Intelligent decision and intervention system for chronic disease health management based on deep learning

Abstract

The application discloses a chronic disease health management intelligent decision and intervention system based on deep learning, relates to the technical field of intelligent medical treatment, and comprises a multi-source data acquisition module, which is used for collecting physiological indexes, genomics data, clinical information and related detection data of patients, and performing unified pretreatment and feature extraction to form standardized data input; a comorbidity relationship reasoning module, which references the standardized input feature vector, constructs a heterogeneous graph network, and mines the cascade effect among nodes through a graph neural network algorithm, simultaneously identifies the influence path and potential conflict relationship among different diseases, and outputs topological structure data; a virtual intervention experiment module, which references the topological structure data and the standardized data input, fuses pathological correlation information and individual difference information, constructs a personalized physiological mechanism mapping function, and combines to generate a personalized virtual mirror image; and a personalized drug recommendation module, which is used for constructing a multi-task learning framework through simulation results and metabolic enzyme genotypes of the patients.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to a deep learning-based intelligent decision-making and intervention system for chronic disease health management. Background Technology

[0002] With the aging population and changing lifestyles, chronic diseases have become a major challenge in global public health. According to the World Health Organization, chronic non-communicable diseases such as cardiovascular disease, diabetes, and chronic respiratory diseases account for a continuously rising proportion of the global disease burden. Traditional chronic disease management mainly relies on regular outpatient visits, manual follow-ups, and paper-based health records, which suffer from problems such as lagging information updates, untimely intervention responses, and low personalization. In recent years, with the rapid development of artificial intelligence, big data, and the Internet of Things, health management systems based on deep learning have gradually become a research hotspot. Collecting multimodal physiological data through wearable devices and remote monitoring terminals, and combining this with machine learning algorithms for risk prediction and health assessment, has become an important direction for the development of smart healthcare.

[0003] Current algorithm models mostly focus on single-task prediction, such as only classifying disease risks or generating health advice, lacking consideration for multi-task joint optimization. This leads to isolated decision results and a lack of continuity and systematicity in intervention strategies. Existing systems generally lack in-depth modeling of non-medical factors such as patient behavior compliance, psychological state, and social support, making it difficult to achieve precise behavioral intervention and personalized decision support. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based intelligent decision-making and intervention system for chronic disease health management. This addresses the current problem that algorithm models often focus on single-task prediction, such as only classifying disease risks or generating health recommendations, and lack consideration for multi-task joint optimization. This results in isolated decision-making results and a lack of continuity and systematicity in intervention strategies. Existing systems generally lack in-depth modeling of non-medical factors such as patient behavioral compliance, psychological state, and social support, making it difficult to achieve precise behavioral intervention and personalized decision support.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a deep learning-based intelligent decision-making and intervention system for chronic disease health management, which includes a multi-source data acquisition module for collecting patients' physiological indicators, genomic data, clinical information and related test data, and performing unified preprocessing and feature extraction to form standardized data input; The comorbidity reasoning module uses standardized input feature vectors to construct a heterogeneous graph network and mines the cascading effects between nodes through graph neural network algorithms. At the same time, it identifies the influence paths and potential conflict relationships between different diseases and outputs topological structure data. The virtual intervention experiment module uses topological data and standardized data input, integrates pathological correlation information and individual difference information, constructs a personalized physiological mechanism mapping function, and generates a personalized virtual image. The personalized medication recommendation module uses simulation results and patients' metabolic enzyme genotypes to construct a multi-task learning framework. It uses genetic information as a hard constraint to simultaneously predict drug efficacy and adverse reaction risks, and outputs drug dosage and combination. The dynamic decision-making intervention module generates real-time personalized health management strategies through medication dosage and combination, and pushes intervention plans through wearable devices.

[0007] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the multi-source data acquisition module includes: Perform integrity identification and categorized completion processing on multi-source data, and introduce data reliability indicators; Data is aligned and standardized based on a unified time reference to form standardized data input.

[0008] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the comorbidity reasoning module includes: Based on standardized data input, different disease states and physiological indicators are abstracted into nodes in a graph structure; By combining historical co-occurrence relationships and changing trends, connections between nodes are established, forming multiple types of association structures; A heterogeneous graph network is constructed based on the nodes and their relationships, and the graph neural network is used to train the heterogeneous graph. The influence relationships between nodes are updated through multiple rounds of propagation. In the updated network structure, the transmission paths and influence chains between diseases are identified, and the conflict relationships between different disease intervention paths are further identified, including mutual inhibition relationships and negative influence relationships. Based on the conflict relationships, a conflict constraint structure is constructed, and the conflict constraint structure is organized into topological data containing cooperative and conflict relationships.

