Knowledge graph-based kidney disease patient diet intelligent recommendation method

Abstract

The invention discloses a kidney disease patient diet intelligent recommendation method based on a knowledge graph, and relates to the technical field of medical health data processing and intelligent recommendation. The method comprises the following steps: constructing a food knowledge graph, constraining a semantic relationship among food, nutritional ingredients, patient states and diet rules by ontology, and solidifying nutritional attribute interval values and credibility parameters at food nodes; generating an individualized nutrition threshold vector based on the state of the patient, executing interval nutrition summarization on the candidate dietary scheme to obtain a nutrition upper bound vector, calculating a nutrition risk upper bound, and implementing safety gating with the nutrition risk upper bound being not greater than zero; carrying out directional repair on a gating failure scheme by replacing retrieval with a dominant nutrient-driven atlas; outputting an evidence chain carrying the source identifier, the credibility parameter and the version identifier; and in the feedback updating stage, a safety non-degradation constraint is applied, and collaborative rollback and cooling control is triggered, so that the recommendation safety, interpretability and continuous iteration controllability are improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health data processing and intelligent recommendation technology, specifically to a knowledge graph-based intelligent dietary recommendation method for kidney disease patients. Background Technology

[0002] Dietary management for kidney disease patients is a multi-constraint safety decision-making problem. Constraints include at least daily upper limits for sodium, potassium, phosphorus, and protein intake, and are coupled with the patient's follow-up status. In an information-based context, dietary recommendations must not only generate actionable food combinations but also achieve safety verification before output, interpretable and traceable recommendation criteria, and safe and controllable continuous model updates.

[0003] Current dietary recommendation schemes generally employ a structure that combines nutrient point values ​​with static threshold filtering. This structure suffers from three core flaws in real-world business environments:

[0004] First, the nutritional content of ingredients fluctuates depending on the source, batch, processing stage, and cooking method. The point value superposition cannot provide a conservative judgment on the upper limit of the combined superposition, resulting in the risk of exceeding the threshold not being blocked before output.

[0005] Second, the threshold adjustment relies on fixed period rules or manual experience, lacks a quantitative tightening mechanism based on the trend of indicators through follow-up windows, the threshold update cannot be recalculated, and the audit consistency is insufficient.

[0006] Third, online system updates are often aimed at satisfaction or clicks, lacking acceptance, rollback, and cooling-off control mechanisms with insecure recommendation rates as hard constraints. At the same time, without evidence chains and version identifiers, it is difficult to complete compliance audits and trace responsibilities.

[0007] Therefore, there is a need for a smart diet recommendation scheme that explicitly incorporates nutritional uncertainty into the calculation, implements gating with a unified safety variable, performs targeted repair of failed schemes, and introduces safety non-deterioration constraints and rollback controls during the update phase. Summary of the Invention

[0008] Technical Objective: To address the shortcomings of existing technologies, this invention discloses a knowledge graph-based intelligent dietary recommendation method for kidney disease patients. This method can perform safety assessments and block risks before outputting candidate dietary plans under upper bound conditions, can quantify and map follow-up trends to threshold updates, and can maintain the safety indicators without deterioration during continuous updates while supporting audit rollback.

[0009] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0010] A knowledge graph-based intelligent dietary recommendation method for kidney disease patients includes the following steps:

[0011] S1. Obtain patient health data related to nutritional management of patients with kidney disease and perform structured parsing to obtain patient state vectors;

[0012] S2. Construct and store a food knowledge graph, which includes food nodes, nutrient component nodes, patient status nodes, and dietary rule nodes. The food nodes are associated with nutrient attribute range values ​​and credibility parameters. The nodes and relationships of the food knowledge graph are associated with source identifiers and version identifiers.

[0013] S3. Generate individualized nutrition threshold vectors based on patient state vectors;

[0014] S4. Generate a set of candidate dietary plans based on the food knowledge graph, and perform interval nutrition summarization for each candidate dietary plan to obtain the nutrition upper bound vector.

[0015] S5. Calculate the upper bound of nutritional risk based on the upper bound vector of nutrition and the individualized nutritional threshold vector, and use the upper bound of nutritional risk not being greater than zero as a safety gate condition to screen and obtain a set of safe dietary plans.

[0016] S6. Sort the set of safe diet plans and output the diet recommendation results, and output the graph evidence chain corresponding to the diet recommendation results;

[0017] S7. For candidate dietary plans that fail to pass the safety gating conditions, identify the dominant nutrient that causes the upper bound of nutritional risk to be positive, and use the dominant nutrient as a constraint to perform a map replacement search to repair the candidate dietary plan. After repair, return to step S4 or step S5.

