the group feedback data, iteratively updating the knowledge base. According to the method, the problems of static disability risk assessment, rigid intervention scheme and management scene splitting in the prior art are solved, and dynamic and accurate early warning of the elderly disability risk, depth self-adaptive adjustment of the intervention path and full-period closed-loop management are realized.

Description

Technical Field

[0001] This invention relates to the application of artificial intelligence deep learning in physiological health data processing, specifically to an adaptive decision path planning method and system based on multi-level risk. Background Technology

[0002] With the increasing aging of the population, disability among the elderly has become a challenge for society and families. Effective prevention and control of disability hinges on early and accurate risk identification and the provision of continuous, full-cycle interventions. However, current technologies and management systems used in the prevention and control of disability among the elderly generally suffer from several technical problems.

[0003] First, existing methods for assessing disability risk are mostly static. Assessments typically rely on standardized scales such as the Barthel Index, conducted as a single or intermittent evaluation at specific time points. This approach struggles to capture the dynamic process of disability risk evolving continuously over time. For example, subtle changes in an individual's progression from mild functional impairment to severe disability are often overlooked, making it impossible to effectively warn of those with latent disabilities on the path to rising risk. Furthermore, traditional statistical models or single machine learning models have limited capacity to handle the common characteristics of "small sample size and limited information" in elderly health data—such as atypical early symptoms and sparse data—resulting in insufficient accuracy and guidance in the assessment.

[0004] Secondly, existing intervention programs generally suffer from fragmentation and rigidity. Most intervention recommendation systems provide isolated, project-based suggestions, such as recommending individual rehabilitation exercises or dietary plans, lacking a progressive and systematic decision-making path that evolves according to individual conditions. Furthermore, once these intervention programs are generated, they are usually fixed and lack a mechanism for adaptive and dynamic adjustment based on feedback from individuals during implementation, such as changes in physiological indicators, adherence to the program, and subjective feelings. When intervention effects are unsatisfactory or users find it difficult to adhere, the system lacks the ability to comprehensively weigh multiple factors and cannot finely adjust the intensity, content, and form of the intervention strategy like human experts can.

[0005] Finally, information silos exist between various scenarios related to disability prevention and control, with data and business processes fragmented. Professional assessment results from medical institutions, follow-up data from community health service centers, and the daily care and rehabilitation implementation status of users in their home environments are often scattered across different information systems, failing to form an effective data loop. This fragmentation makes it impossible to build a unified decision support system that connects the "hospital-community-home" three-way linkage, hindering the achievement of a closed-loop management system covering the entire lifecycle from risk monitoring, intelligent decision-making, intervention implementation to effect feedback and program optimization.

[0006] Therefore, how to solve the technical problems of static disability risk assessment, fragmented intervention plans, and disconnected management scenarios in existing technologies, and provide a solution that can achieve dynamic and accurate risk early warning, deep adaptive adjustment of decision-making paths, and closed-loop management throughout the entire life cycle, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, such as discontinuous risk assessment of disability in the elderly, rigid intervention programs, and non-closed-loop management processes, by providing a method and system that can achieve dynamic risk assessment, adaptive path planning, and continuous evolution of the knowledge base.

[0008] To address the aforementioned technical problems, one aspect of the present invention provides an adaptive decision path planning method based on multi-level risk. This method can be executed in a computing device and specifically includes the following steps: First, a full-cycle risk grading and early warning model for disability is constructed and applied. Multi-dimensional risk indicators are collected and extracted from the health data of the target population. Static risk grading of physiological indicators is performed, and the evolution trend of risk scores is predicted, thereby dynamically assessing the risk status of the target population. In a specific execution step, data construction and preprocessing are performed first. Multi-dimensional risk indicators are collected and extracted from elderly health cohort data. Specifically, these indicators can be structured into three-dimensional risk indicators, including disability level (G) to characterize the severity of disability, disability characteristics (F) to characterize specific functional impairments, and time intervals (T) to characterize the time span of disability occurrence or progression. Based on these indicators, a structured dataset in the format of a "risk factor time series-disability state matrix" is constructed for subsequent model training and application.

[0009] Subsequently, a grey clustering algorithm was applied to perform static risk classification on the collected multidimensional physiological indicators. Addressing the "information-poor" and uncertainties in elderly health data, the grey whitening weight function from grey system theory was used to calculate the weight coefficients for each physiological indicator (e.g., X1, X2, ..., Xn) belonging to multiple preset risk levels (e.g., low risk, medium risk, high risk). Based on the calculated weight coefficients, a grey clustering coefficient vector C was constructed, and combined with the preset scores for each level, the comprehensive risk score of the target object at the current moment was calculated, denoted as R(t). This step, by transforming multiple discrete or fuzzy qualitative physiological signs into a continuous, quantitative comprehensive risk score sequence, helps to solve the static and coarse-grained problems of traditional scale assessments.

