A deep learning-based chronic disease personalized development path planning method

Abstract

The application discloses a kind of based on deep learning's slow disease individualized development path planning method, it is related to intelligent health management technical field, including: the construction standardization health status dataset, and generate multidimensional feature set by representation learning process;With multidimensional feature set as the input of association analysis, extract the feature set representing different health goals, assign target weight to generate individual multi-objective health feature table;Individual multi-objective health feature table is input into deep reinforcement learning model, through the adaptive learning of multilevel sub-strategy, corresponding strategy weight is assigned to different health goals, and candidate intervention path is generated;In candidate intervention path training, risk and benefit feedback are combined, and dynamic balance intervention strategy is output.The application realizes the intervention contribution degree coordination and path weight adaptive convergence effect on the whole path scale to multiple health goals by using dynamic balance algorithm in candidate intervention path training process.

Description

Technical Field

[0001] This invention relates to the field of intelligent health management technology, and in particular to a method for planning personalized development paths for chronic diseases based on deep learning. Background Technology

[0002] With the development of digital health management and intelligent medical technologies, dynamic monitoring, temporal modeling, and intelligent extrapolation methods based on multi-source health data have gradually become important research directions in chronic disease management. Deep representation learning, reinforcement learning, and multi-objective decision-making models have begun to be applied to the construction of personalized health intervention programs. By modeling the complex interactions between continuous health state vectors, multi-dimensional features, and intervention behaviors, a more refined characterization of individual health evolution trends can be achieved. Meanwhile, with the improvement of wearable monitoring, continuous collection of physiological parameters, and structured processing capabilities of behavioral data, multi-time-slice and multi-dimensional health state information provides a rich data foundation for intelligent path planning.

[0003] Current chronic disease management technologies, in contexts with multiple health goals, still struggle to simultaneously address the dynamic balance between heterogeneous objectives such as glycemic stability, blood pressure control, and weight management. In particular, they lack a path generation mechanism capable of comprehensively measuring the cumulative effects of interventions, risk-benefit structures, and the weight distribution of health goals over a complete time series. When the evolutionary trends of multiple objectives are inconsistent, traditional methods struggle to dynamically weigh candidate intervention paths at a global level, thus failing to obtain individualized intervention paths with stable convergence across time slices and objective ranges. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a deep learning-based method for planning personalized development paths for chronic diseases, thereby solving the problem of constructing a dynamically balanced individualized intervention path that takes into account temporal effects, target weights, and risk-reward structures when multiple health goals coexist.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a deep learning-based method for personalized chronic disease development path planning, comprising: constructing a standardized health status dataset and generating a multidimensional feature set through a representation learning process; using the multidimensional feature set as input for association analysis, extracting feature sets representing different health goals, assigning goal weights to generate an individual multi-goal health feature table; inputting the individual multi-goal health feature table into a deep reinforcement learning model, and through adaptive learning of multi-level sub-strategies, assigning corresponding strategy weights to different health goals to generate candidate intervention paths; combining risk and benefit feedback in the training of candidate intervention paths to output a dynamically balanced intervention strategy; executing the intervention strategy, collecting and organizing observation data during the execution period to form risk and benefit feedback records, and using the deep reinforcement learning model's reinforcement learning algorithm to self-adjust and optimize learnable parameters to generate a joint intelligent model; and using the joint intelligent model for path execution and real-time feedback to continuously track and update the patient's health status. Based on the continuously tracked and updated health status of the patients, multidimensional time-series data including blood glucose, blood pressure and weight indicators are extracted. A multi-objective evolution trend model is constructed using a long short-term memory network to predict the evolution trajectory of each indicator within a preset time window. The predicted evolution trajectory is mapped to a preset time series space, and the Monte Carlo tree search algorithm is used to simulate the cumulative effect of different intervention actions in the time series space. The risk-reward structure matrix of each intervention action in the cross-time slice dimension is calculated. Combining the risk-return structure matrix with the preset health target weight distribution, a multi-objective evolutionary algorithm is used to perform a global dynamic trade-off analysis on candidate intervention paths to identify the non-dominated solution set when the evolution trends of multi-objective indicators are inconsistent. Calculate the convergence index of each path in the non-dominated solution set across the target range, select the preferred path that meets the preset convergence condition based on the convergence index, and generate an individualized intervention path.

[0006] As a preferred embodiment of the deep learning-based personalized development path planning method for chronic diseases described in this invention, the specific steps for constructing a standardized health status dataset and generating a multi-dimensional feature set through a representation learning process are as follows: The system collects patients' raw health data and performs semantic alignment, unit conversion, and anomaly handling to form a standardized health data record set. The standardized health data record set is sorted by timestamp and sliced ​​according to a preset time granularity. The health indicators within the same time slice are aggregated and calculated to form a standardized health status dataset. Based on the standardized health status dataset, the health indicators of each time slice are arranged in order and formed into a fixed-dimensional health status vector to generate a multi-dimensional feature set.

[0007] As a preferred embodiment of the deep learning-based personalized chronic disease development path planning method of the present invention, the specific steps for generating an individual multi-objective health feature table by allocating target weights are as follows: Based on a multidimensional feature set, health status vectors are classified into blood glucose stability, blood pressure control, and weight management to construct a health target set. The difference between the health status vectors and corresponding health indicators within the health target set is used as input, and a target sensitivity value is formed through nonlinear enhancement and normalization compression. Based on the target sensitivity value, weights are assigned to all health targets in the health target set, and the weights of each health target are bound to the feature components to generate an individual multi-target health feature table.

[0008] As a preferred embodiment of the deep learning-based personalized chronic disease development path planning method of the present invention, the specific steps for generating candidate intervention paths are as follows: An individual's multi-objective health feature table is input into a deep reinforcement learning model. The weights and feature components of each health objective are combined to form the state description of the deep reinforcement learning model. Based on the state description, the health objectives are ranked according to their weights to obtain their weight priorities. Using these weight priorities, a corresponding set of sub-policies is generated for each health objective. Each sub-policy generates candidate action sequences for different health objectives. Based on these candidate action sequences, the deep reinforcement learning model is used to extrapolate changes in health status and generate candidate intervention paths.

[0009] As a preferred embodiment of the deep learning-based personalized chronic disease progression path planning method described in this invention, the specific steps of outputting a dynamically balanced intervention strategy are as follows: The improvement magnitude of health goals is calculated using the health indicator components in the health status vector. Based on candidate intervention paths, the health status vector and the improvement magnitude of health goals for each path are matched to evaluate the responsiveness of each path to the health goals. Based on the responsiveness value and the weight priority of the health goals, an optimization score is assigned to each candidate intervention path. The optimization scores are then used to calculate a weighted average of all candidate intervention paths, and a dynamic balancing algorithm is executed. Based on the output of the dynamic balancing algorithm, the candidate intervention path with the optimized score and weighted average is selected, and the dynamically balanced intervention strategy is output.

[0010] As a preferred embodiment of the deep learning-based personalized development path planning method for chronic diseases described in this invention, the output of the dynamic balancing algorithm refers to the coordinated intervention contribution among multiple health goals and the generation of balanced path weights through the optimized scoring of candidate intervention paths.

[0011] As a preferred embodiment of the deep learning-based personalized development path planning method for chronic diseases described in this invention, the observation data during the execution period includes health indicator change data, health behavior and execution records, biochemical test data, physiological monitoring data, and patient subjective feedback data.

[0012] As a preferred embodiment of the deep learning-based personalized development path planning method for chronic diseases described in this invention, the learnable parameters within the deep reinforcement learning model refer to the set of parameters that are updated based on observation data during the execution period in the deep reinforcement learning process.