[0009] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the virtual intervention experiment module includes: Based on topological data, the association paths between diseases and their influence relationships at different transmission levels are extracted, and a multi-level disease association structure is formed according to path length and node connection strength. Combining data reliability identifiers and time identifiers, patient individual characteristics are stratified. The input data is divided into current status data layer, recent data layer and historical data layer according to time order, and different participation levels are set for each layer of data based on data reliability. Before performing path mapping, the candidate transmission paths in the multi-level disease association structure are screened, and the paths are prioritized according to the time location and reliability level of the data involved in the path. Higher-priority propagation pathways are selected as the primary pathways to participate in the construction of individual pathological association expression, while lower-priority propagation pathways are selected as auxiliary pathways to participate in supplementary calculations. In the process of constructing individual pathological association expression, data in the current state data layer participates in the calculation of the main action path, data in the recent data layer participates in the calculation of the auxiliary path, and data in the historical data layer participates in the calculation of the marginal path. The individual characteristics after stratification are mapped to the selected propagation paths to form individual pathological association expressions with path priority. Based on the individual pathological association expression, a physiological mechanism mapping model is constructed, so that different input data in the physiological mechanism mapping model correspond to different path ranges and order of action.

[0010] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the virtual image construction includes: Based on the mapping results of individual physiological mechanisms, a virtual individual model representing individual differences among patients is constructed, and the parameters of the virtual individual model are dynamically calibrated using real data; By using calibrated virtual individual models, different intervention programs are simulated and extrapolated to obtain results of physiological state changes and generate personalized virtual images for decision support.

[0011] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the personalized medication recommendation module includes: A set of candidate drug regimens is constructed based on virtual mirrors, and the action path of each candidate drug regimen in the individual physiological mechanism mapping model is simulated to obtain the position and range of action of the drug in different disease transmission paths. The path priority relationship is invoked to perform path-level filtering on the input data of candidate drug regimens, so that the current state data layer participates in the calculation of the main action path, the recent data layer participates in the calculation of the auxiliary path, and the historical data layer participates in the trend reference. Based on data reliability identification, priority participation rules are set for data participating in the main action path, and data with lower reliability is restricted from entering the critical path calculation process; Based on the conflict relationships in the topological data, path-level conflict detection is performed on the role of candidate drug regimens in different disease transmission pathways to identify the mutual inhibition and negative influence relationships of drug combinations in different pathways. Under the constraints of conflict detection results, the candidate drug schemes are screened and reconstructed at the path level, drug combinations with path conflicts are eliminated, and the remaining schemes are ranked according to the degree of path matching. Based on the screening and sorting results, a medication dosage and drug combination plan is generated; Based on the generated medication dosage and drug combination plan, the disease transmission path is marked with feedback, and the path priority is adjusted according to the conflict and matching degree in the path.

[0012] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the genetic information constraints include: The metabolic enzyme genotype information is transformed into constraints that the model can recognize. These constraints are then introduced into the model prediction process to make the prediction results conform to the individual's metabolic characteristics. The prediction results are screened, and medication regimens that do not match the gene characteristics are eliminated. The remaining regimens are then optimized to determine the medication combination.

[0013] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the dynamic decision-making and intervention module includes: Initial management strategies are established by using drug treatment parameters, and real-time physiological data of patients are acquired and integrated. Based on the physiological state data, the initial management strategy is corrected and optimized in real time to generate personalized intervention instructions that are adapted to the current physiological state.

[0014] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the dynamic adjustment includes: Evaluate the effectiveness of current intervention strategies based on patient feedback and identify deviations in strategy implementation; Generate strategy adjustment instructions based on the evaluation results; In response to the adjustment instruction, the current intervention strategy is iteratively optimized, and an updated intervention plan is output.

[0015] As a preferred embodiment of the deep learning-based intelligent decision-making and intervention system for chronic disease health management described in this invention, the proposed intervention scheme includes: Convert the updated intervention strategy into executable instructions; The instructions are encapsulated and sent to the user terminal via a wearable device; The terminal parses and displays the instructions, thereby providing patients with personalized health management plans.

[0016] The beneficial effects of this invention are as follows: By utilizing graph neural network algorithms to mine the cascading effects between nodes, identifying potential conflicting relationships of synergistic deterioration or mutual inhibition between diseases and outputting topological structure data, it realizes the transformation from surface statistical co-occurrence analysis to deep pathological causal logic mapping. It integrates pathological association information and individual difference information to construct a personalized physiological mechanism mapping function, and selects effective pathological pathways based on the path net effect coefficient to generate personalized virtual mirrors. This achieves high-precision simulation of patient intervention response in digital space and pruning of invalid search space, thereby defining precise search boundaries for subsequent drug recommendations, reducing computational redundancy and ensuring that intervention plans are supported by pathological mechanisms. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a schematic diagram of a deep learning-based intelligent decision-making and intervention system for chronic disease health management. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] 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.

[0022] Reference Figure 1This is one embodiment of the present invention, which provides a deep learning-based intelligent decision-making and intervention system for chronic disease health management, comprising the following steps: S1, Multi-source data acquisition module, is used to collect patients' physiological indicators, genomic data, clinical information and related test data, and perform unified preprocessing and feature extraction to form standardized data input.

[0023] Furthermore, the system performs integrity identification and categorized completion processing on multi-source data, and introduces data reliability identifiers.

[0024] Data is aligned and standardized based on a unified time reference to form standardized data input.

[0025] It should be noted that in actual clinical settings, patients’ genomic data, physiological indicators and clinical medical records are often stored in different systems, and the update frequency and data accuracy vary.

[0026] If directly input into a complex inference model, the inference results are easily biased due to data noise or timing misalignment.

[0027] In real-world healthcare settings, patient health data typically originates from various systems, including hospital testing systems, wearable devices, patient self-reports, and historical electronic medical records.