[0010] After obtaining the comprehensive risk score sequence, a Long Short-Term Memory (LSTM) network is further applied for time-series early warning analysis. The comprehensive risk score sequence within a historical time window, such as {R(tk), ..., R(t)}, is used as input to the LSTM network. Leveraging the network model's ability to capture long-term temporal dependencies, the evolution trend of the risk scores is learned and modeled to predict the risk score at the next time step, denoted as R_hat(t+1). After obtaining the actual observed value R(t+1) at the future time step, the residual between the predicted and observed values ​​can be calculated, defined as ε(t) = |R(t+1) - R_hat(t+1)|. Based on the statistical distribution of the residual or the slope of the predicted trend, the warning level can be divided into multiple levels. For example, it can be divided into Level 0 warning (corresponding to fluctuations in factors, indicating the need for attention), Level 1 warning (corresponding to abnormal risk syndromes, indicating the need for screening), Level 2 warning (corresponding to a significant increase in risk, indicating the need for intervention), and Level 3 warning (corresponding to disability having occurred or progressed, indicating the need for care), thereby achieving dynamic and forward-looking monitoring of disability risk.

[0011] Secondly, a structured knowledge base for disability prevention and control is constructed, and intervention measures are extracted and quantified with multi-dimensional attributes to form an entity-relationship knowledge graph that supports machine reasoning. This step includes extracting and organizing intervention measures from clinical guidelines, academic literature, and expert experience to form an intervention measure library. The intervention measure library can cover various intervention types, such as exercise interventions (e.g., balance training, resistance training), nutritional interventions (e.g., protein supplementation, specific dietary adjustments), cognitive interventions (e.g., memory training), and appropriate traditional Chinese medicine techniques (e.g., acupuncture, massage).

[0012] To enable algorithms to understand and invoke these interventions, each intervention in the library needs multi-dimensional attribute labeling and quantification. Specifically, each intervention can be labeled with its applicable scenario attributes, such as home scenarios (typically requiring simple and easy-to-operate equipment), community scenarios (potentially requiring professional guidance), or hospital scenarios (relying on specialized equipment and environment). Furthermore, each intervention is assigned a series of computable quantitative indicators, such as the expected effectiveness index (I_eff), used to quantify the expected effect of the intervention on improving specific disability characteristics; the burden index (I_bur), used to quantify the physiological or psychological burden that users may experience when implementing the intervention; and resource demand levels, etc. Through the above structured processing, an entity-relationship knowledge graph supporting machine reasoning is constructed.

[0013] Furthermore, based on the risk status and the disability prevention knowledge base, an initial decision path is generated for the target object by solving a multi-objective optimization function. The multi-objective optimization function evaluates the overall value of the candidate decision path by combining the expected total effectiveness, comprehensive executability score, total burden required for execution, and corresponding weight coefficients. In a specific execution step, firstly, based on the risk status output by the risk classification and early warning model (e.g., "moderate risk - impaired walking function") and the user's scenario (e.g., "at home"), a set of candidate intervention measures is selected from the knowledge base.

[0014] Subsequently, a multi-objective optimization function, denoted as J, is established to evaluate the merits of decision paths. This function comprehensively evaluates the overall value of a candidate path (P) composed of multiple intervention measures arranged in a time sequence. The optimization function J can be expressed as: J = w1 F_effect(P) + w2 F_exec(P) - w3 F_burden(P). In this formula, F_effect(P) represents the expected total effectiveness of path P in improving the target disability characteristics, and its value can be calculated by summing the expected effectiveness indices I_eff of all interventions in the path; F_exec(P) represents the comprehensive executability score of path P in the current scenario, which can be obtained by weighted average of the executability scores of each intervention in the path; F_burden(P) represents the total burden required for the user to execute the entire path P, and its value can be calculated by summing the burden indices I_bur of all interventions in the path. w1, w2, and w3 are the weight coefficients corresponding to effectiveness, executability, and burden, respectively, and w1+w2+w3=1. Their initial values ​​can be set according to general experience or user preferences, for example, set to a balanced configuration (0.4, 0.3, 0.3). Finally, heuristic search algorithms such as genetic algorithms and particle swarm optimization can be used to solve for the sequence of interventions that maximizes the objective function J value under the current constraints. This sequence is the initial decision path P_opt.

[0015] Another important aspect of this invention is the establishment of a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback. This mechanism adjusts the weight coefficients and reconstructs the initial decision path. Specifically, it includes: calculating the feedback deviation based on collected feedback data, outputting the adjustment amount for the weight coefficients through a fuzzy weight controller, updating the weight coefficients accordingly, and reconstructing the decision path. Specifically, this method includes periodically collecting feedback data during the intervention execution process and reconstructing the decision path based on the feedback data. In a specific execution step, the feedback deviation needs to be calculated first. The feedback deviation can include effect deviation and execution deviation. For example, the effect deviation (defined here as e1) can be determined by the difference between the target improvement rate of Activities of Daily Living (ADL) and the actual monitored ADL improvement rate. Simultaneously, the execution deviation (defined here as e2) can be determined by the difference between the preset target intervention execution rate and the actual recorded execution rate.

[0016] After obtaining the aforementioned feedback deviations, the effect deviation e1 and execution deviation e2 are used as inputs to the fuzzy weight controller. The fuzzy weight controller is a core adjustment module with a pre-set fuzzy rule base. This rule base can simulate the decision-making logic of domain experts, and its output is the adjustment amount of each weight coefficient in the aforementioned multi-objective optimization function, namely Δw1, Δw2, and Δw3. For example, the technical logic of the fuzzy rule base is: when user cooperation is high but the effect is poor, stronger and more targeted intervention measures should be provided; conversely, if the execution deviation is large, it indicates that the current path may be too difficult or inconvenient for the user to execute, and in this case, the executability weight should be increased. This invention, by introducing this fuzzy weight controller, achieves parameter-level adaptive optimization of the decision path, thereby enabling a balance between the effectiveness, executability, and burden of intervention based on individual dynamic feedback.

