Intelligent decision-making system for nutrition metabolism collaborative management of senile chronic disease patients

Abstract

The invention relates to an intelligent decision-making system for nutrition metabolism collaborative management of elderly chronic disease patients, in particular to the field of nutrition metabolism collaborative management of elderly chronic disease, which is characterized in that drug molecule characteristics and nutrient metabolism paths are dynamically integrated through a multi-modal knowledge graph, and a cross-domain associated three-dimensional knowledge network is constructed; based on a reinforcement learning real-time optimization rule confidence threshold value, the early warning sensitivity is adaptively adjusted according to the degree that the metabolic index of the patient deviates from the safety interval; the streaming conflict detection engine accurately identifies the potential risk of asynchronously input medication and diet data, and triggers graded early warning through space-time alignment and sub-graph matching; the closed-loop evolution mechanism fuses patient compliance feedback and blood potassium change trend, drives the taboo rule base to continuously and autonomously evolve under the constraint of renal function layering, and realizes personalized risk prevention and control and metabolic state collaborative optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of old chronic disease nutrition and metabolism collaborative management, more particularly, the present application relates to an old chronic disease patient nutrition and metabolism collaborative management intelligent decision system. BACKGROUND

[0002] With the acceleration of population aging, the coexistence of old chronic diseases is becoming more and more common. Taking typical old patients as an example, they often suffer from multiple basic diseases such as diabetes, hypertension and chronic kidney disease. In the family health management scene, such patients need to take complex drug combinations such as antihypertensive drugs and hypoglycemic drugs for a long time, and adjust the dietary structure according to the nutritionist's advice to control the metabolic indicators. Due to the diversity of dietary choices in the family environment and the dynamic adjustment of the drug regimen, the patient may inadvertently ingest nutrients that have contraindications with the current medication, for example, when the patient uses angiotensin-converting enzyme inhibitor antihypertensive drugs, if at the same time a large amount of food rich in potassium elements such as bananas and spinach is eaten, the interaction of drugs and nutrients may cause electrolyte disorders. This complex scenario of multiple diseases coexisting, multiple drug combinations and variable dietary environment poses a severe challenge to the real-time, accuracy and cross-domain collaboration capabilities of nutrition and metabolism management.

[0003] The prior art mainly relies on a static rule base to realize drug-nutrient conflict detection. A typical scheme is offline matching of the DrugBank drug database and the standard nutrition database. Such a system has three inherent defects: first, the rule base has a long update cycle and cannot incorporate new drug clinical data or individualized medication adjustment information in a timely manner, resulting in lagging conflict recognition; second, the rule design is based on single disease logic and fails to establish a nutrient metabolism correlation model under cross-disease drug combination, especially ignoring the superposition effect of drug interactions when multiple chronic diseases coexist; third, the medication data in the electronic health record system and the patient's self-recorded nutrition log belong to independent modules, lacking a real-time data collaborative computing framework, making it difficult for the system to dynamically integrate the patient's latest biochemical indicators, medication records and dietary intake information. These defects collectively result in the inability of traditional systems to adapt to the individualized and highly dynamic management needs of old co-morbid patients, and there is an urgent need to build an intelligent decision-making mechanism that can integrate multi-source real-time data and support dynamic evolution of cross-domain rules. SUMMARY

[0004] The present application provides an old chronic disease patient nutrition and metabolism collaborative management intelligent decision system to solve the problems raised in the above background art by using multi-modal knowledge graph, reinforcement learning, streaming conflict detection and closed-loop rule evolution technology to realize precise nutrition and metabolism collaborative management of old chronic disease patients.

[0005] The technical scheme for solving the above technical problems is as follows: specifically comprising: a multi-modal knowledge graph construction module, a reinforcement learning rule optimization module, a streaming conflict detection engine module, and a closed-loop rule evolution module, wherein

[0006] The multi-modal knowledge graph construction module, when receiving drug chemical structure data and nutrient metabolism parameters, processes the drug simplified molecular linear input specification sequence through a graph convolution network to generate a drug molecule topological feature vector, simultaneously fuses nutrient molecule descriptors and bioavailability parameters to generate a nutrient metabolism feature vector, and constructs a dynamically updated drug-nutrient-metabolic pathway three-dimensional knowledge graph.

[0007] The reinforcement learning rule optimization module, when patient real-time medication records, dietary intake data, and blood potassium creatinine biochemical indicators are input, first constructs a multi-modal state space to describe the patient's metabolic state, then uses a dual-channel deep Q network agent to generate a confidence threshold based on an action selection strategy, and dynamically adjusts the confidence threshold of the contraindication rule by calculating a reward value according to the degree of deviation of the biochemical indicators from the safety interval.

[0008] The streaming conflict detection engine module, when the medication records and dietary logs are input in the form of streaming data, performs spatio-temporal alignment on the asynchronously arriving medication records and dietary logs, queries the drug-nutrient-metabolic pathway three-dimensional knowledge graph through subgraph matching and calculates the conflict probability, and if the conflict probability exceeds the dynamic confidence threshold, a hierarchical early warning instruction is triggered.

[0009] The closed-loop rule evolution module, when receiving the patient compliance feedback after the early warning and the blood potassium change data within 24 hours, performs feedback data fusion and performance evaluation to generate a rule performance index, updates the dual-channel deep Q network strategy parameters based on the index, and drives the autonomous evolution of the contraindication rule through the high-risk state dynamically constrained by the patient's kidney function classification.

[0010] In a preferred embodiment, in the multi-modal knowledge graph construction module, the specific operation of generating the drug molecule topological feature vector is as follows:

[0011] The drug simplified molecular linear input specification sequence is received as input data, first, an extended connectivity fingerprint algorithm is used to convert the sequence into a one thousand twenty-four dimensional binary fingerprint vector, which is used to describe the local topological environment structure of the atoms in the drug molecule; then the binary fingerprint vector is input into a graph convolution network encoder, weighted sum calculation is performed through a linear transformation matrix and a bias vector, and a linear rectifier activation function is applied for nonlinear feature extraction, and finally a one hundred twenty-eight dimensional dense feature vector is output as the drug molecule topological feature vector; wherein the dimension of the linear transformation matrix is one hundred twenty-eight rows by one thousand twenty-four columns, and the dimension of the bias vector is one hundred twenty-eight.

[0012] In a preferred embodiment, the specific operation of generating the nutrient metabolism feature vector is:

[0013] The set of nutrient molecular descriptors and the bioavailability parameters are received as input data, the molecular descriptors at least including three physicochemical properties of the lipid-water partition coefficient, the number of hydrogen bond donors, and the molecular weight; the four parameters are combined into a four-dimensional original feature vector and input into a self-encoder neural network; the self-encoder compresses the original feature vector into a 64-dimensional latent space through an encoder subnetwork, generating a fusion feature vector that fuses physicochemical properties and metabolic kinetics parameters; at the same time, the original input is reconstructed through a decoder subnetwork, and the fidelity of the feature extraction process is constrained by a mean square error loss function; finally, the 64-dimensional fusion feature vector generated by the encoder is output as the nutrient metabolism feature vector.