[0028] These data not only come from heterogeneous sources, but also exhibit significant differences in collection accuracy and reliability.

[0029] Existing technologies typically process data using uniform weighting or simple averaging, but this approach ignores the reliability differences between different data sources.

[0030] For example, hospital test reports and patient self-reported information are obviously different in terms of credibility. If they are directly used as equally weighted inputs in subsequent reasoning models, the reasoning results will deviate from the true pathological state.

[0031] In the process of fusion of multi-source medical data, there are differences in the credibility of different data sources.

[0032] If data reliability is not differentiated, low-reliability data may have an excessive impact on key diagnostic indicators, thereby reducing overall inference accuracy.

[0033] The reliability between different data sources is not a simple discrete classification relationship, but rather exhibits a continuously changing characteristic.

[0034] Construct a quantitative mechanism that can continuously characterize the credibility of data, so that data forms an adjustable influence weight when it enters the system.

[0035] For any medical data collected The system assigns a data reliability identifier. The range of values ​​is: .

[0036] Among them: Test reports from top-tier hospitals: Wearable device test data: Patient subjective description data: .

[0037] By introducing reliability parameters, data from different sources can form a credibility hierarchy structure when entering the system.

[0038] To address the issue of inconsistent data collection times, a unified timeline is introduced. Time alignment of the data.

[0039] Multi-source medical data typically exhibits significant temporal asynchrony during the acquisition process, with differences in data update frequency and time accuracy across different systems.

[0040] If data from different time points are directly input into the model as if they were at the same time, it will cause a temporal misalignment in the expression of pathological states, thus affecting the determination of causal relationships.

[0041] Construct a unified time reference system so that data from different sources can be expressed under the same time benchmark, thereby ensuring that various indicators have consistent time semantics when participating in calculations.

[0042] For asynchronous data, it is mapped to a unified time base using interpolation or forward padding: ; By using credibility identifiers to differentiate the reliability of different data sources, high-credibility data can have a greater impact in subsequent model calculations.

[0043] By unifying the timeline mechanism, the temporal discrepancies between different medical data are eliminated, thereby providing a consistent data foundation for subsequent pathological reasoning.

[0044] S2, the comorbidity reasoning module, uses standardized input feature vectors to construct a heterogeneous graph network, and uses graph neural network algorithms to mine the cascading effects between nodes. At the same time, it identifies the influence paths and potential conflict relationships between different diseases and outputs topological structure data.

[0045] Furthermore, based on standardized data input, different disease states and physiological indicators are abstracted into nodes in a graph structure.

[0046] By combining historical co-occurrence relationships and changing trends, connections between nodes are established, forming multi-type association structures.

[0047] A heterogeneous graph network is constructed based on nodes and their relationships, and a graph neural network is used to train the heterogeneous graph. The influence relationships between nodes are updated through multiple rounds of propagation.

[0048] In the updated network structure, the transmission paths and influence chains between diseases are identified, and the conflict relationships between different disease intervention paths are further identified, including mutual inhibition and negative influence relationships.

[0049] A conflict constraint structure is constructed based on conflict relationships, and the conflict constraint structure is organized into topological data containing cooperative and conflict relationships.

[0050] It should be noted that traditional comorbidity analysis is often limited to statistical co-occurrence rates and is difficult to capture the deep pathological logic of how disease A indirectly inhibits the treatment of disease C through intermediate indicator B.

[0051] In real-world medical settings, there are often complex interactions between various diseases.

[0052] For example, some medications used to treat disease A may worsen disease B.

[0053] Existing technologies typically rely on statistical co-occurrence probability analysis to identify comorbidity relationships, but this approach cannot identify causal transmission pathways or potential conflicting relationships between diseases.

[0054] In scenarios where multiple diseases coexist, the influence relationships between diseases typically present as a multi-level, multi-path network structure.

[0055] Traditional statistical co-occurrence-based analysis methods can only reflect superficial associations and are insufficient to characterize the complex mechanisms by which diseases are indirectly transmitted through intermediate indicators.

[0056] By constructing a structured model that can express multi-node, multi-path propagation relationships, we can reveal the potential chain of pathological impacts.

[0057] Constructing a graph model: : in, For disease or physiological indicator nodes, These represent pathological association edges between diseases.

[0058] To distinguish between different types of disease relationships, an edge weight symbol mechanism is defined: If the two diseases have a co-exacerbating relationship: .

[0059] If there is a treatment conflict or inhibitory relationship between two diseases: .

[0060] During propagation in a graph network, the node state is updated through the message passing mechanism of the graph neural network: ; in, For node status, Weighting for pathological impact. It is the set of adjacent nodes.

[0061] By using the pathological relationship weight propagation mechanism, the influence between diseases can spread layer by layer along the network path. By distinguishing between synergistic and conflicting relationships through positive and negative weights, the system can identify potential treatment conflicts, thereby providing structural constraints for subsequent medication decisions.

[0062] Node states in a graph network and border rights It is not only used for path identification, but also serves as an input variable for subsequent path priority calculation.

[0063] The node state will participate in the data weight function calculation process to correct the strength of the data's influence on the path, so that the path priority depends not only on the data itself, but also on the pathological transmission structure.