[0017] Furthermore, to enhance the optimization capability and generalization of the fuzzy weight controller, this invention introduces an Adaptive Neural Fuzzy Inference System (ANFIS) to optimize the fuzzy weight controller online, achieving adaptive updates to the adjustment logic itself. Specifically, the system continuously gathers feedback data from group users and the long-term intervention effects of corresponding weight adjustment strategies. Utilizing this data, the Adaptive Neural Fuzzy Inference System can act as a learning mechanism to adjust and optimize the membership functions and fuzzy rules within the fuzzy weight controller. For example, by analyzing the final ADL improvement and path adjustment stability indicators produced by different fuzzy control strategies in groups with similar risk characteristics, the system uses algorithms such as gradient descent to iteratively update the internal parameters of the fuzzy system. In this way, the decision logic of the fuzzy controller no longer relies solely on pre-set expert knowledge but can learn and evolve a better weight adjustment model from large-scale empirical data, thereby optimizing the system's decision parameters.

[0018] Finally, based on empirical data from the group of users, the disability prevention knowledge base is self-evolved and updated, including updating intervention parameters and discovering implicit taboo rules. This method utilizes real-world data from all users to periodically iteratively update the structured disability prevention knowledge base, correcting biases in prior knowledge and discovering new knowledge. In a specific implementation step, the collected execution records of the group of users are first cleaned and stratified to select high-confidence samples, such as intervention periods with execution rates exceeding a certain threshold (e.g., 85%). These are then divided into positive samples (significant improvement in ADL) and negative samples (stagnant or worsening ADL improvement) based on the intervention effect.

[0019] This invention achieves dual self-evolution of the knowledge base by constructing a population-based learning system. Firstly, it updates intervention parameters using a Bayesian posterior correction method. For the prior parameters of each intervention in the knowledge base, such as the expected effectiveness index I_eff and the burden index I_bur, they are corrected using the observed means from the population empirical data. Specifically, the Bayesian smoothing formula can be used: I_new = λ I_prior + (1-λ) I_observed, where I_prior is an initial value set based on guidelines or expert experience, I_observed is the measured mean from the population, and λ is a confidence factor. This method addresses the potential discrepancy between theoretical parameters in evidence-based medicine guidelines and real-world application effects. Secondly, it utilizes association rule mining algorithms to discover implicit taboo rules. By performing association analysis on features in the negative sample set, such as using the Apriori algorithm, strong association combinations leading to adverse intervention outcomes can be identified. For example, it might be found that the combination "a patient with grade II hypertension performing a specific high-intensity activity B in the morning" is strongly associated with "a sudden drop in compliance" or "the occurrence of adverse events." The system can automatically generate new taboo rules from such findings and store them in the knowledge base, proactively avoiding such high-risk combinations when planning pathways for similar users in the future.

[0020] Another aspect of the present invention provides an adaptive decision-making path planning system based on multi-level risk. The system may include: a multi-source data acquisition module for collecting users' health data and intervention feedback data; a risk calculation engine configured with the aforementioned gray clustering and long short-term memory network coupling model; an adaptive decision engine internally deployed with the aforementioned multi-objective optimization algorithm and fuzzy weight controller; a knowledge evolution engine for executing the aforementioned Bayesian parameter correction and association rule mining algorithms; and a multi-terminal interaction module for providing interactive interfaces for different roles (such as doctors and users).

[0021] This invention constructs a system encompassing dynamic risk assessment, personalized path generation, adaptive feedback adjustment, and population knowledge evolution through the aforementioned technical solutions. First, the adoption of a "grey clustering-LSTM" coupled model helps address the challenges of accurate risk identification and dynamic early warning in the context of limited data on elderly health. Second, the introduction of a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback, particularly through optimization of the fuzzy controller itself using ANFIS, achieves deep adaptation of the decision logic. This enables intervention programs to be intelligently and precisely adjusted based on individual feedback, alleviating the rigidity of traditional intervention programs. Finally, the establishment of a dual self-evolutionary method for the knowledge base based on population empirical data helps ensure continuous iteration and optimization of system knowledge and supports data integration and management across multiple scenarios, including hospitals, communities, and home care. This method and system can provide accurate, dynamic, and personalized decision support for the prevention and control of disability in the elderly. Attached Figure Description

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] Figure 1 This is a functional architecture diagram of an adaptive decision path planning system according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the implementation process of the adaptive decision path planning method according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the hardware structure of an adaptive decision path planning system according to an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram of a dual self-evolution mechanism for a knowledge base according to an embodiment of the present invention.

[0028] Figure 6 This is a training error convergence curve of the Adaptive Neural Fuzzy Inference System (ANFIS) according to an embodiment of the present invention.

[0029] Figure 7 This is a system data acquisition and device connection topology diagram according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Example 1 This embodiment provides an adaptive decision path planning method based on multi-level risk, which can be executed on a computing device, such as... Figure 2 As shown, the specific steps may include: Step S100 involves constructing and applying a full-cycle risk grading and early warning model for disability to dynamically assess the risk status of the target object. This step addresses the static nature of traditional assessment methods, enabling accurate capture and forward-looking early warning of risk evolution trends.