[0013] As a preferred embodiment of the deep learning-based personalized chronic disease development path planning method described in this invention, the specific process of generating the joint intelligent model is as follows: The execution-period observation data is paired with candidate intervention paths to construct feedback samples for the deep reinforcement learning model. Based on the feedback samples, the learnable parameters within the deep reinforcement learning model are updated using the execution-period observation data. After the learnable parameters are updated, the candidate intervention paths are evaluated to generate intervention strategy content, value assessment content, and path benefit assessment content, which are then integrated. Based on the integrated intervention strategy content, value assessment content, and path benefit assessment content, a joint intelligent model is generated.

[0014] As a preferred embodiment of the deep learning-based personalized chronic disease development path planning method described in this invention, the specific process of continuously tracking and updating the patient's health status is as follows: Based on the joint intelligent model, candidate intervention paths are executed, and execution-period observation data is generated during the execution process. The execution-period observation data is compared with the health status vector to generate real-time feedback data. Combining the real-time feedback data and the balanced path weights, candidate intervention paths are selected after path execution, and the dynamically balanced intervention strategy path execution mode is output and adjusted. The adjusted intervention strategy path execution mode and real-time feedback data are used as feedback and input into the joint intelligent model to continuously track and update the patient's health status.

[0015] The beneficial effects of this invention are as follows: By employing a dynamic balancing algorithm during the training of candidate intervention paths, the coordination of intervention contributions for multiple health goals across the entire path scale and the adaptive convergence of path weights are achieved. By constructing a comprehensive response vector across time slices using the response degree values ​​of health goals, goal weight coefficients, and path optimization scores, and generating balanced path weights based on an iterative normalization proportional adjustment mechanism, the contribution distribution of intervention paths in the three directions of blood glucose stabilization, blood pressure control, and weight management is dynamically corrected. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart for a deep learning-based method for planning personalized development paths for chronic diseases; Figure 2 A flowchart for constructing a standardized health status dataset; Figure 3 A flowchart generated for candidate intervention paths; Figure 4 A flowchart generated for the dynamic equilibrium intervention strategy. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for planning personalized development paths for chronic diseases based on deep learning, including the following steps: S1: Construct a normalized health status dataset and generate a multidimensional feature set through a representation learning process.

[0022] S1.1: Collect patients' raw health data, and perform semantic alignment, unit conversion, and anomaly handling to form a standardized health data record set.

[0023] Specifically, the process involves collecting patients' raw health data, semantically aligning the names, field meanings, and recording formats of health indicators from different sources, and unifying indicators with the same or similar meanings into preset standard labels. Based on these standard labels, a unified conversion is performed on the units of measurement for health indicators, transforming different recording units into a unified recording method. Missing values, obvious duplicate records, and inconsistencies in the original health data records, particularly within the same timeframe and for the same patient, related to the same health indicator, are identified, corrected, and removed. This ensures that health data remains consistent in semantic expression, unit representation, and completeness, forming a standardized health data record set.

[0024] It should be noted that health indicators include blood glucose, blood pressure, and weight.

[0025] S1.2: Sort the standardized health data record set by timestamp and slice it according to the preset time granularity. Perform aggregation calculation on health indicators within the same time slice to form a standardized health status dataset.

[0026] Specifically, each record in the standardized health data record set is sorted from morning to evening according to its timestamp, and the entire timeline is divided into continuous time slice intervals according to a preset time granularity; each health data record in the standardized health data record set is assigned to its respective time slice according to its timestamp.

[0027] Within the same time slice, health data records are grouped according to the type of health indicator. The average value of multiple health data records for each type of health indicator within the time slice is calculated as the representative value of the health indicator under the time slice. Health status records corresponding to each time slice are generated sequentially, and the health status records of all time slices are organized according to a unified structure to form a standardized health status dataset.

[0028] It should be noted that the preset time granularity is configured according to the chronic disease management plan and the frequency of health data collection, and is determined by selecting a fixed time unit of day, week or month in combination with the follow-up cycle and monitoring interval.

[0029] S1.3: Based on the standardized health status dataset, the health indicators of each time slice are arranged in order and formed into a fixed-dimensional health status vector to generate a multi-dimensional feature set.

[0030] Specifically, based on the standardized health status dataset, for each time slice, the values ​​of health indicators are extracted sequentially according to the order of health indicators, and the values ​​of health indicators are filled into the corresponding positions to form a health status vector of fixed dimension; health status vectors are formed for all time slices, and the health status vectors are arranged in chronological order to obtain a health status vector sequence.

[0031] Using the health status vector sequence as input to the representation learning process, weighted accumulation and nonlinear transformation iterative calculations are performed on the numerical differences and correlation between adjacent time slice vectors. The weight coefficients used in the weighted accumulation are real numbers greater than 0 and less than or equal to 1, and are normalized to a sum of 1 within the same combination. The setting is based on a comprehensive assessment of the importance of each health indicator component in the health target set, its sensitivity to changes in health status, and its stability in historical health data. Multidimensional representations reflecting the temporal changes in health status are extracted, and the multidimensional representations corresponding to each time slice are organized according to a unified structure to generate a multidimensional feature set.

[0032] It should be noted that weighted accumulation refers to multiplying the numerical differences of each health indicator component by a preset weight coefficient between health status vectors in adjacent time slices and summing them continuously over time. The weight coefficient is a real number greater than 0 and less than or equal to 1, and the sum of the weight coefficients of all health indicator components is constrained to be 1 within the same time slice. The specific value is determined based on the importance of each health indicator component in the health target set, the magnitude of numerical fluctuation in the historical health status vector sequence, and the correlation with the magnitude of improvement of health targets.

[0033] The nonlinear transformation iterative calculation combines the weighted cumulative output obtained between the health state vectors of adjacent time slices with the intermediate feature representation obtained in the previous round. After compression, amplification and interactive operations are completed in the input nonlinear mapping process, the current feature representation is updated, and the update process is repeated in multiple rounds in time order.

[0034] S2: Using a multidimensional feature set as input for association analysis, extract feature sets representing different health goals, assign goal weights, and generate an individual multi-goal health feature table.

[0035] S2.1: Based on the multidimensional feature set, the health status vector is classified into blood glucose stability, blood pressure control and weight management to construct a set of health goals.

[0036] Specifically, based on the correspondence between each feature component in the health status vector and blood glucose, blood pressure, and weight indicators, feature index positions are used for blood glucose stability, blood pressure control, and weight management. For each health status vector in the time series, blood glucose-related components, blood pressure-related components, and weight-related components are extracted according to the feature index. The extracted blood glucose-related components, blood pressure-related components, and weight-related components are assigned to the blood glucose stability subset, blood pressure control subset, and weight management subset, respectively. The blood glucose stability subset, blood pressure control subset, and weight management subset are labeled and organized according to the health goal name to construct a health goal set containing the three directions of blood glucose stability, blood pressure control, and weight management.

[0037] S2.2: Taking the difference between the health state vector and the corresponding health indicator within the health target set as input, the target sensitivity value is formed through nonlinear enhancement and normalization compression.

[0038] Specifically, taking each health target in the set of health targets as a unit, the difference between the health state vector and the corresponding health target in the set of health targets is extracted one by one to form a sequence of difference quantities of health targets; a nonlinear transformation operation is performed on each difference quantity in the sequence of difference quantities to change the value of the difference quantity itself according to its magnitude, resulting in a sequence of difference quantities after nonlinear enhancement processing.