[0064] By embedding the graph network output into the data weight calculation process, path evaluation is extended from "data-driven" to "data and structure-driven".

[0065] S3, Virtual Intervention Experiment Module, uses topological structure data and standardized data input, integrates pathological correlation information and individual difference information, constructs personalized physiological mechanism mapping functions, and generates personalized virtual images.

[0066] Furthermore, based on topological data, the association paths between diseases and their influence relationships at different transmission levels are extracted, and a multi-level disease association structure is formed according to path length and node connection strength.

[0067] By combining data reliability identifiers and time identifiers, patient individual characteristics are stratified. The input data is divided into current status data layer, recent data layer and historical data layer according to time order, and different participation levels are set for each layer of data based on data reliability.

[0068] Before performing path mapping, candidate transmission paths in the multi-level disease association structure are screened and prioritized according to the time location and reliability level of the data involved in the path.

[0069] Higher-priority propagation pathways are selected as the primary pathways for constructing individual pathological association expressions, while lower-priority propagation pathways are selected as secondary pathways for supplementary calculations.

[0070] In the process of constructing individual pathological association expression, data in the current state data layer participates in the calculation of the main action path, data in the recent data layer participates in the calculation of the auxiliary path, and data in the historical data layer participates in the calculation of the marginal path.

[0071] The individual characteristics after stratification are mapped to the selected propagation paths to form individual pathological association expressions with path priority.

[0072] A physiological mechanism mapping model is constructed based on individual pathological association expression, so that different input data in the physiological mechanism mapping model correspond to different path ranges and order of action.

[0073] Based on the mapping results of individual physiological mechanisms, a virtual individual model representing individual differences among patients is constructed, and the parameters of the virtual individual model are dynamically calibrated using real data.

[0074] By using calibrated virtual individual models, different intervention programs are simulated and extrapolated to obtain results of physiological state changes and generate personalized virtual images for decision support.

[0075] It should be noted that when multi-source medical data is used in complex disease reasoning, if all data are used in the calculation with the same weight, two problems are likely to arise: Historical data may overly influence current pathological judgments, and low-reliability data may interfere with key diagnostic indicators.

[0076] This is achieved through a data layering and path priority calculation mechanism.

[0077] When designing the data participation mechanism, we first analyze how multi-source medical data plays a role in the actual reasoning process.

[0078] Data at different time scales do not play an equivalent role in pathological diagnosis: current data directly reflects the patient's immediate physiological state and plays a decisive role; recent data reflects short-term trends and has an auxiliary corrective role in state assessment; historical data is mainly used to depict long-term evolutionary background and its role is more reflected in trend constraints rather than direct decision-making.

[0079] First, the data is structurally split, transforming it from a "flat set" into a "hierarchical structure".

[0080] This structural transformation transforms data that originally participated in calculations without distinction into an input system with a clear hierarchy of roles.

[0081] After determining the hierarchical structure, it is necessary to solve the problem of how data from different levels participate in the calculation.

[0082] If a fixed weight allocation is used (for example, manually setting the weight of the current layer to 0.7), it will not be able to adapt to the data differences between different patients.

[0083] By introducing a "computable weighting mechanism," data weights are no longer manually set but dynamically determined by data attributes, thus completing the shift from "rule-driven" to "data-driven" approaches. Based on the time attribute of the data, medical data is divided into three layers: Current data layer Recent historical level and long-term historical layer .

[0084] The importance of medical data is usually influenced by both the credibility of the data source and the time of collection.

[0085] Considering only one of these factors may lead to bias in weighting assessment, for example: Highly reliable but outdated data or recent but low-reliability data.

[0086] By constructing a unified weighting mechanism that can simultaneously integrate "credibility" and "time decay", the degree of data influence can dynamically change with time and quality.

[0087] To avoid a single factor having a dominant influence on the outcome, a fusion approach that reflects the dual constraints should be adopted.

[0088] Define the data weighting function: ; in, For data reliability coefficient, This is the time decay coefficient.

[0089] When constructing the data weighting function, first consider two basic variables: data credibility. and data time difference .

[0090] If a simple linear relationship is used to construct the weights, when lower and Even at higher levels, low-confidence data may still be assigned higher weights. When the impact is high but the data is outdated, it is difficult to effectively suppress the effect in a timely manner.

[0091] This demonstrates that the additive model cannot form a "double constraint".

[0092] Adjust the model to a product form: ; pass By screening data sources for quality, the overall weight of low-reliability data is limited. We apply decay control to the time dimension to gradually reduce the impact of historical data.

[0093] By coupling the two, data only has high weight when it simultaneously meets the requirements of "high reliability + high timeliness".

[0094] This will enable the data participation mechanism to shift from "single-factor evaluation" to "multi-factor coupled evaluation".

[0095] The importance of medical data is usually influenced by two factors: the credibility of the data source and the time of data collection.

[0096] The two factors are combined by multiplication.

[0097] The reference value of medical data typically exhibits a non-linear decay characteristic over time, meaning it changes slowly in the short term but decreases rapidly over a long period.

[0098] We adopt a function form that reflects the nonlinear decay characteristics to better align with actual clinical understanding.

[0099] The time decay function is defined as: ; When determining the form of the time decay function, two types of models, linear decay and exponential decay, were compared: Linear model: The weights decrease uniformly over time, but it cannot reflect the actual medical phenomenon of "short-term stability and long-term sharp decline".