[0014] In a preferred embodiment, the specific operation of constructing a dynamically updated drug-nutrient-metabolic pathway three-dimensional knowledge graph is:

[0015] The drug molecular topological feature vector and the nutrient metabolism feature vector are received as input data, the metabolic pathway subgraph of the target drug and nutrient co-action is extracted from the Kyoto Encyclopedia of Genes and Genomes database, and the metabolic entity feature matrix of the subgraph is also extracted; the connection relationship between the drug nodes and the metabolic pathway entity nodes is extracted to generate a first adjacency submatrix, the connection relationship between the nutrient nodes and the metabolic pathway entity nodes is extracted to generate a second adjacency submatrix, the first adjacency submatrix and the second adjacency submatrix are spliced along the main diagonal direction of the matrix to generate a complete adjacency matrix; the adjacency matrix and the node features are input into a relational graph convolution network, the node embedding state is iteratively updated through a multi-relation message passing mechanism, and finally a drug-nutrient-metabolic pathway three-dimensional knowledge graph containing node embedding matrix and adjacency matrix is output; the relational graph convolution network distinguishes between three connection types of drug association relationship, nutrient association relationship, and metabolic entity relationship in each iteration to perform differential information aggregation;

[0016] When new drug data, new nutrient data, and metabolic research literature are input, an incremental update operation is performed:

[0017] A1, the drug molecular topological feature vector generation process is performed on the new drug data, and a new drug molecular topological feature vector is output;

[0018] A2, the nutrient metabolism feature vector generation process is performed on the new nutrient data, and a new nutrient metabolism feature vector is output;

[0019] A3, based on the metabolic research literature, a new metabolic relationship edge set is extracted, and the edge set is converted into a sparse tensor format;

[0020] A4. Combine the newly added drug molecule topological feature vectors, the newly added nutrient metabolism feature vectors, and the edge set represented by sparse tensors with the existing knowledge graph and perform a union operation on the nodes and edges. Input the relationship graph convolutional network to generate an updated three-dimensional knowledge graph of drugs, nutrients, and metabolic pathways.

[0021] In a preferred embodiment, the specific operation of constructing the multimodal state space in the reinforcement learning rule optimization module is as follows:

[0022] The system receives real-time patient medication records, dietary intake data, and serum potassium and creatinine biochemical indicators as input data. Medication records are vectorized by converting drug names into unique caloric codes and concatenating them with dose normalization values ​​to generate a medication record encoding vector. Nutrient intake data undergoes logarithmic transformation, adding small positive values ​​to generate a nutrient intake transformation vector. Individualized serum potassium safety range baselines are dynamically calculated based on patient age and estimated glomerular filtration rate, while 70% of the patient's historical mean creatinine is used as the creatinine safety threshold. Finally, the medication record encoding vector, nutrient intake transformation vector, real-time serum potassium value, and real-time creatinine value are combined into a quaternion state vector.

[0023] In a preferred embodiment, the specific operation of processing the state vector using the dual-channel deep Q-network architecture is as follows: the first channel outputs a confidence threshold adjustment amount, which is a continuous value in the range of 0.80 to 0.99; the second channel outputs a state value score; and a comprehensive Q value is generated by merging the state value score and the action advantage score.

[0024] The action selection strategy employs an exploration-exploitation balance mechanism, specifically: receiving the comprehensive Q-value output from the dual-channel deep Q-network, subtracting the exploration rate from a probability of 1 to select a confidence threshold adjustment amount that maximizes the comprehensive Q-value; uniformly and randomly selecting a confidence threshold adjustment amount within the range of 0.80 to 0.99 with a probability of exploration rate; and directly assigning the selected adjustment amount as the dynamic confidence threshold; wherein the initial value of the exploration rate is 0.3 and it decays linearly over time.

[0025] The specific operations for reward calculation and network update are as follows:

[0026] The immediate reward is calculated based on the serum potassium and creatinine values ​​in the state vector at the next moment. The serum potassium deviation penalty is calculated by compressing the absolute deviation between the actual serum potassium value and the individualized safety baseline using a hyperbolic tangent function, and then multiplying it by a metabolic sensitivity weighting coefficient. The creatinine deviation penalty is calculated by dividing the absolute deviation between the actual creatinine value and the creatinine safety threshold baseline by the patient's baseline creatinine value and then multiplying it by a second weighting coefficient. The two penalty terms are negatively summed and then superimposed on the safety state reward term to generate the total reward value.

[0027] The empirical priority is calculated using time-series differential error, specifically as follows:

[0028] B1, calculate the time difference error value: according to the difference between the current state action value function and the next state maximum action value function output by the target network, superimpose the reward value and take the absolute value to generate the time difference error value;

[0029] B2, assign priority sampling probability: normalize the time difference error value raised to the power of zero point six to generate the sampling weight of each experience tuple;

[0030] Gradually synchronize network parameters through target network soft update strategy: update the target network parameters at a rate of zero point one, and retain the historical target network parameters at a rate of zero point nine nine, wherein the value network parameters and the advantage network parameters are updated independently.

[0031] In a preferred embodiment, the specific operation of spatiotemporal alignment of asynchronously arriving medication records and meal logs in the stream conflict detection engine module is as follows:

[0032] C1, receive medication record stream and meal log stream as input data, extract drug name and dose value, generate drug name semantic vector through pre-trained drug encoder, generate normalized dose scalar by dividing dose value by maximum safe dose of the drug, and concatenate the two to generate drug semantic dose vector;

[0033] C2, extract food material name and weight data, query the nutrient database to obtain the nutrient standard vector of each food material, weight sum and normalize to generate the food ingredient aggregation vector;

[0034] C3, take the medication time as the center point, extract the meal data within one hour before medication to two hours after medication, concatenate the drug semantic dose vector and all food ingredient aggregation vectors within the window into a spatiotemporal alignment matrix.

[0035] In a preferred embodiment, the specific operation of querying the drug-nutrient-metabolic pathway three-dimensional knowledge graph and calculating the conflict probability through subgraph matching is as follows:

[0036] The spatio-temporal alignment matrix is received as input data, and the drug node is taken as the root to expand the two-hop range along the metabolic pathway edge in the drug-nutrient-metabolic pathway three-dimensional knowledge graph, and the subgraph structure containing the associated nutrient nodes is extracted; the adjacency relationship between the drug nodes, nutrient nodes and metabolic entity nodes in the subgraph is extracted, classified into three types of relationships: drug-enzyme interaction, nutrient-enzyme interaction and enzyme-enzyme cascade reaction, stacked into a three-dimensional heterogeneous adjacency tensor along the third dimension, and the node feature matrix of the subgraph is aligned by a multi-layer perception machine; the relationship-aware graph convolutional network is used to aggregate neighbor information by relationship type, and iteratively update the node embedding state; the tensor outer product of the drug molecule topological feature vector and all nutrient metabolic feature vectors is calculated to generate a high-dimensional interaction tensor;

[0037] The interaction tensor is input into a multi-layer perception machine to generate a conflict probability value through an S-shaped function mapping; if the conflict probability exceeds the dynamic confidence threshold, a hierarchical early warning instruction is triggered according to the probability interval, and the hierarchical early warning instruction is specifically: a probability value of 0.8 to 0.89 corresponds to a first-level warning, a probability value of 0.9 to 0.94 corresponds to a second-level warning, and a probability value of 0.95 and above corresponds to a third-level warning.

[0038] In a preferred embodiment, the specific operation of the closed-loop rule evolution module for performing feedback data fusion and performance evaluation to generate a rule performance index is: receiving patient compliance labels and blood potassium concentration time series data, verifying the effectiveness of the compliance labels and removing outliers; constructing a four-dimensional feedback feature vector, including compliance state, absolute change amount of blood potassium concentration, current confidence threshold and blood potassium change direction indicator value; calculating the rule performance index based on the four-dimensional feedback feature vector and the historical conflict probability value, wherein the clinical performance component is composed of the product of the compliance state, the saturation compression amount of the blood potassium change and the significant improvement reward item, and the stability penalty component is the KL divergence value of the current conflict probability value and the historical conflict probability value; the rule performance index value is generated by dividing the clinical performance component by the confidence threshold and adding the stability penalty item.