[0032] In a specific implementation step, the first step is data construction and preprocessing. Multi-dimensional risk indicators are collected and extracted from elderly health cohort data. These indicators are not limited to traditional medical examination indicators but can also include daily living behavior data, environmental data, etc. Specifically, these indicators can be structured into a three-dimensional risk index, including a disability level G to characterize the severity of disability, a disability feature F to characterize specific functional impairments, and a time interval T to characterize the time span of disability occurrence or progression. For example, disability level G can be divided into "no disability," "mild disability," and "moderate disability"; disability feature F can be specified as "decreased walking ability," "memory decline," and "fine motor impairment"; and the time interval T records the time point of each data collection. Based on these indicators, a structured dataset in the format of a "risk factor time series-disability state matrix" is constructed, which provides effective data input for subsequent model training.

[0033] Subsequently, a grey clustering algorithm was applied to perform static risk classification on the collected multidimensional physiological indicators. Health data of the elderly often exhibits "information-poor" characteristics such as small sample size, ambiguous relationships between indicators, and incomplete data. Grey system theory is suitable for handling such uncertainties. Specifically, using the grey whitening weight function in grey system theory, the weight coefficients of each physiological indicator, such as X1 (e.g., gait speed), X2 (e.g., grip strength), ..., Xn (e.g., Mini-Mental State Examination score), belonging to multiple preset risk levels (e.g., low risk, medium risk, high risk), are calculated. Based on the calculated weight coefficients, a grey clustering coefficient vector C is constructed, and combined with the preset scores for each level (e.g., 1 point for low risk, 3 points for medium risk, and 5 points for high risk), the comprehensive risk score R(t) of the target object at the current time t is calculated. This step uses a mathematical model to transform multiple discrete or fuzzy qualitative physiological signs into a continuous, quantitative comprehensive risk score sequence. This helps to solve the static and coarse-grained problems of traditional scale assessments, enabling the quantification of even small changes in risk.

[0034] After obtaining the comprehensive risk score sequence, a Long Short-Term Memory (LSTM) network is further applied for time-series early warning analysis. LSTM is a special type of recurrent neural network whose internal gating mechanism enables it to effectively learn and remember long-term dependencies in time-series data, which is crucial for capturing the development trend of chronic diseases. Specifically, the comprehensive risk score sequence within a historical time window k, such as {R(tk), ..., R(t)}, is used as the input to the LSTM network. By training on a large amount of historical data, the network model can learn the evolution pattern of the risk score, thereby predicting the risk score at the next moment, defined here as R_hat(t+1). After obtaining the actual observed value R(t+1) at the future moment, the residual between the predicted value and the observed value can be calculated, defined as ε(t) = |R(t+1) - R_hat(t+1)|. When the residual ε(t) exceeds a preset statistical threshold, it indicates that the risk status has experienced unexpected and drastic fluctuations. Based on the magnitude of this residual or based on the slope of the predicted risk trend, the early warning level can be divided into multiple levels. For example, it can be divided into Level 0 warning (corresponding to fluctuations in factors, indicating the need to pay attention to daily behavior), Level 1 warning (corresponding to abnormal risk syndromes, indicating the need for community-level screening), Level 2 warning (corresponding to a significant increase in risk, indicating the need to initiate professional intervention), and Level 3 warning (corresponding to the occurrence or progression of disability, indicating the need to be transferred to the care process), thereby achieving dynamic and forward-looking monitoring of disability risk.

[0035] The data processing flow of this grey clustering-LSTM coupled model can be divided into four main layers. First, the data input layer receives multidimensional physiological indicators, including gait speed X1, grip strength X2, and raw data up to MMSE score Xn. Then, it enters the grey clustering processing layer, using a grey whitening weight function to calculate the membership degree of each indicator to low, medium, and high risk levels, thereby constructing a grey clustering coefficient vector C=[c1,c2,c3], and combining it with preset scores (1 point for low risk, 3 points for medium risk, and 5 points for high risk) to calculate the comprehensive risk score R(t)=c1×1+c2×3+c3×5. The third layer is the LSTM time-series prediction layer, which inputs the risk score sequence {R(tk),R(t-k+1),...,R(t)} within the historical time window k into the LSTM network. This network models long-term temporal dependencies through three gating units: a forget gate, an input gate, and an output gate, and outputs the predicted risk score R_hat(t+1) for the next time step. Finally, at the early warning output layer, the residual between the actual and predicted values, e(t) = |R(t+1) - R_hat(t+1)|, is calculated. Based on the magnitude of the residual, the early warning is divided into four levels: level zero, level one, level two, and level three, thus realizing a multi-level risk early warning function based on a coupled model. The organic combination of grey clustering and LSTM deep learning technology demonstrates the technical advantages of this invention in risk assessment and prediction.

[0036] Step S200 involves constructing a structured knowledge base for disability prevention and control to provide evidence-based support for subsequent decision-making. This step includes extracting and organizing intervention measures from clinical guidelines, academic literature, and expert experience to form an intervention measure library. Preferably, the intervention measure library can cover various intervention types, such as exercise interventions (e.g., balance training, resistance training, aerobic exercise), nutritional interventions (e.g., protein supplementation, Mediterranean diet adjustment), cognitive interventions (e.g., memory training, executive function training), and appropriate traditional Chinese medicine techniques (e.g., acupuncture, massage, Baduanjin), to ensure the diversity of intervention programs.

