A community chronic disease auxiliary decision system

Abstract

The present application relates to a kind of community chronic disease auxiliary decision system, including knowledge warehouse end, value evaluation end and learning network end;Knowledge warehouse end provides the rule and constraint of disease treatment scheme;Value evaluation end provides the value index of evaluation treatment result, value includes treatment benefit, damage and cost;Learning network end is used for the collection and storage of patient state data, generates treatment scheme collection, obtains patient value collection, obtains the highest comprehensive value from multiple treatment schemes as decision scheme, predicts the possible treatment scheme of patient and the value generated by scheme, feedback to treatment scheme by value evaluation system, to obtain the recommended scheme order and its corresponding value.The above model is established application platform, and is tested and improved by community empirical research.The present application will provide technical tool for the standardized management of community medical institutions T2DM, and provide methodological reference for the research of other chronic disease individualized treatment auxiliary decision system.

Description

Technical Field

[0001] This invention relates to the field of medical system technology, specifically to a community chronic disease auxiliary decision-making system. Background Technology

[0002] Diabetes includes type 1 diabetes, type 2 diabetes, and gestational diabetes mellitus. Type 2 diabetes mellitus (T2DM) accounts for over 90% of all diabetes cases, and diabetes can cause serious health problems such as blindness, amputation, kidney failure, myocardial infarction, and cerebral infarction. Given the current situation in primary healthcare institutions, it is essential to explore and establish an intelligent decision support system for communities to promote the standardization, individualization, and intelligent management of T2DM at the grassroots level. Currently, various intelligent decision support systems mainly include rule-based and real-world data-based machine learning systems.

[0003] Currently, rule-based decision support systems rely on a high level of professional expertise, resulting in limited individual diagnostic capabilities. Decision support systems based on real-world data and machine learning are prone to bias due to physicians' treatment habits, missing or incorrect data entry, and the lack of quality control over retrospective data. Furthermore, both existing systems are simply machine learning models that predict treatment outcomes, requiring multiple inputs of treatment measures based on experience and comparison of predictions. This results in low levels of intelligence, and they do not consider the complexity of community-based treatment evaluation, potential harms, costs, etc. They also fail to adhere to the principles of individualized treatment, thus their applicability in communities is poor. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as low intelligence and lack of adherence to the principle of individualized treatment, by combining evidence-based rules with real-world data to establish an evidence-based rule set for type 2 diabetes mellitus (T2DM). The invention employs a patient preference-corrected stepwise weighted ratio analysis (SWARA) method to study a value assessment system for T2DM patients, predicting possible treatment options and their associated value. Feedback from the value assessment system on treatment options is then used to obtain a ranking of recommended options and their corresponding values ​​(benefits, risks, costs, and contraindications). This model will be established on an application platform and tested and refined through community-based empirical research. This invention will provide a technical tool for the standardized management of T2DM in community healthcare institutions and offer methodological guidance for research on auxiliary decision-making systems for individualized treatment of other chronic diseases.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] A community-based chronic disease decision support system includes a knowledge repository, a value evaluation system, and a learning network.

[0007] The knowledge repository collects existing literature data to provide rules and constraints for disease treatment plans to the decision support system;

[0008] The value evaluation terminal provides the auxiliary decision-making system with value indicators for evaluating treatment outcomes, including treatment benefits, harms, and costs.

[0009] The learning network is used for collecting, classifying, and storing patient status data. The patient status data enters the strategy network module of the learning network, which extracts rules and constraints from the knowledge repository to generate a set of treatment plans. The patient status data and the set of treatment plans then pass through the value network module of the learning network to obtain a set of patient values. The set of patient values ​​is evaluated by the value evaluation module and then returned to the strategy network module to generate a new set of treatment plans. After iterating through the data, the value index of each treatment plan under each status is obtained. The value index is weighted to calculate the comprehensive value of each treatment plan, and the treatment plan with the highest comprehensive value is selected as the decision plan.

[0010] Furthermore, after the decision-making scheme is used by the patient, new value indicators and new patient status are generated. The new value indicators are fed back to the value evaluation terminal, and the new patient status is fed back to the knowledge warehouse terminal, thus forming a real-time update of the knowledge of the community chronic disease auxiliary decision-making system.