[0039] For example, when performing nonlinear transformation operations on the sequence of differences in health targets, each difference corresponding to the health state vector within the set of health targets can be input into a nonlinear mapping structure. Smaller differences are compressed to make them close to zero in the output, while larger differences are amplified to occupy a higher numerical range in the output. This allows the influence of larger differences on the health target to be highlighted in subsequent calculations, while the influence of smaller differences is weakened.

[0040] Within each health target, the differential quantity sequence after nonlinear enhancement is normalized to compress the value of the differential quantity itself into a uniform range, and the normalized differential quantity sequence is used as the target sensitivity value corresponding to that health target.

[0041] Normalization is expressed by the following formula: ; Before performing normalization, each variance is divided by the dimensional normalization factor corresponding to each health indicator, using the preset dimensional normalization factor as a benchmark. This process standardizes the units of the variance, ensuring that variances of health indicators with different units of measurement are uniformly represented. Before calculation, it is mapped to a uniform dimensionless scale.

[0042] In the formula, Indicates by the first The target sensitivity value is obtained by normalizing the individual differences. The first term after nonlinear enhancement processing One difference quantity, This represents the index of the difference in the difference sequence. This represents the smallest variance among all variance sequences after nonlinear enhancement processing, within the same health target range. This represents the minimum value. This represents the largest variance among all variance sequences after nonlinear enhancement processing within the same health target range. This represents the maximum value.

[0043] S2.3: Based on the target sensitivity value, assign weights to all health targets in the health target set, bind the weights of each health target to the corresponding feature components, and generate an individual multi-target health feature table.

[0044] Specifically, the target sensitivity value corresponding to each health target in the health target set is compared within the health target set. Based on the magnitude of the target sensitivity value, it is converted into a weight coefficient of the health target according to the proportion. The weight coefficient is then bound one-to-one with the feature components belonging to the health target in the multi-dimensional feature set. The recording format of each feature component is expanded to the feature component value plus the corresponding health target weight.

[0045] For example, for the blood glucose stability health goal in the health goal set, if a certain feature component belonging to the blood glucose stability health goal in the multidimensional feature set has a value of 0.62 at a certain time slice, and the corresponding health goal weight coefficient is 0.30, then the recording format of the feature component is expanded from simply recording the feature component value: 0.62 to the feature component value: 0.62, and the blood glucose stability health goal weight: 0.30. For the blood pressure control health goal and the weight management health goal, the same method is used to combine and record their respective feature component values ​​with the corresponding health goal weight coefficients, so as to achieve a one-to-one correspondence between the feature component values ​​and the health goal weights.

[0046] Following the three health goals of stable blood glucose, blood pressure control, and weight management, all the bound feature components are organized and arranged to generate an individual multi-objective health feature table to describe the patient's health status.

[0047] S3: Input the individual's multi-objective health feature table into the deep reinforcement learning model, and establish a corresponding set of sub-policies inside to form candidate intervention paths.

[0048] S3.1: Input the individual multi-objective health feature table into the deep reinforcement learning model, and combine the weights and feature components of each health objective as the state description of the deep reinforcement learning model.

[0049] Specifically, according to a fixed order in the set of health goals, the feature component values ​​and weights corresponding to the health goals of stable blood glucose, blood pressure control, and weight management are read sequentially. For each health goal, all feature component values ​​under the health goal are extracted, and the feature component values ​​are concatenated with the weights of the health goals to form a state sub-vector. The state sub-vectors corresponding to all health goals are then concatenated to form a complete state vector containing all health goal information. This complete state vector is used as the state description of the deep reinforcement learning model.

[0050] S3.2: Based on the state description of the deep reinforcement learning model, sort the health objectives according to their weights to obtain the weight priority of the health objectives.

[0051] Specifically, in the state description of the deep reinforcement learning model, the weights of the health goals of stable blood glucose, blood pressure control, and weight management are read. The three health goal weights are combined into a comparable weight list and sorted from largest to smallest according to the value of the health goal weight coefficient. According to the sorting order, the health goal with the largest weight is marked as the first priority, the health goal with the middle weight is marked as the second priority, and the health goal with the smallest weight is marked as the third priority. This sorting and marking is recorded as the health goal weight priority in the current state.

[0052] S3.3: Utilize the weights and priorities of health goals to generate a set of corresponding sub-policies for each health goal, and generate candidate action sequences for each sub-policy for different health goals.

[0053] Specifically, each health goal is selected one by one according to its weight priority in the set of health goals. Based on the values ​​and changes of the relevant characteristic components of each health goal, the current health status and expected improvement direction of the health goal are evaluated. The differences, trends and improvement space between the current health indicators and the target values ​​are compared. Based on the evaluation results, the optimization direction of each health goal is constructed. Based on the optimization direction of each health goal, a candidate action sequence consisting of intervention actions such as medication adjustment, dietary adjustment, exercise arrangement and follow-up frequency is generated.

[0054] S3.4: Based on candidate action sequences, use a deep reinforcement learning model to extrapolate changes in health status and generate candidate intervention paths.

[0055] Specifically, based on the candidate action sequence, the health state vector corresponding to the current time slice is selected as the initial health state. The intervention actions in the candidate action sequence are input into the deep reinforcement learning model one by one in chronological order. At each time an intervention action is input, the deep reinforcement learning model outputs the predicted health state vector value for the next time slice and updates the predicted health state vector value to the input health state corresponding to the next intervention action. The same candidate action sequence is repeated until the end of the sequence. The initial health state vector, the intervention actions for each time slice, and the corresponding predicted health state vector values ​​are combined in sequence to generate the candidate intervention path corresponding to the candidate action sequence.

[0056] A superior approach is to employ a deep reinforcement learning model, as existing chronic disease interventions often use fixed procedures or simple prediction models. By using a deep reinforcement learning model to perform stepwise state transition deductions on candidate action sequences, the "current health state vector - intervention action - predicted health state vector value for the next time slice" is continuously linked to form a directly comparable candidate intervention path. This not only explicitly depicts the temporal cumulative impact of the combination of intervention actions on health status, but also embeds multiple health goal weight information, allowing for the simultaneous measurement of benefits and risks at the path level. This enables more precise screening of individualized intervention plans that achieve dynamic balance among multiple goals such as blood glucose stability, blood pressure control, and weight management.

[0057] S4: In the training of candidate intervention paths, combine risk and benefit feedback to output a dynamically balanced intervention strategy.

[0058] S4.1: Calculate the improvement magnitude of health goals using the health indicator components in the health status vector; based on candidate intervention paths, match the health status vector and the improvement magnitude of health goals for each path, and evaluate the responsiveness of each path to the health goals.

[0059] Specifically, for each candidate intervention path, the health status vector corresponding to the start and end time slices is selected in the candidate intervention path. The health indicator components in the health status vector corresponding to the health goals of blood glucose stability, blood pressure control, and weight management are compared one by one. The direction and magnitude of change of the health indicator component values ​​in the end time slice relative to the start time slice are calculated. The changes of each health indicator component under the same health goal are merged according to the weights within the health goal to obtain the improvement magnitude of the health goal.

[0060] It should be noted that for each health indicator component corresponding to a health goal, the value of the health indicator component in the health state vector at the end of the time slice is calculated by subtracting the value of the health indicator component in the health state vector at the beginning of the time slice. When the difference is greater than zero, it is recorded as an increase in the value of the health indicator component; when the difference is less than zero, it is recorded as a decrease in the value of the health indicator component. The absolute value of the difference is taken as the change range of the health indicator component.

[0061] Within the same candidate intervention path, the magnitude of improvement in health goals is correlated with the health status vector corresponding to each time slice on the path. The performance of the candidate intervention path in terms of blood glucose stability, blood pressure control, and weight management is scored according to the magnitude and direction of the improvement in health goals, generating a response value that characterizes the degree of influence of the candidate intervention path on each health goal.