[0018] S8. Obtain feedback data related to the dietary recommendation results, and update the food knowledge graph and ranking process based on the feedback data. Apply safety and non-deterioration constraints during the update process to ensure that the unsafe recommendation rate after the update is not higher than the unsafe recommendation rate before the update.

[0019] In one embodiment, the patient health data includes kidney disease staging information, kidney function indicators, blood biochemistry indicators, body weight indicators, medication information, and dietary intake records; the kidney function indicators include estimated glomerular filtration rate values; and the blood biochemistry indicators include serum sodium, serum potassium, and serum phosphorus values.

[0020] In one embodiment, the individualized nutrient threshold vector consists of the daily upper limits for sodium, potassium, phosphorus, and protein intake, and is determined by a baseline threshold and a follow-up trend.

[0021] In one embodiment, the follow-up trend is calculated from the changes in test indicators within a continuous follow-up window; when the follow-up trend indicates an increase in serum potassium levels, the upper limit of daily potassium intake is tightened; when the follow-up trend indicates an increase in serum phosphorus levels, the upper limit of daily phosphorus intake is tightened; and when the follow-up trend indicates a decrease in estimated glomerular filtration rate, the upper limit of daily protein intake is tightened.

[0022] In one embodiment, the food knowledge graph constrains node types and relation types with ontology. The relation types include relationships such as food containing nutrients, nutrient association test index relationships, patient state triggering dietary rules relationships, food satisfying dietary rules relationships, food substitutability relationships, and cooking methods affecting nutrient retention relationships. The graph evidence chain is composed of nodes and relations corresponding to the dietary recommendation results in the order of the reasoning path. The graph evidence chain includes node identifiers, relation identifiers, source identifiers, credibility parameters, and version identifiers.

[0023] In one embodiment, the interval nutrition summary includes linearly superimposing the interval values ​​of nutritional attributes of each ingredient in the candidate dietary plan by portion size, and superimposing the source fluctuation upper bound coefficient and the cooking retention upper bound coefficient to obtain a nutrition upper bound vector; for any nutrient n, the upper bound intake value of the candidate dietary plan m. satisfy:

[0024] ,

[0025] in, This refers to the quantity of ingredient i, in grams. The content of nutrient n units in food i. This is the upper bound coefficient for source fluctuation. The upper bound coefficient is retained for cooking.

[0026] According to claim 1, a knowledge graph-based intelligent dietary recommendation method for kidney disease patients is characterized in that the upper bound of nutritional risk is calculated as follows: a standardized excess value is calculated for each nutrient n.

[0027] ,

[0028] in, To standardize excess values, This represents the upper limit of daily intake of nutrient n. Standardized factors; upper limit of nutritional risk satisfy

[0029] ,

[0030] and with As a safety gating condition; the dominant nutrient is to enable The nutrient with the highest value is selected; the map replacement retrieval and repair of candidate dietary plans includes: calculating the upper bound contribution value of each ingredient to the dominant nutrient in the candidate dietary plan.

[0031] ,

[0032] in, For the quantity of ingredient i, The content of nutrient n units in food i. This is the upper bound coefficient for source fluctuation. To preserve the upper bound coefficient for cooking, the ingredient with the largest contribution value is selected as the ingredient to be replaced; a set of alternative ingredients is obtained by searching along the food substitutability relationship; for each alternative ingredient, replacement is performed and recalculation is performed. , choose to The replacement result with the largest decrease is taken as the repair result, and the repaired candidate dietary plan is returned to step S4 or step S5.

[0033] A knowledge graph-based intelligent dietary recommendation system for kidney disease patients, used to implement the knowledge graph-based intelligent dietary recommendation method for kidney disease patients as described above, includes:

[0034] The health data parsing module is used to acquire patient health data and output patient state vectors;

[0035] The knowledge graph management module is used to build and store food knowledge graphs and maintain the source identifiers, credibility parameters and version identifiers of nodes and relationships;

[0036] The threshold calculation module is used to generate individualized nutrition threshold vectors based on the patient state vector;

[0037] The candidate generation module is used to generate a set of candidate dietary plans based on a food knowledge graph.

[0038] The upper bound summary module is used to perform interval nutrient summary on candidate dietary plans and output the nutrient upper bound vector.

[0039] The risk gating module is used to calculate the upper bound of nutritional risk and screen a set of safe dietary plans based on safety gating conditions.

[0040] The repair retrieval module is used to identify the dominant nutrients for candidate dietary plans that have not passed the gating and perform a map substitution retrieval to complete the repair.