[0011] Furthermore, based on the obtained rule set, a knowledge graph neural network is established, and an unsupervised learning module and a transfer learning technology module are used to build a knowledge warehouse. The transfer learning technology module sets the input of the knowledge graph neural network to patient data, enabling the neural network to learn from patient data and form a knowledge warehouse. The unsupervised learning module continuously classifies and summarizes the knowledge in the knowledge warehouse, constantly updating the knowledge by learning from patient data.

[0012] Furthermore, the knowledge repository includes a literature research module and a Delphi expert consultation module. The literature research module retrieves relevant medical guidelines or expert consensus, drug instructions, standards for potentially inappropriate drug use, clinical trials, systematic reviews, and other materials related to drug treatment for literature quality evaluation and rule extraction, thereby obtaining a preliminary set of evidence-based rules. The Delphi expert consultation module is used to evaluate the extracted rules through the Delphi method to obtain an evidence-based rule set and form a rule base.

[0013] Furthermore, the value index also includes applicability, which includes injectable and non-injectable treatment. The value index of non-injectable treatment is rated higher than that of injectable treatment. The greater the treatment benefit, the higher the value index is rated. The smaller the harm, the higher the value index is rated, including fatal harm such as acute pancreatitis and heart failure risk, and non-fatal harm such as fracture risk and weight gain. The smaller the cost, the higher the value index is rated.

[0014] Furthermore, the value evaluation module includes a value indicator determination module and a value indicator weight determination module. The value indicator determination module generates a preliminary value indicator list through literature retrieval. The preliminary value indicators include benefits, harms, costs, and applicability in treatment. Physicians, nurses, pharmacists, and patients rate the importance and necessity of the items in the preliminary value indicator list. The list is modified based on consultation opinions, and multiple rounds of consultation are conducted until the evaluation results of the survey participants are consistent. The obtained indicators serve as the initial value evaluation indicators, which can be updated and verified based on feedback from physicians and patients. The value indicator weight determination module initially ranks the importance of the value evaluation indicators, and simultaneously selects pharmacists, physicians, nurses, and patient representatives to adjust the ranking. It then assigns weights to each indicator based on its importance, and calculates the weights of the value indicators.

[0015] Furthermore, the weights of the value indicators are ranked and arranged in descending order based on the importance scores of physicians, nurses, pharmacists, and patients. The median of the ranks corresponding to each value indicator is taken as the final rank of the value indicator. Value indicator j and value indicator j+1 are listed. If value indicator j+1 is statistically significant compared to value indicator j, the relative importance value of value indicator j+1 is obtained. The relative importance values ​​form a vector S, and the number of value indicators is n.

[0016] The coefficient vector K is

[0017] ,

[0018] The transition weight vector P of the value indicator is

[0019] ,

[0020] The final weight vector of the value index is FW.

[0021] .

[0022] Furthermore, the value evaluation module also includes a patient preference module. Patient preferences are input into the value evaluation system during consultation. These preferences include acceptance of injectable dosage forms, affordability, and sensitivity to gastrointestinal adverse reactions. The weights of various value indicators are adjusted based on the patient's preference options, with the patient preference correction accounting for 30%-40% of the total weight. This results in an individualized treatment decision output. All indicators are standardized to a range of 0-1, where 0 is the worst and 1 is the best. The product of the standardized value and the weight is the comprehensive value. The comprehensive value corresponding to different treatment plans is calculated, and the treatment plans are ranked by value using a confidence interval upper bound algorithm to form a decision plan. Based on the rules and constraints in the knowledge warehouse, the predicted treatment effect, daily cost, risks, and contraindications of each possible treatment plan are output for reference.

[0023] Furthermore, the strategy network module takes the patient's state at time t as input, including examination information, demographic characteristics, complication status, and current treatment plan, and outputs a set of treatment plans for the patient from time t to time t+1; the value network module takes the patient's state at time t and the treatment plans for the patient from time t to time t+1 as input, and outputs a value index of the treatment plans for the patient at time t+1.