[0100] Exponential model: It shows small changes over short time periods, but declines rapidly as the time span increases, which is more consistent with the characteristics of changes in the reference value of clinical data.

[0101] Using an exponential decay function: By using an exponential approach, recent data maintains a stable weight over a short period, avoiding excessive fluctuations. The exponential decay accelerates the decrease in the weight of longer-term data, making the system pay more attention to the current state.

[0102] Through parameters Adjusting the decay rate allows the model to adapt to the time sensitivity of different disease types.

[0103] This enables non-linear control of data weights by the time factor.

[0104] The time decay coefficient gradually reduces the weight of earlier collected data, while the reliability coefficient gives higher computational weight to high-quality medical data.

[0105] This ensures that historical information remains valuable for reference while allowing the system to focus more on the patient's current condition.

[0106] The importance of a disease transmission route is usually determined by multiple nodes in the route, rather than a single key node.

[0107] If the evaluation is based only on local nodes, the overall impact of the path may be underestimated or overestimated.

[0108] By constructing an evaluation mechanism that can comprehensively reflect the impact of the entire path.

[0109] The data weights of each node in the path are accumulated, and a path priority function is defined: ; During the path priority modeling process, it was found that path evaluation based solely on the accumulation of node weights is easily affected by the abnormal weight of a single node, which can lead to deviations in the path scoring results.

[0110] Especially when there are fluctuations or noise in multi-source medical data, the stability of path scoring is insufficient.

[0111] To avoid path scores being influenced by only a single data distribution, a path stability coefficient is introduced: ; in, This represents the variance of the node weights in the path.

[0112] By introducing a path stability coefficient, paths with more uniform node weight distribution within the path can receive higher evaluations, thereby suppressing paths with large weight fluctuations. This reduces the impact of abnormal data on path evaluation results and improves the stability and reliability of path selection.

[0113] In the process of path evaluation, it is still difficult to reflect the effect of the path in the actual intervention process based solely on the path structure and data weights.

[0114] By characterizing the stability of a path by measuring the degree of fluctuation in the internal weights of the path, highly volatile paths are suppressed, making path selection more robust. This upgrades the path priority evaluation from "intensity evaluation" to "joint evaluation of intensity and stability".

[0115] Introducing the path net effect coefficient: ; in, This indicates the output result of the virtual image when the path is active. This indicates the output result of the virtual image under path suppression.

[0116] By simulating the differences in system output under two states of path activation and path inhibition, the actual strength of the path's role in individual physiological mechanisms is characterized, thus enabling path evaluation to not only rely on structural relationships but also to be corrected in conjunction with actual intervention results.

[0117] Path scoring is defined as: ; By jointly calculating path weights, path stability, and net path effects, the path score can simultaneously reflect the overall impact capacity, internal structural stability, and actual intervention effects of the path, thereby improving the accuracy of path priority evaluation.

[0118] In the process of path priority modeling, the first consideration is the goal of path evaluation, namely, to identify the pathological transmission chain that has the greatest impact on the patient in the current state.

[0119] By introducing the path net effect coefficient, the path priority evaluation is no longer based solely on the static graph topology and node weights, but is further dynamically corrected based on the actual effect of the path in the individual virtual mirror.

[0120] Based on the calculated path net effect coefficient Construct a "virtual intervention feasible domain constraint set".

[0121] Set a net effect threshold If a candidate path If so, the path is determined to be a "low-response pathological pathway".

[0122] The set of nodes identified as "low-response pathological pathways" is mapped to a logical constraint mask Mconstraint, and the constraint mask is passed to the personalized medication recommendation module as a pre-filter condition.

[0123] This restricts or eliminates low-response paths in subsequent medication regimen calculations, thereby reducing the interference of invalid paths on the decision-making process, narrowing the scope of medication regimen search, and improving the targeting and computational efficiency of medication recommendations.

[0124] Only the largest node weight in the path is selected as the path weight, or a penalty adjustment is applied based on the path length.

[0125] However, there are limitations: the single-node method ignores the synergistic effect of multiple nodes, and the path length method cannot reflect the differences in node quality.

[0126] The path evaluation problem is transformed into a "multi-node comprehensive contribution problem," and a cumulative approach is adopted: By aggregating the weights of each node in the path, the path score reflects the overall impact capability. By introducing the aforementioned weight function, the path evaluation is simultaneously affected by both time and credibility.

[0127] This enables a model shift from "local judgment" to "global path judgment".

[0128] Since the importance of a disease transmission route is usually determined by multiple indicators within the route, an additive approach is used to calculate the overall weight of the route.

[0129] The importance of the disease transmission chain is quantified by accumulating path weights, and the most critical pathological transmission path is identified by prioritizing the path, thereby providing a structural basis for subsequent treatment decisions.

[0130] S4, Personalized Medication Recommendation Module, uses simulation results and patients' metabolic enzyme genotypes to construct a multi-task learning framework. It uses genetic information as a hard constraint to simultaneously predict drug efficacy and adverse reaction risks, and outputs drug dosage and combination.

[0131] Furthermore, a set of candidate drug regimens is constructed based on virtual images, and the action path of each candidate drug regimen in the individual physiological mechanism mapping model is simulated to obtain the position and range of action of the drug in different disease transmission paths.