[0037] To enable these interventions to be understood and invoked by the algorithm, each intervention in the library needs to be labeled and quantified with multi-dimensional attributes. Specifically, each intervention can be labeled with its applicable scenario attributes, such as home scenario (usually requiring simple and easy-to-operate equipment), community scenario (may require professional guidance or specific venues), or hospital scenario (relying on specialized medical equipment and environment). Furthermore, each intervention is assigned a series of calculable quantitative indicators, such as the expected effectiveness index I_eff, which quantifies the expected effect of the intervention on improving a specific disability characteristic F based on meta-analysis results or expert scoring, for example, a score ranging from 1 to 10; the burden index I_bur, which quantifies the physiological or psychological burden that users may incur when implementing the intervention, comprehensively considering factors such as physical exertion, operational difficulty, and time cost; and the resource requirement level, used to characterize the external resource support required to implement the intervention. Through the above structured processing, unstructured intervention knowledge is transformed into an entity-relationship knowledge graph that supports machine reasoning, where each intervention is an entity with clearly defined attributes and relationships.

[0038] Step S300: Based on the risk assessment results and the disability prevention knowledge base, an initial decision path is generated for the target object using a multi-objective optimization algorithm. In a specific execution step, firstly, based on the risk status output by the risk grading and early warning model in step S100 (e.g., "Level 2 Early Warning - Moderate Risk - Impaired Walking Function") and the user's scenario (e.g., "at home"), a set of candidate intervention measures is selected from the knowledge base. This selection process excludes intervention measures that do not meet the current risk level, disability characteristics, or scenario conditions.

[0039] Subsequently, a multi-objective optimization function J is established to evaluate the merits of decision-making paths. This function comprehensively evaluates the overall value of a path P composed of multiple intervention measures arranged in a time sequence. The optimization function J can be expressed as: J = w1 F_effect(P) + w2 F_exec(P) - w3 F_burden(P). In this formula, each term is defined as follows: F_effect(P) represents the expected total effectiveness of path P in improving the target disability characteristics, which can be calculated by summing the expected effectiveness indices I_eff of all interventions in the path; F_exec(P) represents the comprehensive executability score of path P in the current scenario, which can be obtained by weighted average of the executability scores of each measure in the path in a specific scenario; F_burden(P) represents the total burden required for the user to execute the entire path P, which can be calculated by summing the burden indices I_bur of all interventions in the path. w1, w2, and w3 are the weight coefficients corresponding to effectiveness, executability, and burden, respectively, and w1+w2+w3=1. Their initial values ​​can be set according to general experience or user preferences. For example, for conservative users, they can be set to (0.3, 0.5, 0.2) to prioritize executability.

[0040] Finally, heuristic search algorithms, such as genetic algorithms or particle swarm optimization, can be used to find the sequence of intervention measures that maximizes the objective function J under the current constraints. This sequence is the initial decision path P_opt. For example, genetic algorithms can simulate the natural selection process, treating different combinations of intervention measures as chromosomes. Through operations such as crossover and mutation, they iteratively generate a population of better decision paths, eventually converging to a solution that satisfies preset conditions in terms of effectiveness, feasibility, and burden.

[0041] Specifically, the solution process for the multi-objective optimization function includes three main stages: input preparation, candidate measure screening, and genetic algorithm solution. In the input preparation stage, the system receives three types of input: risk status input (including risk level, disability degree, and functional characteristics), user scenario input (such as home, hospital, or community), and a set of candidate intervention measures stored in the knowledge base. The candidate measure screening module pre-screens the measure set based on risk level, disability characteristics, and scenario conditions to obtain a subset of applicable candidate measures.

[0042] The core module of the genetic algorithm is responsible for finding the optimal path. Its process includes: first, generating an initial population by randomly generating multiple candidate paths as initial solutions; then, proceeding with an iterative optimization process, including four stages: fitness evaluation, selection, crossover, and mutation. Fitness evaluation is based on the objective function J = w1. F_effect + w2 F_exec -w3 F_burden calculates the fitness value of each path, taking into account intervention effectiveness, feasibility, and user burden. The selection operation retains paths with high fitness, the crossover operation exchanges intervention measures between paths, and the mutation operation randomly replaces some measures to increase population diversity. After each iteration, it checks if the termination condition (maximum number of iterations or convergence criterion) is met. If not, it returns to the fitness evaluation stage to continue iterating; if the condition is met, it outputs the optimal decision path P_opt.

[0043] Through the aforementioned multi-objective optimization algorithm, the system can automatically find the optimal intervention path that balances the intervention effect, execution feasibility, and user burden from a large set of candidate measures under complex constraints by utilizing the iterative search mechanism of the genetic algorithm. This demonstrates the intelligent and adaptive characteristics of the decision-making method of this invention.

[0044] Step S400: After the initial decision path is generated and executed, a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback is established. For example... Figure 3 As shown, the core of this mechanism lies in adjusting the decision-making logic based on actual user feedback.

[0045] Specifically, the method involves periodically (e.g., weekly) collecting feedback data during the intervention process and reconstructing the decision path based on the feedback data. In a specific implementation step, the feedback bias is first calculated. This feedback bias can include effect bias and execution bias. For example, effect bias e1 can be determined by the difference between the target improvement rate in activities of daily living (ADL) and the actual monitored ADL improvement rate. The target ADL improvement rate can be set based on clinical guideline data in a knowledge base or preset by the physician according to the patient's condition. Simultaneously, execution bias e2 can be determined by the difference between the preset target intervention execution rate (e.g., planned for 10 sessions, actually completed 8 sessions, execution rate 80%) and the actual recorded execution rate. The preset target intervention execution rate is typically set to 100% or as required by the rehabilitation plan.