[0062] S4.2: Assign an optimization score to each candidate intervention path based on the response level value and the weight priority of health goals; calculate a weighted average of all candidate intervention paths based on the optimization scores and execute a dynamic balancing algorithm.

[0063] Specifically, for each candidate intervention path, the response values ​​of the candidate intervention path to the health goals of stable blood glucose, blood pressure control, and weight management are read. The response values ​​of each health goal are paired with the corresponding weight coefficients of the health goals to form a one-to-one numerical pair. The optimized score of the candidate intervention path is calculated according to the weighted summation formula.

[0064] The optimization scores of all candidate intervention paths are normalized to obtain the weight ratios for path combination. The response values ​​of the candidate intervention paths on each health goal are then weighted and averaged according to the weight ratios to form a comprehensive response vector. The comprehensive response vector is then input into the dynamic balancing algorithm.

[0065] S4.3: Based on the output of the dynamic equilibrium algorithm, select candidate intervention paths after optimizing the scores and weighted averaging, and output the dynamic equilibrium intervention strategy.

[0066] Specifically, based on the output of the dynamic balancing algorithm, the candidate intervention path with the highest optimization score and whose contribution distribution among the health goals of stable blood glucose, blood pressure control, and weight management meets the balance requirements is selected from all candidate intervention paths. The corresponding path weights are used as the screening criteria to retain one or more candidate intervention paths. The retained candidate intervention paths are weighted and synthesized according to their path weights, and the intervention actions in each time slice are reorganized to form an intervention action sequence that can take into account the improvement needs of multiple health goals as a whole. The intervention action sequence is used as the dynamic balancing intervention strategy.

[0067] It should be noted that the balance requirement refers to the fact that, within the same candidate intervention path, the response values ​​corresponding to the health goals of stable blood glucose, blood pressure control, and weight management should not be significantly lower than their respective health goal weights.

[0068] S4.4: The output of the dynamic balancing algorithm refers to the coordinated intervention contribution among multiple health goals and the generation of balanced path weights through the optimized scores of candidate intervention paths.

[0069] Specifically, in the dynamic balancing algorithm, the sum of the optimization scores of all candidate intervention paths on the health goals of stable blood glucose, blood pressure control, and weight management is calculated to obtain the overall intervention contribution value corresponding to each health goal. The overall intervention contribution value is then compared with the corresponding health goal weights. Based on the difference in the comparison values, the adjustment coefficient of the path weights is calculated. The path weights of each candidate intervention path are increased or decreased proportionally. After each round of adjustment, all path weights are normalized to ensure that the sum of the path weights remains at 1. By repeatedly calculating the overall intervention contribution value and adjusting the path weights and performing normalization, the overall intervention contribution value obtained by weighting the path weights is matched with the health goal weights in terms of numerical distribution, thus generating the balanced path weights.

[0070] It should be noted that the dynamic balancing algorithm adopts a normalized proportional iteration method based on minimizing the loss function. In each iteration, the overall intervention contribution value vector is calculated based on the current path weight, and a loss function in the form of the sum of squared differences is constructed with the health target weight vector. The path weights of the candidate intervention paths are then proportionally adjusted and normalized according to the loss function, so that the loss function gradually decreases until it converges, and the balanced path weights are obtained.

[0071] S5: Implement intervention strategies, collect and organize observation data during the implementation period, form risk and benefit feedback records, and use deep reinforcement learning models and reinforcement learning algorithms to self-adjust, optimize learnable parameters, and generate joint intelligent models.

[0072] S5.1: Observational data during the implementation period includes data on changes in health indicators, health behavior and implementation records, biochemical test data, physiological monitoring data, and patient subjective feedback data.

[0073] Specifically, during the implementation of the intervention strategy, the changes in the patient's blood glucose, blood pressure, and weight health indicators are recorded in chronological order to form health indicator change data; the patient's medication time and dosage, diet and meal time, and exercise type and duration are recorded simultaneously to form health behavior and implementation records.

[0074] Biochemical test results and continuous monitoring records of heart rate, blood oxygen, and blood pressure obtained during the follow-up period were compiled into biochemical test data and physiological monitoring data. At the same time, patient scores on physical condition, sleep status, and emotional feelings were collected to form patient subjective feedback data. Health indicator change data, health behavior and execution records, biochemical test data, physiological monitoring data, and patient subjective feedback data were summarized according to a unified time stamp and field structure to form execution period observation data.

[0075] S5.2: The learnable parameters within a deep reinforcement learning model refer to the set of parameters that are updated based on observation data during the execution phase of deep reinforcement learning.

[0076] Specifically, the learnable parameters within a deep reinforcement learning model refer to the set of parameters used to establish correspondences between health state vectors, intervention actions and intervention strategies, value assessment content, and path benefit assessment content within the deep reinforcement learning model structure.

[0077] The parameter set includes at least the parameters for encoding and transforming the health state vector to form an internal state representation, the parameters for establishing a mapping relationship between the internal state representation and intervention actions to output the intervention strategy content, and the parameters for calculating the value assessment content and path benefit assessment content based on the internal state representation and intervention actions.

[0078] The establishment process involves taking the health state vectors and intervention actions of different time slices as inputs during deep reinforcement learning, and using the risk and benefit feedback records obtained from the observation data during the execution period as a reference. The learnable parameters inside the deep reinforcement learning model are iteratively updated in multiple rounds, so that the deep reinforcement learning model can output the corresponding intervention strategy content, value assessment content, and path benefit assessment content when given the health state vector and intervention action.

[0079] During the deep reinforcement learning process, the health indicator change data, health behavior and execution records, biochemical test data, physiological monitoring data and patient subjective feedback data contained in the execution period observation data are used as references. The execution period observation data is compared with the intervention strategy content, value assessment content and path benefit assessment content generated by the deep reinforcement learning model based on the current learnable parameters, and the learnable parameters are updated accordingly.

[0080] It should be noted that, for each feedback sample, the reference intervention strategy content, reference value assessment content, and reference path benefit assessment content, which are compiled based on the observation data during the execution period, are aligned item by item with the intervention strategy content, value assessment content, and path benefit assessment content output by the deep reinforcement learning model on the time slice and health target set, respectively. The difference is calculated for the numerical indicators at the corresponding positions, and the consistency of the descriptive content at the corresponding positions is judged. The difference is then used to update the bias information of the learnable parameters.

[0081] S5.3: Pair the observation data during the execution period with the candidate intervention paths to construct feedback samples for the deep reinforcement learning model.

[0082] Specifically, for each executed candidate intervention path, the intervention actions corresponding to each time slice in the candidate intervention path are aligned one by one with the health indicator change data, health behavior and execution records, biochemical test data, physiological monitoring data and patient subjective feedback data in the observation data during the execution period, based on the timestamp. The health status vector before execution, the actual executed intervention action, the health status vector after execution and the corresponding risk and benefit feedback records are extracted respectively, and the feedback samples are combined according to the structure of health status vector, intervention action, health status vector and risk and benefit feedback records.

[0083] S5.4: Based on the feedback samples, the learnable parameters inside the deep reinforcement learning model are updated using the observation data during the execution period.

[0084] Specifically, the health status vector and intervention actions from the feedback samples are input into the deep reinforcement learning model. The model then uses the currently learnable parameters to generate corresponding intervention strategy content, value assessment content, and path benefit assessment content. These generated content are compared with the risk and benefit feedback records compiled from observation data during the execution period. Through comparative analysis, the deviation between the intervention strategy and the actual feedback is calculated, yielding deviation information. Finally, based on this deviation information, the learnable parameters within the deep reinforcement learning model are updated using the backpropagation algorithm.