[0024] Since the input variables of the strategy network module and the value network module are the same, namely the state at time t, the patient state variables can be input into a shared deep unsupervised network for induction before inputting the patient state at time t into the two networks, thus reducing the computational burden on the networks. An attention mechanism is used to filter key variables, reducing the number of key variables in each learning iteration. During learning, the strategy network module and the value network module will extract experiential knowledge from the knowledge repository, thereby improving the learning speed.

[0025] Furthermore, the decision support system establishes a multi-community chronic disease decision support network system by providing at least one of the SDK interface and HTTP interface for use by community hospitals.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] In this application, evidence-based evidence and real-world evidence are organically combined through feature classification or rule constraints. This not only compensates for the shortcomings of large-sample RCT evidence in guiding dosing regimens for individual patients but also reduces the probability of erroneous recommendations that may occur during network learning. Furthermore, using evidence-based rules to classify patients by features ensures that the model's results are within the framework of guidelines and legally recognized documents, avoiding legal risks and thus possessing greater practical value for the community.

[0028] A value assessment system is proposed for evaluating treatment plans. This method addresses the problem that single outcome indicators are unsuitable as decision-making references. Furthermore, this invention innovatively incorporates patient preferences, enabling individualized and dynamic adjustments to the system, thus achieving individualized evaluation of treatment plans within the model. Attached Figure Description

[0029] Figure 1 Technical flowchart for the decision support system;

[0030] Figure 2 A graph showing the benefit evaluation indicators for patients with type 2 diabetes mellitus (T2DM);

[0031] Figure 3 The graph shows the weighting algorithm.

[0032] Figure 4 A schematic diagram of the value evaluation end for adding the patient preference module;

[0033] Figure 5 A schematic diagram of the decision support platform;

[0034] Figure 6 This is a schematic diagram of unsupervised learning.

[0035] Figure 7 This is a diagram illustrating transfer learning.

[0036] Figure 8 This is a schematic diagram of the attention mechanism. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0038] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0039] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0040] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0041] In the description of this invention, it should be noted that the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. These terms are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0042] Example 1, see Figure 1 As shown:

[0043] This embodiment provides a community-based chronic disease auxiliary decision-making system, including a knowledge repository, a value evaluation system, and a learning network.

[0044] The knowledge repository collects existing literature data and provides rules and constraints for disease treatment plans to the auxiliary decision-making system through these literature data. These rules and constraints help avoid bias risks caused by physicians' treatment habits, missing or incorrect data entry, or retrospective data that has not undergone quality control.

[0045] The value evaluation terminal provides the auxiliary decision-making system with value indicators for evaluating treatment outcomes, including treatment benefits, harms, and costs.

[0046] The learning network is used for collecting, classifying, and storing patient status data. The patient status data enters the strategy network module of the learning network, which extracts rules and constraints from the knowledge repository to generate a set of treatment plans. The patient status data and the set of treatment plans then pass through the value network module of the learning network to obtain a set of patient values. The set of patient values ​​is evaluated by the value evaluation module and then returned to the strategy network module to generate a new set of treatment plans. After iterating through the data, the value index of each treatment plan under each status is obtained. The value index is weighted to calculate the comprehensive value of each treatment plan, and the treatment plan with the highest comprehensive value is selected as the decision plan.

[0047] Furthermore, after the decision-making scheme is used by the patient, new value indicators and patient status are generated. The value indicators are fed back to the value evaluation terminal, and the patient status is fed back to the knowledge warehouse terminal, thus forming a real-time update of the knowledge of the community chronic disease auxiliary decision-making system.

[0048] Furthermore, based on the obtained rule set, a knowledge graph neural network is established, and an unsupervised learning module and a transfer learning technology module are used to build a knowledge warehouse. The transfer learning technology module sets the input of the knowledge graph neural network to patient data, enabling the neural network to learn from patient data and form a knowledge warehouse. The unsupervised learning module continuously classifies and summarizes the knowledge in the knowledge warehouse, constantly updating the knowledge by learning from patient data.