[0041] The sorting and interpretation module is used to sort the set of safe diet plans, output dietary recommendations, and output a graph evidence chain.

[0042] The feedback security update module is used to acquire feedback data and update the food knowledge graph and sorting process under the constraint of security and no degradation.

[0043] In one embodiment, it also includes an audit log module for recording candidate dietary plan identifiers, nutritional upper bound vectors, individualized nutritional threshold vectors, nutritional risk upper bounds, safety gating judgment results, dominant nutrient identifiers, graph evidence chains, and corresponding version identifiers.

[0044] In one embodiment, the feedback security update module includes a collaborative rollback module and a cooling control module. The collaborative rollback module is used to roll back the food knowledge graph and sorting process to the corresponding version before the update when the security non-deterioration constraint is violated. The cooling control module is used to enter a cooling period after the rollback and limit the update frequency, while recording the rollback reason and rollback timestamp.

[0045] Beneficial Effects: The intelligent dietary recommendation method for kidney disease patients based on knowledge graphs provided in this invention has the following beneficial effects:

[0046] 1. This invention solidifies the nutritional attribute interval values ​​of food nodes in the food knowledge graph, and performs interval nutritional summarization at the candidate solution level to obtain the nutritional upper bound vector. Then, the nutritional risk upper bound is used as a unified safety variable to implement gating, so that the system can perform pre-output verification on the upper bound situation after the superposition of multiple ingredients, solve the problem of safety omission under the conditions of source fluctuation and cooking fluctuation of point value superposition, reduce the unsafe recommendation rate and improve clinical usability.

[0047] 2. This invention establishes a recalculable mapping between the individualized nutrition threshold vector and the trend of test indicators within the continuous follow-up window through a threshold calculation module, and enables the threshold vector and risk gate to share the same vectorized caliber, realizing a consistent closed loop from state change to threshold tightening and then to gate judgment, improving the traceability and consistency of threshold updates, and avoiding the split of judgment caliber caused by rule drift.

[0048] 3. In the feedback security update module, this invention applies the unsafe recommendation rate as a security indicator to impose a non-deterioration constraint, and triggers a collaborative rollback and cooling control of the graph and ranking process when the constraint is violated. At the same time, the source identifier, credibility parameter and version identifier are recorded with evidence chain and audit log, so as to realize an auditable, rollbackable and controllable iterative engineering closed loop, and enhance its compliance deployment capability in medical and health big data sharing platforms and health-related information systems and cloud platforms. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0050] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0051] Figure 2This is a block diagram of the system structure of the present invention;

[0052] Figure 3 A schematic diagram of the nodes and relationships in a food knowledge graph;

[0053] Figure 4 A schematic diagram of the upper bound gating and targeted repair loop for nutritional risk. Detailed Implementation

[0054] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0055] like Figure 1 As shown, a knowledge graph-based intelligent dietary recommendation method for kidney disease patients includes the following steps:

[0056] S1. Obtain patient health data related to nutritional management of kidney disease patients and perform structured parsing to obtain patient state vectors.

[0057] The system acquires patient health data related to nutritional management of kidney disease patients, performs structured parsing, and generates patient state vectors. The patient status vector is a vectorized representation of a patient's health status within a follow-up window. It includes the numerical values ​​of each indicator, the unit of measurement, the time window boundaries, and the data version identifier, and is used for subsequent threshold calculation and rule triggering.

[0058] Patient health data includes kidney disease staging information, renal function indicators, blood biochemistry indicators, body weight indicators, medication information, and dietary intake records; the estimated glomerular filtration rate is expressed in mL / min / 1.73 mL. 2 Serum sodium, potassium, and phosphorus levels are measured in mmol / L; food portion sizes are measured in grams. The structured parsing process includes field mapping, unit normalization, timestamp consistency checks, and missing data checks. If a key indicator is missing, preventing subsequent threshold calculations, the system generates a non-recommended status and records the reason code to avoid outputting recommended results when there is a lack of safety judgment criteria.

[0059] The output is a patient state vector. This generates a data version identifier bound to the vector, which is then used for reference in audit logs.

[0060] S2. Construct and store a food knowledge graph, which includes food nodes, nutrient component nodes, patient status nodes, and dietary rule nodes. The food nodes are associated with nutrient attribute range values ​​and credibility parameters. The nodes and relationships in the food knowledge graph are associated with source identifiers and version identifiers.