[0085] The backpropagation algorithm calculates the gradient of the loss with respect to each network parameter, and uses these gradients to update the weights and biases, thereby reducing prediction errors and optimizing the deep reinforcement learning model. S5.5: After the learnable parameters are updated, the candidate intervention paths are evaluated to generate intervention strategy content, value assessment content, and path benefit assessment content, and then integrated.

[0086] Specifically, based on learnable parameter updates, the health state vector sequence and intervention action sequence in each candidate intervention path are used as inputs to perform forward computation in the deep reinforcement learning model to obtain the state evaluation value of the candidate intervention path at each time slice. The state evaluation value is calculated as the state value function output of the deep reinforcement learning model, and the real-time benefit value of each stage is accumulated according to the time slice to form the cumulative benefit value along the path.

[0087] Based on the status evaluation quantities, they are organized and summarized in chronological order to form value assessment content corresponding to candidate intervention paths; based on the actual intervention actions and cumulative benefits adopted in each time segment, intervention strategy content and path benefit assessment content corresponding to candidate intervention paths are organized; and the intervention strategy content, value assessment content, and path benefit assessment content are uniformly linked and organized according to candidate intervention paths.

[0088] A better approach is to address the current situation where chronic disease management relies heavily on single follow-up indicators. After updating the learnable parameters, a deeper reinforcement learning model can be used to perform forward computation on each candidate intervention path, taking a complete sequence of health status vectors and intervention action sequences as inputs. This model explicitly obtains the status evaluation values ​​for each time slice and the cumulative benefits along the path, thereby generating value assessment content, intervention strategy content, and path benefit assessment content. These are then uniformly linked and organized according to the candidate intervention paths, elevating path evaluation from a single-point, single-indicator approach to a comprehensive evaluation that is path-wide, cross-target, and comparable. This facilitates fine-tuning among multiple candidate intervention paths, thereby improving the accuracy and interpretability of individualized intervention path planning for chronic diseases.

[0089] S5.6: Generate a joint intelligent model based on the integrated intervention strategy content, value assessment content, and path benefit assessment content.

[0090] Specifically, based on the integrated intervention strategy content, value assessment content, and path benefit assessment content, for each candidate intervention path, the health status vector sequence, intervention action sequence, and corresponding intervention strategy content, value assessment content, and path benefit assessment content on the candidate intervention path are collected and organized according to a unified structure. On the basis of collection and organization, a one-to-one correspondence index is established between each candidate intervention path and the set of health goals and the weights of the health goals, forming a set of candidate intervention paths that can be searched by the set of health goals and the weights of the health goals. The set of candidate intervention paths and the associated intervention strategy content, value assessment content, and path benefit assessment content are used to generate a joint intelligent model.

[0091] It should be noted that the training of the joint intelligent model is carried out by iteratively updating the learnable parameters inside the deep reinforcement learning model based on feedback samples after acquiring new execution period observation data in each round. After the parameter update is completed, the intervention strategy content, value assessment content, and path benefit assessment content corresponding to each candidate intervention path are recalculated and collected into the candidate intervention path set. The training of the joint intelligent model is completed in the process of continuously repeating parameter updates and path re-evaluation and collection.

[0092] S6: Utilizes a joint intelligent model for path execution and real-time feedback to continuously track and update the patient's health status.

[0093] S6.1: Based on the joint intelligent model, the candidate intervention path is executed, and the execution period observation data is generated during the execution process; the execution period observation data is compared with the health status vector to generate real-time feedback data.

[0094] Specifically, based on the joint intelligent model, candidate intervention paths to be executed are selected from the set of candidate intervention paths. The intervention actions corresponding to each time slice in the candidate intervention path are executed in the actual scenario time slice by time slice. In each time slice, health indicator change data, health behavior and execution records, biochemical test data, physiological monitoring data and patient subjective feedback data are recorded to form execution period observation data. The health indicator values ​​recorded in the execution period observation data of each time slice are aligned with the health status vector corresponding to the same time slice in the candidate intervention path. The health indicator components in the health status vector are compared with the corresponding health indicator values ​​in the execution period observation data. The differences and directions of change obtained from the comparison are organized in chronological order to generate real-time feedback data.

[0095] S6.2: Combining real-time feedback data and balanced path weights, select candidate intervention paths after path execution, output the dynamically balanced intervention strategy path execution method, and make adjustments.

[0096] Specifically, by combining real-time feedback data and balanced path weights, each candidate intervention path after execution is evaluated. The real-time feedback data is compared with the health status vector of the corresponding time slice in the candidate intervention path, and a comprehensive reference value for each candidate intervention path is calculated under the constraint of balanced path weights. Based on the comprehensive reference value, the target candidate intervention path for the current stage is selected. Using the sequence of intervention actions in the target candidate intervention path as a basis, the execution arrangement of intervention actions in subsequent time slices is modified according to real-time feedback data, generating and adjusting the dynamic balanced intervention strategy path execution method.

[0097] Before calculating the comprehensive reference value, the feedback assessment values ​​for each health goal are... Convert the values ​​to dimensionless or dimensionless values ​​according to the preset dimensional normalization factor, so that... Before weighted summation, all paths are on the same scale, representing a comprehensive reference value for each candidate intervention path. The formula is as follows: ; In the formula, Indicates the first A comprehensive reference value for candidate intervention pathways. Indicates the sequence number of the candidate intervention path. The first path weight in the balanced path weights The path weight coefficients corresponding to the candidate intervention paths, Indicates the first The weight coefficient of each health goal. Indicates the sequence number of the health goal. This represents the result calculated by combining real-time feedback data. The candidate intervention pathways in the first Feedback evaluation values ​​on health goals This represents the total number of health goals. This means that the feedback evaluation values ​​are weighted and summed over all health goals, based on the weight of each health goal. This represents the sum of the weight coefficients of all health goals, used to normalize the weighted results.

[0098] Health Target Weighting Coefficient The target sensitivity values ​​for each health objective are obtained by normalizing them within the set of health objectives. Specifically, the target sensitivity value for each health objective is divided by the sum of the sensitivity values ​​for all health objectives to obtain the result. and For example, the weighting coefficients for health goals can be set as follows: 0.4–0.6 for stable blood glucose, 0.2–0.4 for blood pressure control, and 0.1–0.3 for weight management; Path weighting coefficients... The optimal scores of each candidate intervention path are obtained by normalizing the proportions through a dynamic balancing algorithm, satisfying the following conditions: and For example, the path weight coefficient of the primary candidate intervention path can be set in the range of 0.3 to 0.7, and the path weight coefficient of the secondary candidate intervention path can be set in the range of 0.05 to 0.3.

[0099] S6.3: The adjusted intervention strategy path implementation method and real-time feedback data are used as feedback and input into the joint intelligent model to continuously track and update the patient's health status.

[0100] Specifically, the adjusted intervention strategy path execution method and real-time feedback data are imported into the joint intelligent model as joint inputs. In the joint intelligent model, the corresponding candidate intervention paths are re-deduced in chronological order, the health status vectors of each time slice are updated, and the intervention strategy content, value assessment content, and path benefit assessment content that match the updated health status vectors are output.

[0101] The updated health status vector is stored as the new health status vector in the joint intelligent model, replacing the original health status vector in the generation and evaluation of candidate intervention paths in the next stage. During subsequent interventions, the input, deduction and update operations are continuously repeated, so that the joint intelligent model can continuously track and update the patient's health status with changes in real-time feedback data.