[0049] The unsupervised learning described herein utilizes a deep autoencoder to summarize and reduce the dimensionality of the input variables. The deep autoencoder first establishes a network, taking the patient's state as input, such as... Figure 6 The input layer in the network is trained through multiple deep networks to obtain a low-dimensional layer, such as... Figure 6 The feature layer in the middle is then multiplied by the transpose of the input and its weights to obtain an output that approximates the input, such as... Figure 6 The output layer in the model approximates the input. During subsequent model training, the low-dimensional layer is used as input to the model, either alone or together with the input variables. This method summarizes and reduces the dimensionality of the input, thus accelerating convergence.

[0050] The transfer learning described here simplifies network computation by leveraging existing experience. Assume the neural network is used to predict a patient's specific metric, such as blood glucose levels three months from now, and that the network has been trained and refined using massive amounts of data. See [link to relevant documentation]. Figure 7 As shown in section a; if you need to predict HbA1c values, you can delete the neurons and parameters (i.e., weights and biases) of the last layer of the trained network, see [link to documentation]. Figure 7 As shown in section b; replace with new neurons and parameters, and train to quickly predict HbA1c based on the original network's parameters. See [link to documentation]. Figure 7 As shown in section c. Similarly, during the knowledge warehouse establishment process, the neural network input is replaced to achieve transfer learning of the input.

[0051] Furthermore, the knowledge repository includes a literature research module and a Delphi expert consultation module. The literature research module retrieves relevant medical guidelines or expert consensus, drug instructions, standards for potentially inappropriate drug use, clinical trials, systematic reviews, and other materials related to drug treatment for literature quality evaluation and rule extraction, thereby obtaining a preliminary set of evidence-based rules. The Delphi expert consultation module is used to evaluate the extracted rules through the Delphi method to obtain an evidence-based rule set and form a rule base.

[0052] The system's search module retrieves databases, government agency websites, PubMed, Embase, and other resources to obtain T2DM (Treatment-Onset Diabetes Mellitus) related guidelines, criteria for potentially inappropriate medication use, drug instructions, and systematic reviews. The "AgREE II" (Academic Guideline Research and Evaluation System II) is used to evaluate the quality of guidelines and expert consensus, while the Cochrane literature quality assessment tool is used to evaluate the quality of the literature. Based on the AGREE II evaluation results, rules are extracted from guidelines with a recommendation level of "Strong Recommendation" and "Recommended." Based on the literature quality assessment results, low-quality literature is removed. Evidence-based rules are extracted to form a rule set.

[0053] The Delphi method screening module evaluates and filters the rule set: A questionnaire is created, and physicians, pharmacists, and nurses are selected as an expert panel. The questionnaires are distributed, collected, and analyzed "back-to-back" until the experts reach a consensus. The quality of the method is evaluated based on expert engagement, authority, and coordination. A new rule list is generated based on the Delphi method results, establishing the T2DM evidence-based rule set.

[0054] Further, see Figure 2 As shown, the value index also includes applicability, which includes injectable and non-injectable treatment. The value index of non-injectable treatment is rated higher than that of injectable treatment. The greater the treatment benefit, the higher the value index is rated. The smaller the harm, the higher the value index is rated. This includes fatal harm such as acute pancreatitis, heart failure risk, fracture risk, and non-fatal harm such as weight gain. The smaller the cost, the higher the value index is rated.

[0055] Furthermore, the value evaluation module includes a value indicator determination module and a value indicator weight determination module. The value indicator determination module generates a preliminary value indicator list through literature retrieval. The preliminary value indicators include benefits, harms, costs, and applicability in treatment. Physicians, nurses, pharmacists, and patients rate the importance and necessity of the items in the preliminary value indicator list. The list is modified based on consultation opinions, and multiple rounds of consultation are conducted until the evaluation results of the survey participants are consistent. The obtained indicators serve as the initial value evaluation indicators, which can be updated and verified based on feedback from physicians and patients. The value indicator weight determination module initially ranks the importance of the value evaluation indicators, and simultaneously selects pharmacists, physicians, nurses, and patient representatives to adjust the ranking. It then assigns weights to each indicator based on its importance, and calculates the weights of the value indicators.