[0132] The path priority relationship is invoked to perform path-level filtering on the input data of candidate drug regimens, so that the current state data layer participates in the calculation of the main action path, the recent data layer participates in the calculation of the auxiliary path, and the historical data layer participates in the trend reference.

[0133] By combining data reliability identifiers, priority participation rules are set for data participating in the main action path, and data with lower reliability is restricted from entering the critical path calculation process.

[0134] Based on the conflict relationships in the topological data, path-level conflict detection is performed on the role of candidate drug regimens in different disease transmission pathways to identify the mutual inhibition and negative influence relationships of drug combinations in different pathways.

[0135] Under the constraints of conflict detection results, candidate drug schemes are screened and reconstructed at the path level, drug combinations with path conflicts are eliminated, and the remaining schemes are ranked according to the degree of path matching.

[0136] Based on the screening and sorting results, a medication dosage and drug combination plan is generated.

[0137] Based on the generated medication dosage and drug combination plan, the disease transmission path is marked with feedback, and the path priority is adjusted according to the conflict and matching degree in the path.

[0138] Metabolic enzyme genotype information is transformed into constraints that the model can recognize. These constraints are then introduced into the model prediction process to ensure that the prediction results conform to the individual's metabolic characteristics.

[0139] The prediction results are screened, and medication regimens that do not match the gene characteristics are eliminated. The remaining regimens are then optimized to determine the medication combination.

[0140] It should be noted that in the treatment of comorbidities, different drugs may act on multiple disease pathways simultaneously, resulting in complex interactions.

[0141] A drug that treats disease A may affect disease B through a certain pathological pathway. If this pathway conflict is not quantitatively analyzed, it is possible that treating one type of disease may worsen another type of disease.

[0142] The path priority result is not only used for sorting, but also directly participates in the medication regimen generation process as a constraint.

[0143] Only high-priority paths are allowed to participate in the action path matching of candidate drug protocols, while low-priority paths are weakened or eliminated.

[0144] By limiting the drug search space through path priority, the decision-making process is transformed from "global search" to "path-constrained search".

[0145] By designing a path conflict quantification function, we can assume that the drug regimen is on the path... The intensity of the action on is If the path and There is a conflict relationship .

[0146] In the treatment of comorbidities, the risk of drug conflict not only stems from the conflict between pathways, but is also affected by the intensity of the drug's effect on each pathway.

[0147] When a drug acts on multiple conflicting pathways simultaneously, the synergistic effect may amplify the overall risk.

[0148] By constructing a quantitative model that can simultaneously characterize the "degree of path conflict" and the "intensity of drug action".

[0149] Define the conflict cost function: ; The results of the medication conflict assessment will have a reverse effect on the path weight update process.

[0150] Define the path feedback update function: ; in, To mitigate the risk of path conflict, This is the adjustment coefficient.

[0151] By mitigating path weights through conflict risk, the impact of high-risk paths in subsequent calculations is reduced. The feedback mechanism enables the path structure to be dynamically adjusted based on the decision results.

[0152] In the context of co-medication, the risk of conflict is not caused by a single factor, but by the combination of multiple factors.

[0153] By analyzing actual clinical cases, the main sources of conflict risk can be summarized as follows: The conflict between pathways, the intensity of drug action on each pathway, and the synergistic effect of simultaneous action on multiple pathways.

[0154] The initial modeling attempt was to use the following approach: only count the number of conflicting paths or only consider the largest conflicting path.

[0155] However, it fails to reflect the following issues: differences in the intensity of conflicts along the same pathway, the amplifying effect of drug action on conflicts, and the nonlinear risks arising from the superposition of multiple pathways.

[0156] Therefore, conflict modeling is transformed into a three-factor coupling problem, and the following product model is constructed: pass Indicates the intensity of path conflict. It indicates the intensity of the drug's effect along the pathway and reflects the multi-path linkage effect through a product form.

[0157] This transforms the discrete conflict problem into a continuous computable model.

[0158] Drug conflict risk is typically determined by three factors: the intensity of conflict between pathways, the intensity of drug action along the pathway, and the synergistic effect between multiple pathways.

[0159] By quantifying the severity of path conflicts through the absolute value of conflict weights and reflecting the multi-path linkage effect through the product of drug action intensity, the complex comorbid treatment conflict is transformed into a calculable indicator.

[0160] In complex comorbidity treatment scenarios, the choice of medication regimens needs to balance efficacy and safety.

[0161] If we only pursue maximizing the therapeutic effect, we may introduce higher treatment risks; if we only control the risks, we may reduce the therapeutic effect.

[0162] By constructing a unified multi-objective optimization framework, it is possible to coordinate between returns and risks.

[0163] Construct the constrained optimization objective function: ; Limited by: ; in, This refers to the set of pathological pathways involved in the candidate drug regimens.

[0164] Although the drug combination has a theoretical therapeutic effect If the effect is high, but the main pathological pathway is judged as having a "low net effect" (i.e., no significant improvement after intervention) in the virtual simulation, then the drug combination will be judged as not meeting the constraints and thus excluded from the feasible solution set.

[0165] The results of virtual simulation directly guide the search direction, making drug recommendations no longer blind mathematical optimization, but precise decisions based on pathological mechanism verification.