[0102] In summary, this invention achieves coordination of intervention contributions across multiple health goals and adaptive convergence of path weights by employing a dynamic balancing algorithm during candidate intervention path training. By utilizing the response levels of health goals, goal weight coefficients, and path optimization scores to construct a comprehensive response vector across time slices, and generating balanced path weights based on an iterative normalization proportional adjustment mechanism, the contribution distribution of intervention paths in the three directions of blood glucose stabilization, blood pressure control, and weight management is dynamically corrected.

[0103] S7: Based on the continuously tracked and updated health status of the patients, extract multidimensional time-series data including blood glucose, blood pressure and weight indicators, and use long short-term memory network to construct a multi-objective evolution trend model to predict the evolution trajectory of each indicator within a preset time window.

[0104] The process begins by extracting multidimensional time-series data based on continuously tracked and updated patient health status. This program reads the patient's electronic health record through a data interface, filtering out historical data including blood glucose, blood pressure, and weight indicators. During extraction, the system aligns timestamps, mapping data from different sampling frequencies to a preset time granularity, such as 8 AM daily. For the extracted raw data, outlier detection is performed, marking values ​​exceeding medically acceptable ranges as noise and smoothing them. Specifically, the system reads continuous records from the most recent 90 time steps to form a time-series sample. Subsequently, a multi-objective evolutionary trend model is constructed using a Long Short-Term Memory (LSTM) network. This construction process does not directly call encapsulated functions but executes calculations layer by layer: the system first initializes the weight matrices and bias vectors of the forget gate, input gate, and output gate. At each time step, the current input vector is concatenated with the hidden state vector from the previous time step, multiplied by the corresponding weight matrix, and a bias term is added. The result is then processed by an activation function to obtain the control coefficients for each gate.

[0105] Based on this, the system determines the proportion of cell state information retained from the previous time step through a forget gate control coefficient, and the proportion of new information written at the current time step through an input gate control coefficient. The two are added together to update the current cell state. Next, the current cell state is filtered through an output gate control coefficient to generate the hidden state and predicted output for the current time step. This process is repeated at all time steps until the input sequence has been processed. Finally, the model outputs the evolution trajectory of various indicators within a preset time window, which includes the predicted values ​​of blood glucose, blood pressure, and weight for each day over the next 30 days.

[0106] To verify the predictive logic and parameter settings of the aforementioned evolutionary trend model, specific numerical calculations are required. First, the process of acquiring input data is defined: the system collects a patient's fasting blood glucose values ​​for the past 5 days. As shown in Table 1, this table records the patient's original monitoring data and the input data after normalization. Normalization uses the range method. The system obtains the preset safe blood glucose range of a maximum value of 10.0 mmol / L and a minimum value of 4.0 mmol / L, and calculates the difference between the two to obtain the range of 6.0 mmol / L. For the original blood glucose value of 7.0 mmol / L on the first day, the system subtracts the minimum value of 4.0 mmol / L to obtain 3.0 mmol / L, and then divides it by the range of 6.0 mmol / L to obtain the normalized input value of 0.50. Similarly, the normalized values ​​for subsequent days can be calculated.

[0107] Table 1. Patient's Historical Health Indicator Data As shown in Table 1, the system inputs the processed normalized data sequence into the Long Short-Term Memory (LSTM) network. The forget gate weight parameter is set to 0.8, and the cell state value at the previous time step is 0.6. The system performs the forget operation by multiplying the cell state value of 0.6 at the previous time step with the forget gate weight parameter 0.8 to obtain the retained component 0.48. The input gate weight parameter is set to 0.5, and the new information input at the current time step is 0.55. The system performs the write operation by multiplying the current input value of 0.55 at the current time step with the input gate weight parameter 0.5 to obtain the written component 0.275. Subsequently, the system adds the retained component 0.48 to the written component 0.275 to obtain the updated cell state value 0.755. The output gate weight parameter is set to 0.9. The system maps the updated cell state value 0.755 through a hyperbolic tangent activation logic mapping (simplified here as a linear mapping) and multiplies it with the output gate weight parameter 0.9 to calculate the normalized predicted output value of 0.6795 for the current step. Finally, the system performs inverse normalization, multiplying 0.6795 by the range of 6.0 mmol / L and adding the minimum value of 4.0 mmol / L to calculate the predicted blood glucose level for the first future time step as 8.077 mmol / L. This result indicates that, under the current trend, the patient's future blood glucose levels are likely to rise, and the model has successfully captured the evolutionary characteristics of the data sequence.

[0108] S8: Map the predicted evolution trajectory to a preset time series space, use the Monte Carlo tree search algorithm to simulate the cumulative effect of different intervention actions in the time series space, and calculate the risk-reward structure matrix of each intervention action across the time slice dimension.

[0109] The predicted evolutionary trajectory is mapped to a predefined time series space. This mapping process discretizes continuous physiological indicator predictions into state nodes, constructing a time tree representing the future development of health status. Within this space, the cumulative effect of different intervention actions is simulated using a Monte Carlo tree search algorithm. Starting from the current state node, the system performs a selection operation: based on a predefined confidence interval upper limit algorithm, it calculates the priority weight of each child node and selects the node with the highest weight for expansion. During the expansion phase, the system randomly selects an untried intervention action, such as "increasing exercise duration" or "adjusting drug dosage," to generate a new child node.

[0110] Subsequently, the simulation phase begins. Starting from newly generated child nodes, the system extrapolates forward along the timeline until the end of a preset time window is reached. During the extrapolation, the system calculates the immediate reward and risk value for each time step based on a physiological model. The immediate reward is calculated based on how close the health indicator is to the target value, while the risk value is based on whether the indicator's fluctuation exceeds a safety threshold. After the extrapolation, a backpropagation operation is performed, incrementing the visit count of all nodes along the simulation path by 1 and updating the cumulative benefit value to each parent node. Through thousands of simulation iterations, the system statistically analyzes the performance of each intervention action across different time slices and calculates the risk-reward structure matrix of each intervention action across time slices. Each row of this matrix represents an intervention action, and each column represents a future time slice. The matrix elements contain the expected average benefit and potential risk probability of the action in that time slice.

[0111] To illustrate the calculation logic of the risk-reward structure matrix, the simulation parameters are set as follows: the preset time window is the next 3 days, i.e., 3 time slices. For the intervention action "walking 30 minutes daily," the system recorded data for days 1, 2, and 3 during the simulation. First, the benefit calculation logic is defined: the system obtains the predicted and target blood glucose values ​​for a given time slice, calculates the absolute value of the difference, and subtracts this absolute value from the baseline constant 10 to obtain the base score. If the base score is less than 0, it is recorded as 0. Next, the risk calculation logic is defined: the system obtains the variance of blood glucose fluctuations within the given time slice. If the variance is greater than the preset safe fluctuation threshold of 2.0, the risk value is 1; otherwise, it is 0. For example, on day 1, the predicted blood glucose value is 8.0 mmol / L, and the target value is 6.0 mmol / L. The absolute value of the difference is 2.0, and the base score is calculated as 10 minus 2.0, which equals 8.0. The variance is 0.5, which is less than 2.0, so the risk value is 0. On Day 2, the predicted blood glucose level is 7.0 mmol / L, the target level is 6.0 mmol / L, the absolute difference is 1.0, and the baseline score is calculated as 10 minus 1.0 equals 9.0; the variance is 1.2, which is less than 2.0, and the risk is 0. On Day 3, the predicted blood glucose level is 5.5 mmol / L, the target level is 6.0 mmol / L, the absolute difference is 0.5, and the baseline score is calculated as 10 minus 0.5 equals 9.5; the variance is 2.5, which is greater than 2.0, and the risk is 1.