[0056] Further, see Figure 3As shown, the weights of the value indicators are ranked and arranged in descending order based on the importance scores of physicians, nurses, pharmacists, and patients. The median of the ranks corresponding to each value indicator is taken as the final rank of the value indicator. Value indicator j and value indicator j+1 are listed. If value indicator j+1 is statistically significant compared to value indicator j, the relative importance value of value indicator j+1 is obtained. The relative importance values ​​form a vector S, and the number of value indicators is n.

[0057] The coefficient vector K is

[0058] ,

[0059] The transition weight vector P of the value indicator is

[0060] ,

[0061] The final weight vector of the value index is FW.

[0062] .

[0063] Further, see Figure 4 As shown, the value evaluation module also includes a patient preference module. Patient preferences are input into the value evaluation system during medical visits. These preferences include acceptance of injectable dosage forms, affordability of costs, and sensitivity to gastrointestinal adverse reactions. The weights of various value indicators are adjusted based on the patient's preference options, with the patient preference correction accounting for 30%-40% of the total weight. This results in an individualized treatment decision output. All indicators are standardized to a range of 0-1, where 0 is the worst and 1 is the best. The product of the standardized value and the weight is the comprehensive value. The comprehensive value corresponding to different treatment plans is calculated, and the treatment plans are ranked by value using a confidence interval upper bound algorithm to form a decision plan. Based on the rules and constraints in the knowledge warehouse, the predicted treatment effect, daily cost, risks, and contraindications of each possible treatment plan are output.

[0064] Furthermore, the strategy network module takes the patient's state at time t as input, including examination information, demographic characteristics, complication status, and current treatment plan, and outputs a set of treatment plans for the patient from time t to time t+1; the value network module takes the patient's state at time t and the treatment plans for the patient from time t to time t+1 as input, and outputs a value index of the treatment plans for the patient at time t+1.

[0065] Since the policy network module and the value network module share the same input variable—the state at time t—before the patient's state is input into the two networks, the patient's state variable can be first input into a shared deep unsupervised network for induction, reducing the computational burden on the networks. An attention mechanism is used to filter key variables, reducing the number of key variables learned each time. During learning, the policy network module and the value network module will extract experiential knowledge from the knowledge repository, improving learning speed. The attention mechanism (Attention) is an attention model that learns the importance of each element in the input vector, such as... Figure 8 As shown, the attention model is used to learn the weights w1-wn of each input variable from the input variables X1-Xn. Each input variable corresponds one-to-one with each weight. The model uses the SOFTMAX function to find the derivative and performs backpropagation to update the weights w1-wn, thereby selecting the input variables that are important for predicting the output layer results and achieving the selection of key variables.

[0066] Further, see Figure 5 As shown, the decision support system establishes SDK and HTTP interfaces for use by community hospitals, thereby creating a multi-community chronic disease decision support network system.