[0046] After obtaining the aforementioned feedback bias, the effect bias e1 and execution bias e2 are used as inputs to the fuzzy weight controller. Fuzzy logic can be used to process and quantify fuzzy human language concepts such as "larger" and "smaller." The fuzzy weight controller has a pre-set fuzzy rule base that can simulate the decision-making logic of domain experts. Its output is the adjustment amount of each weight coefficient in the aforementioned multi-objective optimization function J, namely Δw1, Δw2, and Δw3. For example, the fuzzy rule base may contain the following rule: "If the effect bias e1 is a large positive value (indicating that the actual improvement is far lower than expected) and the execution bias e2 is a small value (indicating that the user's execution is good), then the fuzzy weight controller infers and outputs a large positive value Δw1 and a large negative value Δw3." The technical principle of this rule is: when the user's cooperation is high but the effect is poor, stronger and more targeted intervention measures should be provided. Therefore, it is necessary to increase the effectiveness weight w1 and reduce the consideration of burden, that is, reduce the burden weight w3. Conversely, if the execution deviation e2 is large, it indicates that the current path may be too difficult or inconvenient for the user to execute. In this case, the fuzzy weight controller will infer and output an adjustment amount mainly to increase the executability weight w2. By introducing this fuzzy weight controller, parameter-level adaptive optimization of the decision path is achieved, enabling the system to weigh the effectiveness, executability, and burden of intervention based on the dynamic feedback of individuals.

[0047] Furthermore, to enhance the optimization capability and generalization of the fuzzy weight controller, this invention introduces an Adaptive Neural Fuzzy Inference System (ANFIS) for online optimization. ANFIS combines the learning capabilities of neural networks with the inference capabilities of fuzzy systems. The system continuously aggregates feedback data (e1, e2) from a group of users, along with the long-term intervention effects (such as the total improvement in ADL after three months) resulting from corresponding weight adjustment strategies (Δw1, Δw2, Δw3). Utilizing this data, the Adaptive Neural Fuzzy Inference System can act as a learning mechanism, employing algorithms such as gradient descent to inversely adjust and optimize the membership functions and fuzzy rules within the fuzzy weight controller. In this way, the decision logic of the fuzzy controller no longer relies solely on pre-set static expert knowledge but can learn from large-scale empirical data and evolve a better weight adjustment model, thereby achieving adaptive updates to the adjustment logic itself.

[0048] like Figure 6The figure shows the training error convergence curve of the Adaptive Neural Fuzzy Inference System (ANFIS) during the optimization process. The horizontal axis represents the number of iterations (Epoch), and the vertical axis represents the mean squared error (MSE). As can be seen from the figure, the model, starting from the initial membership function parameters, is in a rapid learning phase during the first 100 iterations, with the MSE rapidly decreasing from around 0.85 to around 0.25, indicating that the system can quickly correct model parameters using feedback data. As learning progresses, the curve gradually flattens out, reaching a stable parameter state after approximately 300 iterations, with the final MSE stabilizing at 0.080, below the preset convergence threshold (MSE=0.1). This verifies that the ANFIS mechanism used in this invention has good convergence and stability, and can effectively learn and optimize decision logic from population data.

[0049] Step S500 provides a self-evolutionary knowledge base method based on empirical data from a group. This method utilizes real-world data from all users to periodically iteratively update the structured disability prevention knowledge base, correcting biases in prior knowledge and discovering new knowledge.

[0050] In a specific implementation step, the collected user execution records are first cleaned and stratified to select high-confidence samples. For example, only the data of the intervention period with an execution rate higher than a certain threshold (such as 85%) are retained. Based on the intervention effect, they are divided into positive samples (significant improvement in ADL) and negative samples (stagnation or deterioration in ADL improvement).

[0051] This invention achieves dual self-evolution of the knowledge base by constructing a population-based learning system. Firstly, it updates intervention parameters using a Bayesian posterior correction method. For the prior parameters of each intervention in the knowledge base, such as the expected effectiveness index I_eff and the burden index I_bur, they are corrected using the observed means from the population empirical data. Specifically, the Bayesian smoothing formula can be used: I_new = λ I_prior + (1-λ) I_observed. In this formula, I_prior is an initial value set based on guidelines or expert experience, I_observed is the measured mean from the population, and λ is a confidence factor whose value can be positively correlated with the sample size used to calculate the mean. This method addresses the potential discrepancy between theoretical parameters in evidence-based medicine guidelines and real-world application effects, making the knowledge base parameters more realistic.

[0052] Secondly, association rule mining algorithms are used to discover implicit taboo rules. By performing association analysis on features in the negative sample set, such as using the Apriori algorithm, strong association combinations leading to adverse intervention outcomes can be identified. For example, the combination "when the user's characteristic is stage II hypertension and the intervention scenario is in the morning, perform specific high-intensity resistance training B" may be found to be strongly associated with "sudden drop in compliance" or "occurrence of adverse events." The system can automatically generate new taboo rules from such findings and store them in the knowledge base. When planning paths for users with similar characteristics in the future, the system will proactively avoid such high-risk combinations, thereby improving the safety and effectiveness of decision-making.