[0039] In a preferred embodiment, the specific operation of the high-risk state constrained rule generation and stability verification includes: calculating the strategy gradient expected value with the rule performance index as the weight, and preferentially sampling the blood potassium concentration abnormal high-risk state; performing difference chemical rate update on the value network and the advantage network, and the advantage network uses a learning rate twice that of the value network;

[0040] The specific definition of the high-risk state constraint is:

[0041] When the estimated glomerular filtration rate of the patient is less than 30, the high-risk blood potassium interval is blood potassium concentration greater than or equal to 3.0 mmol / L to less than or equal to 5.5 mmol / L;

[0042] When the patient's estimated glomerular filtration rate is between thirty and sixty (inclusive), the high-risk blood potassium range is greater than or equal to three point three millimoles per liter to less than or equal to five point two millimoles per liter;

[0043] When the patient's estimated glomerular filtration rate is greater than or equal to sixty, the high-risk blood potassium range is greater than or equal to three point five millimoles per liter to less than or equal to five point zero millimoles per liter;

[0044] Generate new contraindication rules based on updated network parameters, and only generate rules for high-risk states where blood potassium concentration exceeds the safety threshold;

[0045] Evaluate the compatibility of new rules with the historical rule library through the rule mutation stability scoring mechanism, and if the score exceeds zero point eight, the new rule will be incorporated into the rule library.

[0046] The beneficial effects of the present application are: through the dynamic integration of drug molecular characteristics and nutrient metabolism pathways by multi-modal knowledge graph, a three-dimensional knowledge network with cross-domain association is constructed; based on reinforcement learning, the rule confidence threshold is optimized in real time, and the early warning sensitivity is adaptively adjusted according to the degree of deviation of patient metabolic indicators from the safety interval; the flow conflict detection engine accurately identifies the potential risks of asynchronous input of medication and dietary data, and triggers hierarchical early warning through spatio-temporal alignment and subgraph matching; the closed-loop evolution mechanism integrates patient compliance feedback and blood potassium change trend, and drives the contraindication rule library to continuously and autonomously evolve under the constraint of kidney function stratification, realizing personalized risk prevention and control and metabolic state optimization. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The method flowchart of the present application;

[0048] Figure 2 The system structure block diagram of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] In the description of the present application, the terms "first", "second" are only used for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0051] In the description of the present application, the term "for example" is used to indicate "serving as an example, instance, or illustration." Any embodiment described as "for example" in this application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to practice the application as claimed. In the following description, for purposes of explanation, specific details are set forth to provide a thorough understanding of the present application. It will be apparent to one skilled in the art, however, that the present application can be practiced without using these specific details. In other instances, well-known structures and processes are not elaborated in detail in order not to obscure the description of the present application with unnecessary detail. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0052] Embodiment 1

[0053] The present embodiment provides an intelligent decision-making system for the coordinated management of nutrition and metabolism of elderly patients with chronic diseases, as shown in Figures 1-2 The present embodiment provides an intelligent decision-making system for the coordinated management of nutrition and metabolism of elderly patients with chronic diseases, as shown in

[0054] The multi-modal knowledge graph construction module, when receiving drug chemical structure data and nutrient metabolism parameters, processes the drug simplified molecular linear input specification sequence through a graph convolution network to generate a drug molecule topological feature vector, and simultaneously fuses nutrient molecule descriptors and bioavailability parameters to generate a nutrient metabolism feature vector, thereby constructing a dynamically updated drug-nutrient-metabolic pathway three-dimensional knowledge graph;

[0055] The reinforcement learning rule optimization module, when patient real-time medication records, dietary intake data, and blood potassium creatinine biochemical indicators are input, first constructs a multi-modal state space to describe the patient's metabolic state, then uses a dual-channel deep Q network agent to generate a confidence threshold based on an action selection strategy, and updates the network through a reward mechanism to calculate a reward value based on the degree of biochemical indicator deviation from the safe interval, thereby dynamically adjusting the confidence threshold of the contraindication rule;

[0056] The streaming conflict detection engine module, when medication records and dietary logs are input in the form of streaming data, performs spatio-temporal alignment on asynchronously arriving medication records and dietary logs, queries the drug-nutrient-metabolic pathway three-dimensional knowledge graph through subgraph matching and calculates the conflict probability, and if the conflict probability exceeds the dynamic confidence threshold, triggers a hierarchical early warning instruction;

[0057] The closed-loop rule evolution module, when receiving the patient compliance feedback after the early warning and the blood potassium change data within 24 hours, performs feedback data fusion and performance evaluation to generate a rule performance index, updates the policy parameters of the double-channel deep Q network based on the index, and evolves the contraindication rule autonomously driven by the high-risk state of the patient's renal function classification dynamic constraint.

[0058] In this embodiment, it is specifically required to explain that in the multi-modal knowledge graph construction module, the specific operation of generating the drug molecule topological feature vector is as follows:

[0059] The drug simplified molecular linear input canonical sequence is received as input data, and first, the extended connectivity fingerprint algorithm is used to convert the sequence into a 1024-dimensional binary fingerprint vector, and the conversion formula of the extended connectivity fingerprint algorithm is as follows:

[0060] b drug = ECFP4(SMILES);

[0061] Wherein, SMILES represents the drug simplified molecular linear input canonical sequence, a standard text format describing the structure of the drug molecule, ECFP4(·) represents the extended connectivity fingerprint algorithm, which is used to convert the SMILES sequence into a fixed-length binary vector, b drug represents the drug molecule ECFP fingerprint vector (i.e. a 1024-dimensional binary fingerprint vector), which represents the local topological environment structure of the atoms in the drug molecule, and is used as the input of the graph convolution network encoder; then the binary fingerprint vector is input into the graph convolution network encoder, weighted summation calculation is performed through a linear transformation matrix and a bias vector, and a linear rectifier activation function is applied for nonlinear feature extraction, and the extraction formula of the linear rectifier activation function is as follows:

[0062]

[0063] Wherein, W enc represents the weight matrix (128x1024-dimensional real matrix), which is used for linear transformation of the input vector, b enc represents the bias vector (128-dimensional real vector), which is used to increase the fitting ability of the graph convolution network encoder, ReLU(·) represents the linear rectifier activation function, which is used to introduce the nonlinear feature extraction capability, represents the drug molecule topological feature vector (128-dimensional real vector), which is the intermediate feature representation extracted by the graph convolution network encoder, and is used as the drug node embedding feature of the drug-nutrient-metabolic pathway three-dimensional knowledge graph, and finally outputs a 128-dimensional dense feature vector as the drug molecule topological feature vector v drug ; wherein the dimension of the linear transformation matrix is one hundred and twenty-eight rows by one thousand and twenty-four columns, and the dimension of the bias vector is one hundred and twenty-eight;

[0064] The specific operation of generating the nutrient metabolism feature vector is:

[0065] The nutrient molecular descriptor set and the bioavailability parameter (F ∈ [0, 1]) are received as input data, and the molecular descriptor D nut At least three physicochemical properties including the lipid-water partition coefficient LogP, the number of hydrogen bond donors HBD, and the molecular weight MW are included; the four parameters are combined into a four-dimensional original feature vector and input into the autoencoder neural network, and the expression of the four-dimensional original feature vector is:

[0066] M raw =[LogP,HBD,MW,F] Ta ;

[0067] Wherein, M raw represents the four-dimensional original feature vector (4-dimensional column vector) of the nutrient, which is used to integrate the physicochemical properties and metabolic parameters and provide input for dimension reduction, LogP represents the lipid-water partition coefficient (characterizing lipophilicity), HBD represents the number of hydrogen bond donors (affecting solubility), MW represents the molecular weight (determining the metabolic rate), F represents the bioavailability (value range [0, 1], representing the absorption efficiency), and Ta represents the matrix transposition operation; the autoencoder compresses the original feature vector to a sixty-four-dimensional latent space through the encoder subnetwork, generates a fusion feature vector that fuses the physicochemical properties and metabolic kinetic parameters, and the dimension reduction formula of the autoencoder is:

[0068] v nut =Encoder AE (M raw ;θ AE );

[0069] Wherein, v nut represents the fusion feature vector that fuses the physicochemical properties and metabolic kinetic parameters (sixty-four-dimensional), which characterizes the high-order interaction characteristics of the nutrient in the metabolic pathway and is used for knowledge graph correlation modeling, Encoder AE (·) represents the encoder subnetwork of the autoencoder, θ AE represents the parameter matrix (including weights and biases) of the autoencoder, which controls the feature compression and reconstruction process and ensures that the key metabolic properties are retained after dimension reduction; meanwhile, the original input is reconstructed through the decoder subnetwork, and the fidelity of the feature extraction process is constrained by the mean square error loss function, and the expression of the mean square error loss function is:

[0070] L AE =||M raw -Decoder AE (v nut )|| 2 ;

[0071] Wherein, LAE The mean squared error loss function of the autoencoder constrains encoder-decoder learning and prevents feature distortion (such as loss of bioavailability F information). AE (·) denotes the decoder subnetwork of the autoencoder, ||·|| 2 The L2 norm squared (measures the reconstruction error); the final output is a 64-dimensional fused feature vector generated by the encoder, which serves as the nutrient metabolism feature vector v. nut ;

[0072] The specific steps for constructing a dynamically updated three-dimensional knowledge graph of drugs, nutrients, and metabolic pathways are as follows:

[0073] The system receives drug molecule topological feature vectors and nutrient metabolism feature vectors as input data, and extracts metabolic pathway subgraphs (G) of the target drug and nutrient interaction from the Kyoto Encyclopedia of Genetics and Genomes database. metab = (Va, Ea), where Va represents the set of nodes in the metabolic pathway subgraph (i.e., the number of metabolic pathway entity nodes), and Ea represents the set of edges in the metabolic subgraph (i.e., edges represent the biochemical reaction relationships between nodes). Simultaneously, the metabolic entity feature matrix H of this subgraph is extracted. kegg It includes original biological characteristics such as enzyme catalytic activity and receptor binding sites; the connection relationship between drug nodes and metabolic pathway entity nodes is extracted to generate a first adjacency sub-matrix, and the connection relationship between nutrient nodes and metabolic pathway entity nodes is extracted to generate a second adjacency sub-matrix. The first and second adjacency sub-matrixes are concatenated along the main diagonal of the matrix to generate a complete adjacency matrix, expressed as:

[0074]

[0075] Among them, A drug-path Let A represent the first adjacency matrix between drug nodes and metabolic pathways. nut-path Let A represent the second adjacency submatrix between nutrient nodes and metabolic pathways, and let Aw represent the complete adjacency matrix. The adjacency matrix and node features are input into a relational graph convolutional network. The node embedding state is iteratively updated through a multi-relationship message passing mechanism, ultimately outputting a three-dimensional knowledge graph of drugs, nutrients, and metabolic pathways containing both the node embedding matrix and the adjacency matrix. H (L) (Represents the embedding matrix of the last layer nodes); In each iteration, the relational graph convolutional network distinguishes three connection types—drug associations, nutrient associations, and relationships between metabolic entities—and aggregates differentiated information. The expression for updating the node embedding state of the relational graph convolutional network is:

[0076]

[0077] in, denotes the set of relation types (drug association, nutrient association, metabolic entity association) used to define the message passing rules for different types of edges in the heterogeneous graph, denotes the inverse matrix of the degree matrix of relation r N denotes the total number of nodes in the drug-nutrient-metabolic pathway three-dimensional knowledge graph, and the calculation method is N = ||Va|| + 2, Va denotes the number of metabolic pathway entity nodes, and + 2 represents drug nodes and nutrient nodes), and the adjacency matrix Aw r is normalized to prevent information propagation bias caused by node degree difference, Aw r denotes the adjacency matrix of relation r (Aw r ∈{0, 1} N×N (sparse matrix)) describes the connection structure between nodes under a specific relation r (such as drug-enzyme interaction), H (l) denotes the node embedding matrix of the l-th layer (H

[0078] N denotes the total number of nodes in the drug-nutrient-metabolic pathway three-dimensional knowledge graph, and d denotes the feature vector dimension (the drug molecular topological feature vector mentioned in the foregoing is d = 128, the nutrient metabolic feature vector is d = 64, and the feature dimension of the metabolic entity node extracted from the Kyoto Encyclopedia of Genes and Genomes database is d = 64 by default. Different node features are unified to d dimensions by zero padding, such as d = 128), which stores the feature representation of all nodes in the current layer. In addition, H (l) is an intermediate variable calculated iteratively, and the initial value H (0) is constructed from the drug molecular topological feature vector and the nutrient metabolic feature vector (i.e. the initial node feature of the graph), denotes the trainable weight matrix of relation r in the l-th layer (W d, de respectively denote the input and output feature dimensions (strictly identical to the d of H (l) )), which performs differential feature transformation on different relation types, and σw(·) denotes a nonlinear activation function (such as ReLU), H (l+1) denotes the node embedding matrix of the l+1-th layer (H (N is the number of nodes, and d is the feature dimension)), and the updated node feature representation;

[0079] When new drug data De new , new nutrient data Qe new and metabolic research literature We new are input, incremental update operation is performed:

[0080] A1, the drug molecular topological feature vector generation process is performed on the new drug data De new , and the new drug molecular topological feature vector v drug-new is output;

[0081] A2, new nutrient data Qe new Execute the nutrient metabolism feature vector generation process to output the new nutrient metabolism feature vector v nut-new ;

[0082] A3, based on metabolic research literature We new Extract the new metabolic relationship edge set, and convert the edge set into a sparse tensor format, with the expression being:

[0083] ΔA = SpareTensor(Ea new );

[0084] Wherein, Ea new represents the new drug-nutrient-metabolic pathway association extracted from We new , and ΔA represents the sparse tensor;

[0085] A4, perform the union operation of the new drug molecule topological feature vector, the new nutrient metabolism feature vector, and the edge set represented by the sparse tensor, and the existing knowledge graph to input the relationship graph convolution network to generate the updated drug-nutrient-metabolic pathway three-dimensional knowledge graph, and the expression of the updated drug-nutrient-metabolic pathway three-dimensional knowledge graph is:

[0086]

[0087] Wherein, represents the updated drug-nutrient-metabolic pathway three-dimensional knowledge graph, which is output by the relationship graph convolution network by fusing the new nodes and edges, and outputs the dynamically evolving knowledge graph, represents the original drug-nutrient-metabolic pathway three-dimensional knowledge graph, represents the union operation of the graph nodes and edges, which merges the existing drug-nutrient-metabolic pathway three-dimensional knowledge graph and the new nodes / edges (not simply superimposed, and the topological consistency is preserved).

[0088] In this embodiment, it is specifically necessary to explain that the specific operation of the multi-modal state space construction in the reinforcement learning rule optimization module is:

[0089] Receive patient real-time medication records, dietary intake data, and blood potassium creatinine biochemical indicators as input data; vectorize the medication records, convert the drug names into one-hot encoding vectors, and splice them with the dose normalized values to generate medication record encoding vectors; perform logarithmic transformation on the nutrient intake data, add a small positive value to avoid zero value anomalies, and generate nutrient intake transformation vectors; combine the patient's age and estimated glomerular filtration rate to dynamically calculate the individualized blood potassium safety interval reference value, and the calculation formula of the individualized blood potassium safety interval reference value is:

[0090] K safe =f age (K std ,eGFR);

[0091] wherein K safe represents the individualized potassium safety range reference value, the safety range dynamically adjusted according to age and kidney function, K std represents the standard potassium safety value, i.e. the potassium reference range of healthy population (usually 3.5-5.0 mmol / L), eGFR represents the estimated glomerular filtration rate, i.e. the key indicator for assessing kidney function (unit: mL / min / 1.73m 2 ), f age represents the medical rule function for dynamically adjusting the potassium safety range according to the kidney function of the patient, and the expression of the creatinine safety threshold reference is:

[0092] C safe =0.7×Cr base ;

[0093] wherein Cr safe represents the creatinine safety threshold reference, the warning line set based on the baseline creatinine of the patient, Cr base represents the baseline creatinine value, the average level of historical creatinine of the patient (reflecting the basic state of kidney function); finally, the medication record coding vector, the nutrient intake amount transformation vector, the real-time potassium value and the real-time creatinine value are combined into a four-tuple state vector, and the expression of the state vector is:

[0094]

[0095] wherein S t represents the state vector, used for integrating the multi-modal features of medication, nutrition, potassium and creatinine, φ(D t ) represents the medication record coding vector, used for converting discrete drug information into continuous features (One-hot + dose normalization), D t represents the real-time medication record, recording the names and doses of the drugs currently taken by the patient, ψ(N t ) represents the nutrient transformation vector, log-compressed nutrient intake amount (solving the problem of long-tail distribution), and the calculation method is ψ(N t ) = N t + εe, εe represents a small positive value (its value is set to 10 -5 ), avoiding abnormal log calculation when N t = 0, N t represents the dietary intake data, recording the types and contents of nutrients currently taken by the patient, represents the real-time potassium value, i.e. the current potassium concentration of the patient (unit: mmol / L), Cr t represents the real-time creatinine value, i.e. the current creatinine concentration of the patient (unit: mg / dL);

[0096] The specific operation of processing the state vector by using the dual-channel deep Q network architecture is as follows: the first channel outputs a confidence threshold adjustment amount (a t ∈[0.80, 0.99]), which is a continuous value in the interval of zero point eight zero to zero point nine nine; the second channel outputs a state value score V(S t ), representing the overall risk level of the current metabolic state; the comprehensive Q value is generated by merging the state value score and the action advantage score, and the formula of the comprehensive Q value is as follows:

[0097] Q(S t ,a t ; θw) = Vn(S t ; θw v ) + An(S t ,a t ; θw a );

[0098] wherein Q(S t ,a t ; θw) represents the current Q value, S t represents the state vector, a t represents the confidence threshold adjustment amount, and θw represents the total parameter set of the DQN (the set of the value network θw v and the advantage network θw a parameters), Vn(S t ; θw v ) represents the state value score (i.e. the overall risk evaluation value of the metabolic state), and An(S t ,a t ; θw a ) represents the action advantage score (the confidence threshold adjustment amount output by the first channel);

[0099] The action selection strategy adopts an exploration and utilization balance mechanism, specifically: receiving the comprehensive Q value output by the dual-channel deep Q network, selecting the confidence threshold adjustment amount that maximizes the comprehensive Q value with a probability of one minus the exploration rate; uniformly and randomly selecting the confidence threshold adjustment amount from the interval of zero point eight zero to zero point nine nine with a probability of the exploration rate; directly assigning the selected adjustment amount as the dynamic confidence threshold (τ t =a t , τ t represents the dynamic confidence threshold); wherein the initial value of the exploration rate is zero point three and linearly decays over time, and the expression of the action selection strategy is as follows:

[0100]

[0101] where εnis the exploration rate, which is initially set to 0.3, U(0.80, 0.99) represents uniform random sampling, and argmax a Q(S t ,a t ;θw) represents the action that maximizes the Q value;

[0102] The specific operations of reward calculation and network update are as follows:

[0103] The immediate reward is calculated based on the potassium value and creatinine value in the next time state vector; the absolute deviation amount of the actual potassium value from the individualized safety benchmark is compressed by using a hyperbolic tangent function in the potassium deviation penalty term calculation, and then multiplied by a metabolic sensitivity weight coefficient, and the calculation formula of the potassium deviation penalty term is:

[0104]

[0105] where K represents the potassium value at the next time, K safe represents the individualized potassium safety benchmark, δ K represents the potassium fluctuation tolerance coefficient, tanh(·) represents the hyperbolic tangent function, and αwrepresents the metabolic sensitivity weight, which is initially set to 2.0, P K represents the potassium deviation penalty term; the absolute deviation amount of the actual creatinine value from the creatinine safety threshold benchmark is divided by the baseline creatinine value of the patient in the creatinine deviation penalty term calculation, and then multiplied by a second weight coefficient, and the calculation formula of the creatinine deviation penalty term is:

[0106]

[0107] where Cr t+1 represents the creatinine value at the next time, Cr safe represents the creatinine safety threshold benchmark, which is a warning line based on the baseline creatinine of the patient, Cr base represents the baseline creatinine value, which is the historical average creatinine level of the patient (reflecting the basic state of kidney function), and βwrepresents the second weight coefficient, which is initially set to 1.5, P Cr represents the creatinine deviation penalty term; the negative sum of the two penalty terms is added to the safety state reward term to generate a total reward value, and the calculation formula of the total reward value is:

[0108] R t =-(P K +P Cr )+γw·Ip safe ;

[0109] where R t represents the total reward value, γwrepresents the safety state reward coefficient, which is set to 0.5, I saferepresents a safety state indicator function, when the value of I safe is set to 1;

[0110] The experience priority is calculated using the timing difference error, specifically:

[0111] B1, calculate the timing difference error value: according to the difference between the current state-action value function and the maximum action value function of the next state output by the target network, superimpose the reward value and take the absolute value to generate the timing difference error value, the calculation formula of the timing difference error is:

[0112] δ t =||R t +γomax a′ Q(S t+1 , a'; θw - )-Q(S t , a t ; θw)||;

[0113] Wherein, δ t represents the timing difference error, quantifying the deviation of the current Q value prediction, R t represents the total reward value, γo represents the discount factor, used to balance the importance of current and future rewards, max a′ Q(S t+1 , a'; θw - ) represents the target Q value (calculating the maximum Q value of the next state through the target network), Q(S t , a t ; θw) represents the current Q value, evaluating the value of the current state-action pair, θw represents the total parameter set of DQN, θw - represents the target network parameter;

[0114] B2, assign priority sampling probability: normalize the timing difference error value raised to the power of zero point six to generate the sampling weight of each experience tuple, the expression of the priority sampling probability is:

[0115]

[0116] Wherein, p i represents the priority sampling probability (generating the sampling weight of each experience tuple), δ i represents the TD-error absolute value (timing difference error value) of the i-th experience tuple, ωe represents the TD-error power weight, controlling the degree of sampling bias to high error samples, and its value is set to 0.6;

[0117] Gradually synchronize network parameters through the target network soft update strategy: update the target network parameters at a rate of zero point zero one, and retain the historical target network parameters at a rate of zero point nine nine, the expression of the target network soft update is:

[0118] θw - ←τwθw+(1-τw)θw - ;

[0119] where τw represents the soft update ratio (τw=0.01), controlling the target network parameter synchronization speed, and θw represents the DQN total parameter set (value network θw v and advantage network θw a parameters are updated independently. - θw represents the target network parameter, which is a stable parameter for calculating the target Q value (0.01% updated every step), where the value network parameter and the advantage network parameter are updated independently.

[0120] In this embodiment, it is specifically required to explain the specific operation of spatiotemporal alignment of asynchronously arriving medication records and meal logs in the streaming conflict detection engine module:

[0121] C1, receiving medication record stream and meal log stream as input data, extracting drug name and dose value, generating drug name semantic vector through pre-trained drug encoder, generating normalized dose scalar by dividing dose value by maximum safe dose of the drug, and splicing the two to generate drug semantic dose vector, the expression of drug semantic dose vector is:

[0122]

[0123] where d y represents the drug name, the drug identification (such as "aspirin") in the medication record, dose y represents the dose value, the drug taking amount (unit: mg / g), maxdose represents the maximum safe dose, the maximum single dose allowed in clinical practice of the drug (predefined constant), BERT drug (·) represents a pre-trained drug encoder for generating a semantic vector (768 dimensions) of the drug name, drug semantic dose vector, the final output drug feature vector (769 dimensions);

[0124] C2, extracting food material name f u and weight data ωm u , querying the nutrition database to obtain the nutrient standard vector of each food material, and generating the food material component aggregation vector by weighted summation and normalization according to the weight proportion, the expression of the food material component aggregation vector is:

[0125]