[0067] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A community chronic disease assistant decision system, characterized in that: The knowledge warehouse end, the value evaluation end and the learning network end are included; The knowledge warehouse end collects literature materials and provides rules and constraints of disease treatment plans; Based on the obtained rules, a knowledge graph neural network is established, and an unsupervised learning module and a transfer learning technology module are used to establish the knowledge warehouse; The transfer learning technology module sets the input of the knowledge graph neural network as patient data, so that the neural network is used for learning of the patient data to form the knowledge warehouse; The unsupervised learning module performs classification and induction of knowledge in the knowledge warehouse, and updates the knowledge by learning patient data; The value evaluation end provides value indicators for evaluating treatment results, and the value indicators include treatment benefits, damages and costs; The learning network end is used for collection and storage of patient state data, the patient state data enters a strategy network module in the learning network end, the strategy network module generates a treatment plan collection by extracting rules and constraints in the knowledge warehouse end, the patient state data and the treatment plan collection pass through a value network module in the learning network end to obtain a patient value collection, the patient value collection is evaluated by the value evaluation end and then returns to the strategy network module to generate a new treatment plan collection, after iteration of data, value indicators of each treatment plan under each state are obtained, the value indicators are weighted to calculate comprehensive values of each treatment plan, and a treatment plan with the highest comprehensive value is selected as a decision-making plan; The unsupervised learning module uses a deep auto-encoder to summarize and reduce dimensions of input variables, a network is established by the deep auto-encoder, the patient state data is taken as an input layer, a low-dimensional feature layer is obtained after training of multiple layers of deep network, the low-dimensional layer is multiplied by a transpose of input and its weight to obtain an approximate input, and in subsequent model training, the low-dimensional layer is taken as an input of the model alone or together with the input variables; The value evaluation end includes a value indicator determination module and a value indicator weight determination module, The value indicator determination module forms a preliminary value indicator list through literature retrieval, the preliminary value indicators include benefits, damages, costs and applicability in treatment, the importance and necessity of items in the preliminary value indicator list are scored by research objects, the research objects include doctors, nurses, pharmacists and patient representatives, the item list is modified according to consultation opinions, continuous consultation is performed until evaluation results of the research objects are consistent, and obtained indicators are taken as initial value indicators, and the initial value indicators can be updated and verified according to feedback of doctors and patients; The value indicator weight determination module sorts the importance of value indicators, the research objects adjust the sorting, and weights of the value indicators are assigned according to the importance of each indicator, and the weights of the value indicators are calculated. The weight calculation method of the value indicators is that all the value indicators are ranked in descending order by the importance score of the value indicators by the research object, the final rank of each value indicator is taken as the median of the corresponding rank of each value indicator, value indicator j and value indicator j+1 are listed, the relative importance value of value indicator j+1 is obtained, the relative importance value forms a vector S, the number of value indicators is n, The coefficient vector K is ， The transition weight vector P of the value indicators is ， The most weighted vector of the value index is FW, and FW is ; The decision scheme forms new value indicators and a new patient state after the patient uses it, the new value indicators are fed back to the value evaluation end, and the new patient state is fed back to the knowledge warehouse end; The value evaluation end further comprises a patient preference module, the patient preference module inputs the value evaluation system at the time of treatment, the patient preference includes the patient's acceptance of injection forms, the patient's ability to bear the cost, and the patient's sensitivity to adverse reactions of the gastrointestinal tract, the weight of each value indicator of the patient is adjusted according to the patient preference, the patient preference accounts for 30% to 40% of the total weight, an individualized treatment decision scheme output is obtained, all indicators are standardized to the interval of 0-1, wherein 0 is the worst and 1 is the best, the product of the value of the indicator after standardization and the weight is the comprehensive value, the comprehensive value corresponding to different treatment schemes is calculated, the treatment schemes are sorted in value by the upper limit algorithm of the confidence interval to form a decision scheme, and the predicted treatment effect, daily cost, risk and contraindication of all treatment schemes are output according to the rules and constraints in the knowledge warehouse end; The value indicators further include applicability, the applicability includes injection treatment and non-injection treatment, the value indicator evaluation of non-injection treatment is higher than that of injection treatment; the greater the treatment benefit, the higher the value indicator evaluation; the smaller the damage, the higher the value indicator evaluation, the damage includes fatal damage of acute pancreatitis, heart failure risk and non-fatal damage of fracture risk and weight gain; the smaller the cost, the higher the value indicator evaluation; The knowledge warehouse end comprises a literature research module and a Delphi expert consultation module, the literature research module is used for collecting literature data and extracting treatment rules to obtain a primary evidence-based rule set, and the Delphi expert consultation module is used for expert evaluation of the extracted rules by the Delphi method to obtain an evidence-based rule set and form a rule library; The input of the strategy network module is the state of the patient at time t, the state includes patient examination information, demographic characteristics, complication conditions and the current treatment scheme, and the output is a treatment scheme set of the patient from time t to time t+1; the input of the value network module is the state of the patient at time t and the treatment scheme of the patient from time t to time t+1, and the output is the value indicator of the treatment scheme of the patient at time t+1; Before the patient state data at the time t is input into the strategy network module and the value network module, the patient state data is input into a common deep unsupervised network for induction, and an attention mechanism is used for screening of key variables; The strategy network module and the value network module extract experience knowledge from the knowledge warehouse end during learning. The attention mechanism learns the importance of each element from the input vector, learns the weight of each input variable from the input variable, each input variable corresponds to each weight, and uses the SOFTMAX function to derive and back-propagate to update the weight, filters out the important input variable for the output layer result prediction, and completes the screening of the key variable.

Images