[0053] like Figure 5 As shown, the knowledge base's dual self-evolution mechanism starts from the execution records of group users. First, it filters high-confidence samples with an execution rate greater than 85% through data cleaning and hierarchical modules, classifying these samples into positive and negative categories. For positive samples showing significant ADL improvement, the system initiates the first evolutionary path—Bayesian posterior correction—using the formula I_new = λ. I_prior + (1-λ) I_observed updates the parameters of the expected effectiveness index I_eff and the burden index I_bur, where λ is the confidence factor. For negative samples where ADL improvement has stagnated or worsened, the system initiates a second evolutionary path—association rule mining. The Apriori algorithm is used to mine strong association combinations leading to adverse events, generating taboo rules such as "Hypertension Grade II + Morning + High-Intensity Training B → Adverse Event". The parameter corrections and taboo rules generated by the two evolutionary paths are ultimately incorporated into the knowledge base update module and jointly written into the disability prevention knowledge base, achieving continuous optimization and improvement of the knowledge base. The diagram uses solid lines to indicate the processing flow of positive samples and dashed lines to indicate the processing flow of negative samples, clearly demonstrating the parallel mechanism of dual self-evolution.

[0054] Example 2 This embodiment provides an adaptive decision-making path planning system based on multi-level risk, such as... Figure 1 As shown, the system can implement the method described in Example 1. Figure 4 As shown, the system may include hardware such as a processor, memory, communication interface, and I / O interface. The communication interface connects the wearable device 401 and the mobile terminal 402 via a network, and the I / O interface connects to the display device 403. The memory stores a computer program, which, when executed by the processor, can implement the following functional modules: The multi-source data acquisition module 104 is used to collect multi-dimensional health data, behavioral data, and feedback data during intervention processes from users through wearable devices, mobile applications, or third-party information system interfaces. For example... Figure 7 As shown, the typical application scenarios of this system cover a variety of IoT devices. The edge gateway connects to the cloud server via WiFi or 4G network, and connects to devices such as smart bracelet 101, smartwatch 102, heart rate belt 103, smartphone, tablet, scale 301, blood pressure monitor 302 and infrared sensor 303 via Bluetooth or WiFi, so as to realize the comprehensive real-time collection and transmission of user physiological, behavioral and environmental data.

[0055] The risk calculation engine is configured with a gray clustering and long short-term memory network (LSTM) coupled model described in Example 1. This engine is responsible for receiving data from the multi-source data acquisition module, performing dynamic risk assessment and multi-level early warning, and outputting the assessment results to the adaptive decision engine.

[0056] The adaptive decision engine is internally configured with the multi-objective optimization algorithm described in Example 1 and a fuzzy weight controller optimized based on the Adaptive Neural Fuzzy Inference System (ANFIS). Based on the output of the risk calculation engine 201 and the content of the disability prevention knowledge base, the engine generates an initial decision path and adaptively and dynamically adjusts the decision path according to user feedback data.

[0057] The knowledge evolution engine is used to execute the Bayesian parameter correction and association rule mining algorithm described in Example 1. This engine periodically processes the collected population data, iteratively updating and optimizing the intervention parameters and rules in the disability prevention knowledge base.

[0058] And a multi-terminal interaction module 304, which provides a visual interactive interface for different roles (such as doctors, caregivers, and users) to display risk assessment results, decision-making path schemes and execution feedback, and to receive user input commands.

[0059] This invention, through the aforementioned technical solution, constructs a closed-loop management system encompassing dynamic risk assessment, personalized path generation, adaptive feedback adjustment, and collective knowledge evolution. This method and system can provide precise, dynamic, and personalized decision support for the prevention and control of disability among the elderly.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive decision path planning method based on multi-level risk, characterized in that, Includes the following steps: Step S1: Construct and apply a full-cycle risk grading and early warning model for disability, collect and extract multi-dimensional risk indicators from the health data of the target object, perform static risk grading on the multi-dimensional physiological indicators in the multi-dimensional risk indicators and predict the evolution trend of the generated risk scores, thereby dynamically assessing the risk status of the target object. Step S2: Construct a structured knowledge base for disability prevention and control, extract intervention measures and perform multi-dimensional attribute labeling and quantification to form an entity-relationship knowledge graph that supports machine reasoning; Step S3: Based on the risk status and the disability prevention and control knowledge base, an initial decision path is generated for the target object by solving a multi-objective optimization function; The multi-objective optimization function evaluates the overall value of a candidate decision path by combining the expected total effectiveness of the candidate decision path, the comprehensive executability score, the total burden required for execution, and the corresponding weight coefficients. Step S4: Establish a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback, adjust the weight coefficients and reconstruct the initial decision path. The reconstruction steps include: calculating the feedback deviation based on the collected feedback data, outputting the adjustment amount for the weight coefficients through the fuzzy weight controller, updating the weight coefficients and reconstructing the decision path accordingly. Step S5: Based on empirical data from the group of users, the disability prevention and control knowledge base is self-evolved and updated, including updating intervention parameters and discovering implicit taboo rules.

2. The method according to claim 1, characterized in that, In step S4, the feedback data collected during the intervention process is collected periodically; the feedback deviation includes effect deviation e1 and execution deviation e2, wherein the effect deviation e1 is determined by the difference between the preset target improvement rate of daily living activities and the actual monitored improvement rate of daily living activities, and the execution deviation e2 is determined by the difference between the preset target intervention execution rate and the actual recorded execution rate. There are at least three weighting coefficients, denoted as w1, w2 and w3; The fuzzy weight controller performs inference based on a preset fuzzy rule base, and outputs adjustment amounts Δw1, Δw2 and Δw3 for each weight coefficient in the multi-objective optimization function according to the magnitude and sign of the effect deviation e1 and the execution deviation e2. The update of the weight coefficients specifically involves adding the weight coefficients w1, w2, and w3 to the adjustment amounts Δw1, Δw2, and Δw3, respectively. The operation of reconstructing the decision path involves resolving the multi-objective optimization function based on the updated weight coefficients to generate the updated decision path.