[0126] where, food material component aggregation vector, the final output nutrient feature vector (300 dimensions), This indicates L2 norm normalization, used to eliminate differences in the dimensions of different nutrients (e.g., vitamin μg vs. mineral mg). Let f represent the nutrient baseline vector, and f represent the food ingredient. u The standard content vector of the kth nutrient (e.g., potassium content 120mg / 100g), ωm u This indicates weight data, representing food intake (unit: g);

[0127] C3. Taking medication time as the baseline center point t d The time window from one hour before medication to two hours after medication (t) is captured. d -1h,t d Dietary data (+2h) was used to concatenate the drug semantic dose vector with the aggregated vectors of all food components within the window to form a spatiotemporal alignment matrix. M align This represents the spatiotemporal alignment matrix, where me represents the number of food items within the time window from one hour before medication to two hours after medication, and Ta represents the matrix transpose operation. (Represents the set of aggregated vectors of all food ingredients);

[0128] The specific steps for querying the three-dimensional knowledge graph of drugs, nutrients, and metabolic pathways by subgraph matching and calculating the probability of conflict are as follows:

[0129] The spatiotemporal alignment matrix is ​​received as input data. Using the drug node as the root, a two-hop range is expanded along the metabolic pathway edge in the three-dimensional knowledge graph of drug-nutrient-metabolic pathway. A subgraph structure containing associated nutrient nodes is extracted. The subgraph query formula is:

[0130] G sub ={v∈Va|dis(v,d)} b )≤2}∪{associated nutrient nodes};

[0131] Among them, G sub Denotes a subgraph containing a set of nodes, including drug nodes, metabolic entities within 2 hops, and associated nutrient nodes, d. b This represents a drug node, the corresponding drug-nutrient-metabolic pathway 3D knowledge graph node (subgraph query root node) for medication records, dis(v,d) b The ) represents the distance between nodes, i.e., the distance between node v and drug node d. b The shortest metabolic pathway hop count, ranging from {0, 1, 2} (limited to 2 hops), where Va represents the set of nodes in the metabolic pathway subgraph, i.e., the number of metabolic pathway entity nodes. Adjacency relationships between drug nodes, nutrient nodes, and metabolic entity nodes within the subgraph are extracted and categorized into three types: drug-enzyme interaction, nutrient-enzyme interaction, and enzyme-enzyme cascade reaction. These relationships are then stacked along the third dimension to form a three-dimensional heterogeneous adjacency tensor. The expression for the three-dimensional heterogeneous adjacency tensor is:

[0132]

[0133] wherein, A sub represents the three-dimensional heterogeneous adjacency tensor, i.e., the adjacency representation of the three types of relationships within the subgraph (drug-enzyme / nutrient-enzyme / enzyme-enzyme), and Awrepresents the complete adjacency matrix, and the expression of the aligned subgraph node feature matrix is:

[0134]

[0135] wherein, H align represents the aligned subgraph node feature matrix, which is the subgraph node representation that fuses the drug and nutrient features, and MLP(·) represents a multi-layer perceptron, represents the vector concatenation operation, v drug represents the drug molecule topological feature vector, v nut represents the nutrient metabolism feature vector, H kegg represents the metabolic entity feature matrix, which contains original biological features such as enzyme catalytic activity and receptor binding site; the relationship-aware graph convolutional network is used to aggregate neighbor information by relationship type, and iteratively update the node embedding state, and the relationship-aware graph convolution formula is:

[0136]

[0137] wherein, Z (l) represents the subgraph node embedding matrix (l is the number of graph convolution layers), and Z (0) = H align , A sub represents the three-dimensional heterogeneous adjacency tensor, represents the relationship convolution weight dedicated to conflict detection in the streaming conflict detection engine module (r = 1, 2, 3 corresponds to the drug-enzyme / nutrient-enzyme / enzyme-enzyme relationship), and in addition, it is independently trained in the streaming conflict detection engine module (not shared with the R-GCN weight of the multi-modal knowledge graph construction module), and σw(·) represents a nonlinear activation function (such as ReLU), Z (l+1) represents the iteratively updated node embedding state; the tensor outer product of the drug molecule topological feature vector and all nutrient metabolism feature vectors is calculated to generate a high-dimensional interaction tensor;

[0138] The interaction tensor is input into a multi-layer perceptron to generate a conflict probability value through a sigmoid function, and the generation formula of the conflict probability value is:

[0139]

[0140] P conflict = sigmoid(MLP(flatten(PE)));

[0141] wherein PE represents a high-dimensional interaction tensor, P conflict represents a conflict probability value, v drug represents a drug molecule topological feature vector, v nut represents a nutrient metabolism feature vector; if the conflict probability exceeds a dynamic confidence threshold, a hierarchical early warning instruction is triggered according to a probability interval, and the hierarchical early warning instruction is specifically: a probability value of 0.80 to 0.89 corresponds to a first-level early warning, 0.90 to 0.94 corresponds to a second-level early warning, and 0.95 and above corresponds to a third-level early warning.

[0142] In the present embodiment, it is specifically required to explain that in the closed-loop rule evolution module, the specific operation of executing feedback data fusion and performance evaluation to generate a rule performance index is: receiving a patient compliance label y comply ∈{0,1} and a blood potassium concentration time series data, verifying the effectiveness of the compliance label and eliminating outliers; constructing a four-dimensional feedback feature vector containing the compliance state, the absolute change amount of blood potassium concentration, the current confidence threshold τ t and the blood potassium change direction indication value, and the calculation formula of the absolute change amount of blood potassium concentration is:

[0143]

[0144] wherein, represents the blood potassium concentration value at the early warning moment, represents the blood potassium concentration value after 24 hours, ΔK + represents the absolute change amount of blood potassium concentration;

[0145] The expression of the four-dimensional feedback feature vector is:

[0146]

[0147] wherein FY represents a four-dimensional feedback feature vector, represents the blood potassium change direction indication value, i.e. the change direction of the final blood potassium value relative to the safety value (rise = 1, fall = -1, unchanged = 0), K safe represents the individualized blood potassium safety interval reference value, i.e. the blood potassium safety reference under the current kidney function of the patient (such as 4.0 mmol / L); the rule performance index is calculated based on the four-dimensional feedback feature vector and the historical conflict probability value, wherein the clinical performance component is composed of the product of the compliance state, the blood potassium change saturation compression amount and the significant improvement reward item, and the calculation formula of the clinical performance component is:

[0148]

[0149] wherein Γ clin represents the clinical performance component, y complyrepresents the compliance state, i.e. whether the patient follows the warning suggestion (0 / 1), ΔK + represents the absolute change of blood potassium concentration, βp represents the blood potassium change sensitivity coefficient, i.e. the proportional factor for compressing the blood potassium change (default 0.5), γp represents the significant improvement reward coefficient, i.e. the additional reward weight for significant improvement of blood potassium (default 0.2), and δp represents the significant improvement threshold, i.e. the absolute change of blood potassium concentration exceeding 0.3 mmol / L triggers the reward, represents the significant improvement indication function, i.e. ΔK + is 1 when ΔK + represents the saturation function for compressing the absolute change of blood potassium concentration, avoiding extreme values from dominating, and the stability penalty component is the KL divergence value of the current conflict probability value and the historical conflict probability value, and the calculation formula of the stability penalty component is:

[0150] Γ stab = KL(P conflict ||P historical );

[0151] wherein Γ stab represents the stability penalty component, KL represents the KL divergence for measuring the deviation degree of the current conflict probability value and the historical conflict probability value, P conflict represents the current conflict probability value (real-time detected generated conflict probability), and P historical represents the historical conflict probability value (long-term accumulated conflict probability distribution), the rule effectiveness index value is generated by dividing the clinical effectiveness component by the confidence threshold and superimposing the stability penalty item, and the calculation formula of the rule effectiveness index value is:

[0152]

[0153] wherein E rule represents the rule effectiveness index value, a scalar index quantifying the rule effect, λp represents the stability weight coefficient, i.e. the weight of punishing the rule mutation (default 0.1), and τ t represents the current confidence threshold, represents the current confidence threshold normalization, so that the high effectiveness rule obtains higher weight at low threshold;