[0112] The system performs cumulative effect calculations: The baseline scores over three days are summed to obtain a cumulative benefit value of 8.0 + 9.0 + 9.5 = 26.5. The risk values ​​over the three days are then either logically ORed or weighted (here, weighted summation is used, with each day's risk weight being 1) to obtain a cumulative risk value of 0 + 0 + 1 = 1. Subsequently, the system constructs the corresponding elements of the risk-reward structure matrix. For the action "walking 30 minutes daily," the benefit structure item is recorded as 26.5, and the risk structure item as 1.0. Similarly, for another action "increasing the metformin dosage," simulation calculations yield a cumulative benefit value of 28.0 and a cumulative risk value of 3.0. This example demonstrates how continuous changes over time can be compressed into a quantitative indicator for evaluating the value of intervention actions, providing data support for subsequent global trade-offs.

[0113] S9: Combining the risk-return structure matrix with the preset health target weight distribution, a multi-objective evolutionary algorithm is used to perform a global dynamic trade-off analysis on candidate intervention paths to identify the non-dominated solution set when the evolution trends of multi-objective indicators are inconsistent.

[0114] Based on the generated risk-reward structure matrix and the preset health goal weight distribution, a multi-objective evolutionary algorithm is used to perform a global dynamic trade-off analysis on candidate intervention paths. The system first initializes a population containing multiple candidate intervention paths, with each individual representing a long-term intervention strategy combination. The preset health goal weight distribution is set by doctors or patients according to disease priority, such as blood glucose control weight, blood pressure control weight, and weight management weight. During the evolutionary process, the system performs a fitness assessment. For each individual in the population, the system uses the risk-reward structure matrix to find the payoff and risk values ​​corresponding to each action in the strategy combination. For each health goal, a weighted payoff sum is calculated. Specifically, the system multiplies the payoff of an action in the blood glucose dimension by the blood glucose control weight, the payoff in the blood pressure dimension by the blood pressure control weight, and the payoff in the weight dimension by the weight management weight. Subsequently, the system performs a non-dominated ranking operation. In cases where the evolutionary trends of multiple goal indicators are inconsistent (e.g., some actions can lower blood glucose but lead to weight gain), the system compares individuals in the population pairwise. If individual A performs better than or equal to individual B in all goals, and is strictly better than individual B in at least one goal, then individual A is determined to dominate individual B. Through this comparison, the system divides the population into different dominance levels, with the first level being the current set of non-dominated solutions. These solutions represent the optimal equilibrium states under the current constraints where one objective cannot be improved without compromising another.

[0115] To illustrate the specific logic of the global dynamic trade-off analysis, a numerical example is introduced. The weight distribution of health goals is set as follows: blood glucose control weight 0.5, blood pressure control weight 0.3, and weight management weight 0.2. Two candidate intervention paths are selected for comparison: Pathway A (focusing on medication): The original benefit in the blood glucose dimension is 90, the original benefit in the blood pressure dimension is 80, and the original benefit in the weight dimension is 60. Pathway B (focusing on lifestyle): The original benefit in the blood glucose dimension is 70, the original benefit in the blood pressure dimension is 85, and the original benefit in the weight dimension is 95.

[0116] The system performs a weighted overall utility calculation (used to assist in ranking or crowding calculation; although non-dominant rankings primarily rely on the Pareto principle, the weighted values ​​can be used to illustrate the trade-off logic): For path A: Blood glucose weighted score = 90 multiplied by 0.5 equals 45.0. Blood pressure weighted score = 80 multiplied by 0.3 equals 24.0. Weight weighted score = 60 multiplied by 0.2 equals 12.0. Overall utility value for path A = 45.0 + 24.0 + 12.0 equals 81.0.

[0117] For Pathway B: Blood glucose weighted score = 70 x 0.5 = 35.0. Blood pressure weighted score = 85 x 0.3 = 25.5. Weight weighted score = 95 x 0.2 = 19.0. Overall utility value of Pathway B = 35.0 + 25.5 + 19.0 = 79.5.

[0118] Although path A has a slightly higher overall weighted score, it needs to be compared item by item in the non-dominated ranking logic of the multi-objective evolutionary algorithm. The comparison logic is as follows: Blood glucose target: Path A (90) is greater than path B (70), path A is superior. Blood pressure target: Path A (80) is less than path B (85), path B is superior. Weight target: Path A (60) is less than path B (95), path B is superior. Since paths A and B have their own strengths and weaknesses, and neither completely dominates the other, both paths A and B are classified into the non-dominated solution set. The system further calculates the crowding distance to maintain the diversity of the solution set. Assuming that the solution density around path A is low and the solution density around path B is high in the solution space, the system will prioritize retaining path A to ensure the breadth of the selection range. Finally, the system identifies a set of non-dominated solution sets, as shown in Table 2, which shows the frontier solutions selected after multiple rounds of evolutionary iterations.

[0119] Table 2 Example Table of Non-Dominated Solution Sets for Multi-Objectives As shown in Table 2, Scheme 001 performs exceptionally well in blood glucose control, Scheme 002 performs well in weight and blood pressure control with low risk, and Scheme 003 provides a balanced compromise among various indicators. This result demonstrates that the algorithm successfully identifies the effective trade-off boundary under multi-objective conflict.

[0120] S10: Calculate the convergence index of each path in the non-dominated solution set across the target range, select the preferred path that meets the preset convergence condition based on the convergence index, and generate an individualized intervention path.

[0121] For the non-dominated solution set generated in the previous step, a convergence screening operation is performed to generate the final individualized intervention path. The system first defines a "convergence index across the target range." This index measures whether various health indicators can stably tend towards the ideal state during the future evolution of the intervention path, rather than experiencing drastic fluctuations. The calculation logic is as follows: For each path in the solution set, the system extracts the predicted variance sequence of various indicators (blood glucose, blood pressure, weight) within a preset future time window (e.g., 30 days). The system first calculates the coefficient of variation for each indicator, i.e., the standard deviation of the predicted value divided by the mean of the predicted value, to eliminate the influence of dimensions. Then, the system calculates the overall convergence index across the target by weighted summing of the coefficients of variation of each indicator (the weights are consistent with the aforementioned health target weights). Next, the system sets a preset convergence condition, which is represented by a threshold, for example, the convergence index must be less than 0.15. The system traverses all paths in the non-dominated solution set, calculates the convergence index for each path, and compares the calculation results with the threshold. If the convergence index of a certain path is less than 0.15, the path is deemed to meet the convergence condition and is added to the preferred candidate pool. Finally, in the preferred candidate pool, the final individualized intervention path is selected based on the principle of maximizing the overall return value.

[0122] To demonstrate the calculation and selection process of the convergence index in detail, "Scheme 003" from Table 2 is selected for calculation. The predicted blood glucose values ​​for Scheme 003 at five key time points in the future are set as follows: 7.0, 6.8, 6.9, 6.7, 6.8 (unit: mmol / L). Step 1: Calculate the mean blood glucose value: Add the five values ​​together to get 34.2, divide by 5 to get the mean of 6.84. Step 2: Calculate the variance and standard deviation of the blood glucose value: Calculate the square of the difference between each value and the mean: (7.0 - 6.84)² = 0.0256; (6.8 - 6.84)² = 0.0016; (6.9 - 6.84)² = 0.0036; (6.7 - 6.84)² = 0.0196; (6.8 - 6.84)² = 0.0016. The summation yields 0.052, which, when divided by 5, gives the variance of 0.0104. Taking the square root gives the standard deviation of approximately 0.102. The third step is to calculate the blood glucose coefficient of variation: the standard deviation of 0.102 divided by the mean of 6.84 yields 0.0149.