3. The method according to claim 1, characterized in that, In step S1, the full-cycle disability risk grading and early warning model is a gray clustering and long short-term memory network coupled model, specifically including: A grey clustering algorithm was applied to perform static risk classification on the collected multidimensional physiological indicators, resulting in a comprehensive risk score sequence; and, The comprehensive risk score sequence is used as input to a long short-term memory network to learn and model the risk evolution trend, so as to predict the risk score at the next moment and perform early warning analysis.

4. The method according to claim 1, characterized in that, In step S5, the empirical data of the group of users is first divided into a positive sample set and a negative sample set, and the disability prevention and control knowledge base is then updated through self-evolution, including: The intervention parameters, including the expected effectiveness index and the burden index, are updated using a Bayesian posterior correction method, specifically employing the Bayesian smoothing formula I_new = λ. I_prior + (1-λ) I_observed is used for correction, where I_new is the corrected parameter value, I_prior is the prior parameter value, I_observed is the measured mean from the population, and λ is the confidence factor; and, Association rule mining algorithms are used to perform association analysis on features in the negative sample set in order to discover and generate implicit taboo rules.

5. The method according to claim 1, characterized in that, In step S2, the multi-dimensional attribute labeling and quantification of the intervention measures includes: Tag each intervention with its applicable scenario attributes; and, Each intervention is assigned a quantitative indicator, which includes an expected effectiveness index to quantify the expected effect of the intervention on improving a specific disability characteristic, and a burden index to quantify the physiological or psychological burden incurred by the target population when implementing the intervention.

6. The method according to claim 1, characterized in that, The method further includes: The fuzzy weight controller in step S4 is optimized online using an adaptive neural fuzzy inference system, specifically including: The long-term intervention effect of continuously collecting feedback data from group users and corresponding weight adjustment strategies; Using the aforementioned adaptive neural fuzzy inference system, the gradient descent algorithm is employed to inversely adjust and optimize the membership function and fuzzy rules within the fuzzy weight controller.

7. The method according to claim 1, characterized in that, In step S3, the multi-objective optimization function adopts the formula J = w1 F_effect(P) + w2 F_exec(P) - w3 F_burden(P) is calculated, where J is the overall value, P is the candidate decision path, w1, w2, and w3 are the weighting coefficients for effectiveness, executability, and burden, respectively, and w1+w2+w3=1; F_effect(P) is the expected total effectiveness, calculated by the sum of the expected effectiveness indices of all interventions in the candidate decision path P; F_exec(P) is the comprehensive executability score, obtained by the weighted average of the executability scores of each measure in the candidate decision path P; F_burden(P) is the total burden, calculated by the sum of the burden indices of all interventions in the candidate decision path P.

8. The method according to claim 3, characterized in that, The steps for applying the grey clustering algorithm to perform static risk classification on the collected multidimensional physiological indicators and obtain a comprehensive risk score sequence include: Using the gray whitening weight function, the weight coefficients of each physiological indicator belonging to multiple preset risk levels are calculated; Based on the weight coefficients, a gray clustering coefficient vector is constructed, and combined with the preset scores of each risk level, the comprehensive risk score of the target object at the current moment is calculated, thereby forming the comprehensive risk score sequence.

9. The method according to claim 5, characterized in that, The disability prevention and control knowledge base includes the following types of intervention measures: exercise intervention, nutritional intervention, cognitive intervention, and traditional Chinese medicine techniques.

10. An adaptive decision-making path planning system based on multi-level risk, characterized in that, include: The multi-source data acquisition module is configured to collect health data and intervention feedback data of the target object; The risk calculation engine is configured to build and apply a full-cycle risk grading and early warning model for disability, collect and extract multi-dimensional risk indicators from the health data of the target object, perform static risk grading of physiological indicators and predict the evolution trend of risk scores, thereby dynamically assessing the risk status of the target object. The structured disability prevention knowledge base is configured to store extracted and multi-dimensional attribute-labeled and quantified intervention measures, forming an entity-relationship knowledge graph that supports machine reasoning; An adaptive decision engine is configured to generate an initial decision path for the target object by solving a multi-objective optimization function based on the risk status and the disability prevention knowledge base. The multi-objective optimization function evaluates the overall value of a candidate decision path by combining the expected total effectiveness of the candidate decision path, the comprehensive executability score, the total burden required for execution, and the corresponding weight coefficients. And establish a path adaptive dynamic adjustment mechanism based on fuzzy logic feedback to adjust the weight coefficients and reconstruct the initial decision path, specifically including: calculating the feedback deviation based on the collected feedback data, outputting the adjustment amount for the weight coefficients through the fuzzy weight controller, updating the weight coefficients and reconstructing the decision path accordingly; A knowledge evolution engine is configured to self-evolve and update the disability prevention knowledge base based on empirical data from a group of users, including updating intervention parameters and discovering implicit taboo rules; and... The multi-terminal interaction module is configured to provide an interactive interface to display results and receive input.

Images