[0154] In this embodiment, it is specifically required to explain that the specific operation of the rule generation and stability verification of the high-risk state constraint includes: the policy gradient expectation value is calculated with the rule effectiveness index as the weight, and the calculation formula of the policy gradient expectation value is:

[0155]

[0156] wherein, represents the policy gradient expectation value (E rulefor the network to optimize the direction of the weights), p π denotes the state distribution of the policy, the probability distribution of the states visited by the policy, the sampling weight of the high-risk state is higher, and v denotes the total parameter set of the DQN (value network a ), S t denotes the state vector, a t denotes the confidence threshold adjustment amount, E rule denotes the rule effectiveness index value, denotes that the high-risk state of abnormal potassium concentration is preferentially sampled; the difference chemical rate update is performed on the value network and the advantage network, the learning rate of the advantage network is twice that of the value network, and the update formula of the value network and the advantage network is:

[0157]

[0158] wherein, ak v denotes the learning rate of the value network, controls the parameter update step of the value network (its value is set to 0.01), ak a denotes the learning rate of the advantage network, controls the parameter update step of the advantage network (its value is set to 0.02, which is twice that of ak v );

[0159] The specific definition of the high-risk state constraint is:

[0160] When the estimated glomerular filtration rate of the patient is less than thirty, the high-risk potassium concentration interval is greater than or equal to three point zero millimoles per liter to less than or equal to five point five millimoles per liter (including the critical value);

[0161] When the estimated glomerular filtration rate of the patient is between thirty and sixty (including thirty and excluding sixty), the high-risk potassium concentration interval is greater than or equal to three point three millimoles per liter to less than or equal to five point two millimoles per liter (including the critical value);

[0162] When the estimated glomerular filtration rate of the patient is greater than or equal to sixty, the high-risk potassium concentration interval is greater than or equal to three point five millimoles per liter to less than or equal to five point zero millimoles per liter (including the critical value);

[0163] Based on the updated network parameters generate a new taboo rule, only for the high-risk state whose potassium concentration exceeds the safety threshold, and the expression of the new taboo rule is:

[0164]

[0165] wherein, S critical denotes the set of high-risk states, Q(S t , a t ; and new) represents a two-channel deep Q-network, i.e. state vector S t a value evaluation function of the lower confidence threshold adjustment amount, represents a new taboo rule, argmax a represents maximizing action selection;

[0166] The compatibility of the new rule with the historical rule library is evaluated by a rule mutation stability score mechanism, if the score is more than 0.8, the new rule is incorporated into the rule library, and its expression is:

[0167]

[0168] wherein, Score stab represents a rule stability score, i.e. the compatibility score of the new rule with the historical rule library (range 0-1), represents an updated rule library, i.e. the taboo rule library after merging the new rule.

[0169] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0170] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0171] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a machine that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0172] ​These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0173] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0174] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such additional variations and modifications as fall within the scope of the application.

[0175] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. An intelligent decision-making system for the collaborative management of nutrition and metabolism in elderly patients with chronic diseases, characterized in that, Specifically, it includes: The module includes a multimodal knowledge graph construction module, a reinforcement learning rule optimization module, a streaming conflict detection engine module, and a closed-loop rule evolution module. The multimodal knowledge graph construction module, when receiving drug chemical structure data and nutrient metabolism parameters, processes the simplified linear input sequence of drug molecules through graph convolutional networks to generate drug molecule topological feature vectors. At the same time, it integrates nutrient molecule descriptors and bioavailability parameters to generate nutrient metabolism feature vectors, thus constructing a dynamically updated three-dimensional knowledge graph of drug-nutrient-metabolic pathways. The reinforcement learning rule optimization module first constructs a multimodal state space to describe the patient's metabolic state when the patient's real-time medication records, dietary intake data, and blood potassium and creatinine biochemical indicators are input. Then, a dual-channel deep Q network agent is used to generate a confidence threshold based on the action selection strategy. The network is updated through a reward mechanism, and the reward value is calculated based on the degree to which the biochemical indicators deviate from the safe range. The confidence threshold of the taboo rules is dynamically adjusted. The streaming conflict detection engine module performs spatiotemporal alignment on asynchronously arriving medication records and dietary logs when they are input as streaming data. It queries the three-dimensional knowledge graph of drug-nutrient-metabolic pathway through subgraph matching and calculates the conflict probability. If the conflict probability exceeds the dynamic confidence threshold, it triggers a graded warning instruction. The closed-loop rule evolution module, upon receiving patient compliance feedback and 24-hour blood potassium change data after an early warning, performs feedback data fusion and efficacy assessment to generate a rule efficacy index. Based on this index, it updates the dual-channel deep Q-network strategy parameters and drives the autonomous evolution of contraindication rules through the high-risk state of dynamic constraints on the patient's renal function classification.

2. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 1, characterized in that: In the multimodal knowledge graph construction module, the specific operation for generating drug molecule topological feature vectors is as follows: The system receives a simplified linear input sequence of a drug molecule as input data. First, it uses an extended connectivity fingerprint algorithm to transform the sequence into a 1024-dimensional binary fingerprint vector, which describes the local topological environment of atoms in the drug molecule. Then, the binary fingerprint vector is input into a graph convolutional network encoder, where a weighted summation is performed using a linear transformation matrix and a bias vector. A linear rectified activation function is then applied for nonlinear feature extraction, and finally, a 128-dimensional dense feature vector is output as the topological feature vector of the drug molecule. The linear transformation matrix has a dimension of 128 rows by 1024 columns, and the bias vector has a dimension of 128.

3. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 2, characterized in that: The specific operation for generating the nutrient metabolism feature vector is as follows: The system receives a set of nutrient molecular descriptors and bioavailability parameters as input data. The molecular descriptors include at least three physicochemical properties: lipid-water partition coefficient, number of hydrogen bond donors, and molecular weight. These four parameters are combined into a four-dimensional original feature vector and input into an autoencoder neural network. The autoencoder compresses the original feature vector into a 64-dimensional latent space through an encoder subnetwork, generating a fused feature vector that integrates physicochemical properties and metabolic kinetic parameters. Simultaneously, the original input is reconstructed through a decoder subnetwork, with the mean squared error loss function constraining the fidelity of the feature extraction process. Finally, the 64-dimensional fused feature vector generated by the encoder is output as the nutrient metabolism feature vector.

4. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 3, characterized in that: The specific steps for constructing the dynamically updated three-dimensional knowledge graph of drug-nutrient-metabolic pathways are as follows: The system receives drug molecule topological feature vectors and nutrient metabolism feature vectors as input data. It extracts a subgraph of the metabolic pathways resulting from the interaction of target drugs and nutrients from the Kyoto Genome Encyclopedia database, and simultaneously extracts the metabolic entity feature matrix of this subgraph. It extracts the connection relationships between drug nodes and metabolic pathway entity nodes to generate a first adjacency sub-matrix, and extracts the connection relationships between nutrient nodes and metabolic pathway entity nodes to generate a second adjacency sub-matrix. The first and second adjacency sub-matrixes are concatenated along their main diagonal to generate a complete adjacency matrix. The adjacency matrix and node features are input into a relational graph convolutional network. The node embedding state is iteratively updated through a multi-relationship message passing mechanism, ultimately outputting a three-dimensional knowledge graph of drugs, nutrients, and metabolic pathways containing both the node embedding matrix and the adjacency matrix. In each iteration, the relational graph convolutional network differentiates between three connection types—drug associations, nutrient associations, and relationships between metabolic entities—for differentiated information aggregation. When new drug data, new nutrient data, and metabolic research literature are entered, an incremental update operation is performed: A1. Execute the drug molecule topological feature vector generation process on the newly added drug data and output the newly added drug molecule topological feature vector. A2. Perform the nutrient metabolism feature vector generation process on the newly added nutrient data and output the newly added nutrient metabolism feature vector. A3. Extract new metabolic relationship edge sets based on metabolic research literature, and convert the edge sets into sparse tensor format; A4. Combine the newly added drug molecule topological feature vectors, the newly added nutrient metabolism feature vectors, and the edge set represented by sparse tensors with the existing knowledge graph and perform a node and edge union operation. Input the relationship graph convolutional network to generate an updated three-dimensional knowledge graph of drug-nutrient-metabolic pathways.

5. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 4, characterized in that: The specific operations for constructing the multimodal state space in the reinforcement learning rule optimization module are as follows: The system receives real-time patient medication records, dietary intake data, and serum potassium and creatinine biochemical indicators as input data. Medication records are vectorized by converting drug names into unique caloric codes and concatenating them with dose normalization values ​​to generate a medication record encoding vector. Nutrient intake data undergoes logarithmic transformation, adding small positive values ​​to generate a nutrient intake transformation vector. Individualized serum potassium safety range baselines are dynamically calculated based on patient age and estimated glomerular filtration rate, while 70% of the patient's historical mean creatinine is used as the creatinine safety threshold. Finally, the medication record encoding vector, nutrient intake transformation vector, real-time serum potassium value, and real-time creatinine value are combined into a quaternion state vector.

6. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 5, characterized in that: The specific operation of processing state vectors using a dual-channel deep Q-network architecture is as follows: the first channel outputs a confidence threshold adjustment, which is a continuous value in the interval from 0.80 to 0.99; the second channel outputs the state value score; and a comprehensive Q-value is generated by merging the state value score and the action advantage score. The action selection strategy employs an exploration-exploitation balance mechanism, specifically: receiving the comprehensive Q-value output from the dual-channel deep Q-network, subtracting the exploration rate from a probability of 1 to select a confidence threshold adjustment amount that maximizes the comprehensive Q-value; uniformly and randomly selecting a confidence threshold adjustment amount within the range of 0.80 to 0.99 with a probability of exploration rate; and directly assigning the selected adjustment amount as the dynamic confidence threshold; wherein the initial value of the exploration rate is 0.3 and it decays linearly over time. The specific operations for reward calculation and network update are as follows: The immediate reward is calculated based on the serum potassium and creatinine values ​​in the state vector at the next moment. The serum potassium deviation penalty is calculated by compressing the absolute deviation between the actual serum potassium value and the individualized safety baseline using a hyperbolic tangent function, and then multiplying it by a metabolic sensitivity weighting coefficient. The creatinine deviation penalty is calculated by dividing the absolute deviation between the actual creatinine value and the creatinine safety threshold baseline by the patient's baseline creatinine value and then multiplying it by a second weighting coefficient. The two penalty terms are negatively summed and then superimposed on the safety state reward term to generate the total reward value. The empirical priority is calculated using time-series differential error, specifically as follows: B1. Calculate the temporal difference error value: Based on the difference between the current state action value function and the target network's output next state maximum action value function, the reward value is added and the absolute value is taken to generate the temporal difference error value; B2. Assign priority sampling probability: After raising the time difference error value to the power of 0.6, normalize it to generate the sampling weight of each empirical tuple; The target network parameters are gradually synchronized using a soft update strategy: the target network parameters are updated at a rate of 0.01, while the historical target network parameters are retained at a rate of 0.

99. The value network parameters and the advantage network parameters are updated independently.

7. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 6, characterized in that: In the streaming conflict detection engine module, the specific operation for performing spatiotemporal alignment of asynchronously arriving medication records and dietary logs is as follows: C1. Receive medication record stream and dietary log stream as input data, extract drug name and dosage value, generate drug name semantic vector through pre-trained drug encoder, divide the dosage value by the maximum safe dose of the drug to generate normalized dose scalar, and concatenate the two to generate drug semantic dose vector. C2. Extract the names and weights of the ingredients, query the nutrition database to obtain the nutrient baseline vectors for each ingredient, and generate an aggregated vector of ingredient components by weight-weighted summation and normalization. C3. Using the medication time as the reference center point, extract dietary data within a time window from one hour before medication to two hours after medication, and concatenate the drug semantic dose vector with the aggregated vectors of all food components within the window to form a spatiotemporal alignment matrix.

8. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 7, characterized in that: The specific operation of querying the three-dimensional knowledge graph of drugs-nutrients-metabolic pathways through subgraph matching and calculating the conflict probability is as follows: The system receives a spatiotemporal alignment matrix as input data. Using drug nodes as the root, it expands the range by two hops along the metabolic pathway edge in the three-dimensional knowledge graph of drug-nutrient-metabolic pathway to extract the subgraph structure containing associated nutrient nodes. It then extracts the adjacency relationships between drug nodes, nutrient nodes, and metabolic entity nodes within the subgraph, classifies them into three types: drug-enzyme interaction, nutrient-enzyme interaction, and enzyme-enzyme cascade reaction, and stacks them along the third dimension to form a three-dimensional heterogeneous adjacency tensor. Finally, it aligns the feature matrix of the subgraph nodes using a multilayer perceptron. A relation-aware graph convolutional network is used to aggregate neighbor information according to relation type and iteratively update the node embedding state; a high-dimensional interaction tensor is generated by calculating the tensor cross product of the drug molecule topological feature vector and the metabolic feature vector of all nutrients. The interaction tensor is input into the multilayer perceptron and mapped by the S-shaped function to generate a conflict probability value. If the conflict probability exceeds the dynamic confidence threshold, a graded warning instruction is triggered according to the probability interval. The graded warning instructions are as follows: a probability value of 0.80 to 0.89 corresponds to a level 1 warning, 0.90 to 0.94 corresponds to a level 2 warning, and 0.95 and above corresponds to a level 3 warning.

9. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 8, characterized in that: In the closed-loop rule evolution module, the specific operations for performing feedback data fusion and efficacy evaluation to generate the rule efficacy index are as follows: receiving patient compliance labels and serum potassium concentration time series data, verifying the validity of compliance labels and removing outliers; constructing a four-dimensional feedback feature vector, including compliance status, absolute change in serum potassium concentration, current confidence threshold, and serum potassium change direction indicator value; calculating the rule efficacy index based on the four-dimensional feedback feature vector and historical conflict probability values, wherein the clinical efficacy component is composed of the product of compliance status, serum potassium change saturation compression, and significant improvement reward, and the stability penalty component is the KL divergence value between the current conflict probability value and the historical conflict probability value; and dividing the clinical efficacy component by the confidence threshold and then superimposing the stability penalty term to generate the rule efficacy index value.

10. The intelligent decision-making system for collaborative management of nutrition and metabolism in elderly patients with chronic diseases according to claim 9, characterized in that: The specific operations for rule generation and stability verification of the high-risk state constraint include: calculating the expected value of the policy gradient with the rule effectiveness index as the weight, prioritizing the sampling of high-risk states with abnormal blood potassium concentration; performing differential chemical rate updates on the value network and the dominant network, with the dominant network using twice the learning rate of the value network. The specific definition of the high-risk state constraint is as follows: When a patient's estimated glomerular filtration rate is less than 30, the high-risk serum potassium range is a serum potassium concentration greater than or equal to 3.0 mmol / L to less than or equal to 5.5 mmol / L. When a patient's estimated glomerular filtration rate is between 30 and 60 (inclusive but exclusive of 60), the high-risk range for serum potassium is a serum potassium concentration greater than or equal to 3.3 mmol / L and less than or equal to 5.2 mmol / L. When a patient's estimated glomerular filtration rate is greater than or equal to 60, the high-risk serum potassium range is a serum potassium concentration greater than or equal to 3.5 mmol / L to less than or equal to 5.0 mmol / L. New taboo rules are generated based on the updated network parameters, and rules are only generated for high-risk states where blood potassium concentration exceeds the safe threshold. The compatibility of new rules with the historical rule base is evaluated through a rule mutation stability scoring mechanism. If the score exceeds 0.8, the new rule will be incorporated into the rule base.

Images