[0123] Similarly, assuming the calculated coefficient of variation for blood pressure is 0.0200 and for weight is 0.0050, the system calls the preset weights: blood glucose weight 0.5, blood pressure weight 0.3, and weight weight 0.2. The fourth step is to calculate the cross-target convergence index: Blood glucose component: 0.0149 multiplied by 0.5 equals 0.00745; Blood pressure component: 0.0200 multiplied by 0.3 equals 0.00600; Weight component: 0.0050 multiplied by 0.2 equals 0.00100. Adding these three together: 0.00745 plus 0.00600 plus 0.00100 equals 0.01445.

[0124] The convergence threshold was set to 0.02. The calculated result 0.01445 was compared with the threshold 0.02. Since 0.01445 is less than 0.02, the logical judgment result is "true". Therefore, the system determines that "Solution 003" has good convergence stability and can be adopted as the preferred path. This calculation process ensures that the final generated intervention path not only performs well in terms of numerical benefits, but also maintains a smooth transition of the patient's physiological state during execution, avoiding the risk of drastic fluctuations in indicators due to excessive intervention.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for personalized chronic disease progression planning based on deep learning, characterized in that, include: A standardized health status dataset is constructed, and a multidimensional feature set is generated through a representation learning process; Using a multidimensional feature set as input for association analysis, feature sets representing different health goals are extracted, and target weights are assigned to generate an individual multi-target health feature table. The individual multi-objective health feature table is input into a deep reinforcement learning model. Through adaptive learning of multi-level sub-policies, corresponding policy weights are assigned to different health objectives to generate candidate intervention paths. By incorporating risk and benefit feedback during the training of candidate intervention paths, a dynamically balanced intervention strategy is output. Implement intervention strategies, collect and organize observation data during the implementation period, form risk and benefit feedback records, and use deep reinforcement learning models and reinforcement learning algorithms to self-adjust, optimize learnable parameters, and generate joint intelligent models. By utilizing a joint intelligent model for path execution and real-time feedback, the patient's health status can be continuously tracked and updated. Based on the continuously tracked and updated health status of the patients, multidimensional time-series data including blood glucose, blood pressure and weight indicators are extracted. A multi-objective evolution trend model is constructed using a long short-term memory network to predict the evolution trajectory of each indicator within a preset time window. The predicted evolution trajectory is mapped to a preset time series space, and the Monte Carlo tree search algorithm is used to simulate the cumulative effect of different intervention actions in the time series space. The risk-reward structure matrix of each intervention action in the cross-time slice dimension is calculated. Combining the risk-return structure matrix with the preset health target weight distribution, a multi-objective evolutionary algorithm is used to perform a global dynamic trade-off analysis on candidate intervention paths to identify the non-dominated solution set when the evolution trends of multi-objective indicators are inconsistent. Calculate the convergence index of each path in the non-dominated solution set across the target range, select the preferred path that meets the preset convergence condition based on the convergence index, and generate an individualized intervention path.

2. The method for personalized chronic disease progression planning based on deep learning as described in claim 1, characterized in that, The specific steps for constructing a standardized health status dataset and generating a multidimensional feature set through a representation learning process are as follows: Collect patients' raw health data, and perform semantic alignment, unit conversion and anomaly handling to form a standardized health data record set; The standardized health data record set is sorted by timestamp and sliced ​​according to a preset time granularity. The health indicators within the same time slice are aggregated and calculated to form a standardized health status dataset. Based on a standardized health status dataset, the health indicators for each time slice are arranged in order and formed into a fixed-dimensional health status vector to generate a multi-dimensional feature set.

3. The method for personalized chronic disease progression planning based on deep learning as described in claim 2, characterized in that, The specific steps for generating an individual multi-objective health feature table by allocating target weights are as follows: Based on the multidimensional feature set, the health status vector is classified into blood glucose stability, blood pressure control and weight management to construct a set of health goals; The target sensitivity value is formed by taking the difference between the health status vector and the corresponding health indicator in the health target set as input, and through nonlinear enhancement and normalization compression. Based on the target sensitivity value, weights are assigned to all health targets in the health target set, and the weights of each health target are bound to the corresponding feature components to generate an individual multi-target health feature table.

4. The method for personalized chronic disease progression planning based on deep learning as described in claim 3, characterized in that, The specific steps for generating candidate intervention paths are as follows: The individual multi-objective health feature table is input into the deep reinforcement learning model, and the weights and feature components of each health objective are combined as the state description of the deep reinforcement learning model. Based on the state description of the deep reinforcement learning model, the health objectives are sorted according to their weights to obtain the weight priority of the health objectives. By utilizing the weights and priorities of health goals, a set of corresponding sub-policies is generated for each health goal, and each sub-policy generates a sequence of candidate actions for different health goals; Based on candidate action sequences, a deep reinforcement learning model is used to extrapolate changes in health status and generate candidate intervention paths.

5. The method for personalized chronic disease progression planning based on deep learning as described in claim 4, characterized in that, The specific steps of the intervention strategy for dynamic output balancing are as follows: The magnitude of improvement in health goals is calculated by using the health indicator components in the health status vector. Based on candidate intervention paths, the health status vector of each path is matched with the improvement magnitude of the health target, and the response degree of each path to the health target is evaluated. Based on the response level value and the weighting priority of health goals, an optimization score is assigned to each candidate intervention path; All candidate intervention paths are weighted and averaged based on the optimization scores, and a dynamic balancing algorithm is then executed. Based on the output of the dynamic balancing algorithm, candidate intervention paths with optimized scores and weighted averages are selected, and dynamic balancing intervention strategies are output.

6. The method for personalized chronic disease progression planning based on deep learning as described in claim 5, characterized in that, The output of the dynamic balancing algorithm refers to the coordinated intervention contribution among multiple health goals and the generation of balanced path weights through the optimized scoring of candidate intervention paths.

7. The method for personalized chronic disease progression planning based on deep learning as described in claim 6, characterized in that, The observation data during the implementation period includes data on changes in health indicators, health behavior and implementation records, biochemical test data, physiological monitoring data, and patient subjective feedback data.

8. The method for personalized chronic disease progression planning based on deep learning as described in claim 7, characterized in that, The learnable parameters within the deep reinforcement learning model refer to the set of parameters that are updated based on observation data during the execution phase of the deep reinforcement learning process.

9. The method for personalized chronic disease progression planning based on deep learning as described in claim 8, characterized in that, The specific process for generating the joint intelligent model is as follows: The observation data during the execution period are paired with candidate intervention paths to construct feedback samples for the deep reinforcement learning model; Based on the feedback samples, the learnable parameters inside the deep reinforcement learning model are updated using the observation data during the execution period. After the learnable parameters are updated, the candidate intervention paths are evaluated to generate intervention strategy content, value assessment content, and path benefit assessment content, and then integrated. A joint intelligent model is generated based on the integrated intervention strategy content, value assessment content, and path benefit assessment content.

10. The method for personalized chronic disease progression planning based on deep learning as described in claim 9, characterized in that, The specific process for continuously tracking and updating the patient's health status is as follows. Based on the joint intelligent model, candidate intervention paths are executed, and execution period observation data is generated during the execution process; The execution period observation data is compared with the health status vector to generate real-time feedback data; By combining real-time feedback data and balanced path weights, candidate intervention paths are selected after path execution, and the dynamic balanced intervention strategy path execution mode is output and adjusted. The adjusted intervention strategy and real-time feedback data are used as feedback and input into the joint intelligent model to continuously track and update the patient's health status.

Images