[0166] S5, the dynamic decision-making intervention module, generates real-time personalized health management strategies through medication dosage and combination, and pushes intervention plans through wearable devices.

[0167] Furthermore, initial management strategies are established through drug treatment parameters, and real-time physiological status data of patients are acquired and integrated.

[0168] Based on physiological state data, the initial management strategy is corrected and optimized in real time, generating personalized intervention instructions that are adapted to the current physiological state.

[0169] The effectiveness of the current intervention strategy is evaluated based on patient feedback, and deviations in strategy implementation are identified.

[0170] Based on the evaluation results, strategy adjustment instructions are generated.

[0171] In response to adjustment instructions, the system iteratively optimizes the current intervention strategy and outputs an updated intervention plan.

[0172] The updated intervention strategy was converted into executable instructions.

[0173] The instructions are encapsulated and sent to the user terminal via wearable devices.

[0174] The terminal parses and displays the instructions, thereby providing patients with personalized health management plans.

[0175] It should be noted that in existing technologies, health management plans are often periodic (such as monthly follow-up visits), which has a significant lag. This solution, by introducing real-time data streams from wearable devices, increases the update frequency of intervention strategies to the minute or even second level.

[0176] It employs a four-step cyclical logic of "initial strategy - real-time correction - performance evaluation - iterative optimization".

[0177] set up Intervention strategies at specific times The system obtains real-time physiological status through wearable devices. and the state predicted by the virtual image. Compare them.

[0178] If there is a deviation If the threshold is exceeded, the system will trigger a correction mechanism, updating the policy parameters using the gradient descent approach.

[0179] Patients' physiological state is dynamic during intervention, and fixed strategies are difficult to adapt to individual changes.

[0180] By introducing a feedback-based adaptive adjustment mechanism, the system can continuously optimize the intervention strategy based on the actual physiological response.

[0181] To achieve continuous optimization, an iterative parameter tuning method is constructed.

[0182] Using parameter update method: ; in This is the learning rate.

[0183] In the process of adjusting intervention strategies, we first analyze existing methods: rule-based adjustment (such as adjusting when a threshold is exceeded) and periodic correction based on human experience.

[0184] However, it has obvious shortcomings: it cannot respond to continuous changes, the adjustment range is uncontrollable, and it is difficult to adapt to individual differences.

[0185] Therefore, the problem is transformed into a "continuous optimization problem" and a parameter update mechanism is introduced.

[0186] In the modeling process, drawing on optimization theory, policy adjustment is viewed as a process of minimizing the loss function: The loss function characterizes the difference between the predicted state and the true state, the gradient term determines the direction of parameter adjustment, and the learning rate controls the adjustment magnitude.

[0187] This enables a shift from a "discrete adjustment" to a "continuous optimization" mechanism.

[0188] The patient's physical response is used as a feedback signal to correct the intervention instructions.

[0189] By encapsulating and parsing instructions through wearable devices, a seamless connection between medical-grade decision-making and consumer-grade terminals is achieved.

[0190] As feedback data accumulates, the system's calibration of individual pathological mechanism mapping models will become more accurate, thus forming a virtuous cycle of "becoming more accurate with use".

[0191] Dynamic feedback results are also used to update data reliability indicators. .

[0192] When a data source exhibits a high deviation in multiple rounds of decision-making, the reliability coefficient is dynamically adjusted downwards; when a data source consistently and stably matches the prediction results, the reliability coefficient is adjusted upwards.

[0193] The credibility of data is corrected by reversing the decision results, so that the data weights are dynamically adjusted as the system operates.

[0194] In summary, this invention achieves a shift from surface-level statistical co-occurrence analysis to deep pathological causal logic mapping by utilizing graph neural network algorithms to mine cascading effects between nodes, identifying potential conflicting relationships of synergistic deterioration or mutual inhibition between diseases, and outputting topological structure data. It integrates pathological association information with individual difference information to construct personalized physiological mechanism mapping functions, and selects effective pathological pathways based on path net effect coefficients to generate personalized virtual mirrors. This enables high-precision simulation of patient intervention responses in digital space and pruning of invalid search spaces, thereby defining precise search boundaries for subsequent drug recommendations, reducing computational redundancy, and ensuring that intervention plans are supported by pathological mechanisms.

[0195] 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 deep learning-based intelligent decision-making and intervention system for chronic disease health management, characterized by: include, The multi-source data acquisition module is used to collect patients' physiological indicators, genomic data, clinical information and related test data, and perform unified preprocessing and feature extraction to form standardized data input; The comorbidity reasoning module uses standardized input feature vectors to construct a heterogeneous graph network and mines the cascading effects between nodes through graph neural network algorithms. At the same time, it identifies the influence paths and potential conflict relationships between different diseases and outputs topological structure data. The virtual intervention experiment module uses topological data and standardized data input, integrates pathological correlation information and individual difference information, constructs a personalized physiological mechanism mapping function, and generates a personalized virtual image. The personalized medication recommendation module uses simulation results and patients' metabolic enzyme genotypes to construct a multi-task learning framework. It uses genetic information as a hard constraint to simultaneously predict drug efficacy and adverse reaction risks, and outputs drug dosage and combination. The dynamic decision-making intervention module generates real-time personalized health management strategies through medication dosage and combination, and pushes intervention plans through wearable devices.

2. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 1, characterized in that: The multi-source data acquisition module includes: Perform integrity identification and categorized completion processing on multi-source data, and introduce data reliability indicators; Data is aligned and standardized based on a unified time reference to form standardized data input.

3. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 2, characterized in that: The comorbidity reasoning module includes: Based on standardized data input, different disease states and physiological indicators are abstracted into nodes in a graph structure; By combining historical co-occurrence relationships and changing trends, connections between nodes are established, forming multiple types of association structures; A heterogeneous graph network is constructed based on the nodes and their relationships, and the graph neural network is used to train the heterogeneous graph. The influence relationships between nodes are updated through multiple rounds of propagation. In the updated network structure, the transmission paths and influence chains between diseases are identified, and the conflict relationships between different disease intervention paths are further identified, including mutual inhibition relationships and negative influence relationships. Based on the conflict relationships, a conflict constraint structure is constructed, and the conflict constraint structure is organized into topological data containing cooperative and conflict relationships.

4. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 3, characterized in that: The virtual intervention experiment module includes: Based on topological data, the association paths between diseases and their influence relationships at different transmission levels are extracted, and a multi-level disease association structure is formed according to path length and node connection strength. Combining data reliability identifiers and time identifiers, patient individual characteristics are stratified. The input data is divided into current status data layer, recent data layer and historical data layer according to time order, and different participation levels are set for each layer of data based on data reliability. Before performing path mapping, the candidate transmission paths in the multi-level disease association structure are screened, and the paths are prioritized according to the time location and reliability level of the data involved in the path. Higher-priority propagation pathways are selected as the primary pathways to participate in the construction of individual pathological association expression, while lower-priority propagation pathways are selected as auxiliary pathways to participate in supplementary calculations. In the process of constructing individual pathological association expression, data in the current state data layer participates in the calculation of the main action path, data in the recent data layer participates in the calculation of the auxiliary path, and data in the historical data layer participates in the calculation of the marginal path. The individual characteristics after stratification are mapped to the selected propagation paths to form individual pathological association expressions with path priority. Based on the individual pathological association expression, a physiological mechanism mapping model is constructed, so that different input data in the physiological mechanism mapping model correspond to different path ranges and order of action.

5. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 4, characterized in that: The virtual image construction includes: Based on the mapping results of individual physiological mechanisms, a virtual individual model representing individual differences among patients is constructed, and the parameters of the virtual individual model are dynamically calibrated using real data; By using calibrated virtual individual models, different intervention programs are simulated and extrapolated to obtain results of physiological state changes and generate personalized virtual images for decision support.

6. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 5, characterized in that: The personalized medication recommendation module includes: A set of candidate drug regimens is constructed based on virtual mirrors, and the action path of each candidate drug regimen in the individual physiological mechanism mapping model is simulated to obtain the position and range of action of the drug in different disease transmission paths. The path priority relationship is invoked to perform path-level filtering on the input data of candidate drug regimens, so that the current state data layer participates in the calculation of the main action path, the recent data layer participates in the calculation of the auxiliary path, and the historical data layer participates in the trend reference. Based on data reliability identification, priority participation rules are set for data participating in the main action path, and data with lower reliability is restricted from entering the critical path calculation process; Based on the conflict relationships in the topological data, path-level conflict detection is performed on the role of candidate drug regimens in different disease transmission pathways to identify the mutual inhibition and negative influence relationships of drug combinations in different pathways. Under the constraints of conflict detection results, the candidate drug schemes are screened and reconstructed at the path level, drug combinations with path conflicts are eliminated, and the remaining schemes are ranked according to the degree of path matching. Based on the screening and sorting results, a medication dosage and drug combination plan is generated; Based on the generated medication dosage and drug combination plan, the disease transmission path is marked with feedback, and the path priority is adjusted according to the conflict and matching degree in the path.

7. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 6, characterized in that: The genetic information constraints include: The metabolic enzyme genotype information is transformed into constraints that the model can recognize. These constraints are then introduced into the model prediction process to make the prediction results conform to the individual's metabolic characteristics. The prediction results are screened, and medication regimens that do not match the gene characteristics are eliminated. The remaining regimens are then optimized to determine the medication combination.

8. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 7, characterized in that: The dynamic decision intervention module includes: Initial management strategies are established by using drug treatment parameters, and real-time physiological data of patients are acquired and integrated. Based on the physiological state data, the initial management strategy is corrected and optimized in real time to generate personalized intervention instructions that are adapted to the current physiological state.

9. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 8, characterized in that: The dynamic adjustment includes: Evaluate the effectiveness of current intervention strategies based on patient feedback and identify deviations in strategy implementation; Generate strategy adjustment instructions based on the evaluation results; In response to the adjustment instruction, the current intervention strategy is iteratively optimized, and an updated intervention plan is output.

10. The intelligent decision-making and intervention system for chronic disease health management based on deep learning as described in claim 9, characterized in that: The push notification intervention scheme includes: Convert the updated intervention strategy into executable instructions; The instructions are encapsulated and sent to the user terminal via a wearable device; The terminal parses and displays the instructions, thereby providing patients with personalized health management plans.

Images