[0061] The system builds and stores a food knowledge graph. .like Figure 3 As shown, the food knowledge graph is a semantic network composed of a set of nodes and a set of relations. It includes at least food nodes, nutritional component nodes, patient status nodes, and dietary rule nodes, as well as inspection indicator nodes and cooking method nodes. The graph uses ontology to constrain node types and relation types, making graph querying, reasoning, and evidence chain derivation semantically closed.

[0062] For each food node i and nutrient n, the graph maintains the following field as input for subsequent secure calculations: content per unit. Source fluctuation upper bound coefficient Cooking retains the upper bound coefficient The system maintains a credibility parameter q. Nodes and relationships in the graph carry source and version identifiers; any change to a node attribute or relationship generates a new version identifier and records the changed object, values ​​before and after the change, reason for the change, and timestamp, ensuring that the recommendation calculation is repeatable and traceable.

[0063] S3. Generate individualized nutrition threshold vectors based on patient state vectors.

[0064] The system is based on patient state vectors Generate individualized nutrient threshold vectors The individualized nutrient threshold vector is a set of daily intake limits expressed in vector form, consisting of at least the daily intake limits for sodium, potassium, phosphorus, and protein, used to safely gate the nutritional upper bounds of candidate dietary plans.

[0065] The threshold is generated using a baseline threshold plus a tightening follow-up trend determination calculation:

[0066]

[0067]

[0068] in, The baseline threshold, This is the tightening coefficient. This represents the trend strength of the corresponding test indicators within the follow-up window. The trend strength is obtained through time series differencing and stability verification; when the blood potassium trend increases, the potassium threshold is tightened; when the blood phosphorus trend increases, the phosphorus threshold is tightened; and when the estimated glomerular filtration rate trend decreases, the protein threshold is tightened. The system assigns a threshold version identifier to each threshold generation and records the triggering indicator and trigger timestamp.

[0069] S4. Generate a set of candidate dietary plans based on the food knowledge graph, and perform interval nutrition summarization for each candidate dietary plan to obtain the nutrition upper bound vector.

[0070] The system is based on a food knowledge graph. Generate a set of candidate dietary plans A candidate meal plan is a structured object consisting of multiple ingredient items, each containing an ingredient identifier and serving size. Including cooking method identifiers. Candidate generation employs a combination of dietary structure templates, rule triggering, constraint search, and pruning to avoid uncontrollable combinations resulting solely from similarity concatenation.

[0071] The candidate generation process includes:

[0072] Dietary structure template definition: A daily diet is expressed as multiple slots, each slot is limited by meal frequency, food category constraints and portion size boundaries; food category constraints are determined by the ontological category hierarchy.

[0073] Rule triggering: The diet rule node is triggered based on the patient status node to obtain the rule set; the rule set is used for slot candidate retrieval and combination constraints.

[0074] Slot candidate retrieval: For each slot, retrieve a set of food items from the graph that satisfy both category and rule constraints, and the retrieval results must have... , , The confidence parameter q ensures that the subsequent upper bound aggregation is computable.

[0075] Constraint Search and Pruning: Gradually fill combinations by slot, and use individualized nutrient threshold vectors. Perform partial upper bound pruning; terminate the branch and record the pruning reason code when a partial combination exceeds the upper threshold under the achievable minimum upper bound estimate.

[0076] Candidate generation outputs a set of candidate dietary plans. It also fixes the referenced map version identifier, rule version identifier, and template version identifier.

[0077] S5. Calculate the upper bound of nutritional risk based on the nutritional upper bound vector and the individualized nutritional threshold vector, and use the upper bound of nutritional risk not being greater than zero as a safety gate condition to screen and obtain a set of safe dietary plans.

[0078] The system calculates the nutritional upper bound vector for each candidate dietary plan m. The nutrient upper bound vector represents a conservative upper bound for intake of candidate options, considering source variation and cooking retention uncertainty. The upper bound intake values ​​for nutrients are:

[0079]

[0080] in, This is a measure of weight, expressed in grams (g). Content per unit; This is the upper bound coefficient for source fluctuations; The upper bound coefficient is retained for cooking.

[0081] The system calculates the standardized excess value and the upper bound of nutritional risk:

[0082]

[0083]

[0084] in, This is to standardize the factors and incorporate them into version management. The set of safe dietary plans was obtained by screening based on safety gating conditions, and the details of each plan were recorded. , , Reference version identifier and gating judgment result.

[0085] S6. Sort the set of safe diet plans and output the dietary recommendations, and output the graph evidence chain corresponding to the dietary recommendations.

[0086] The system sorts the set of safe dietary plans and outputs dietary recommendations. The sorting uses a deterministic sorting key, and the input includes risk margin, aggregated values ​​of confidence parameters, preference matching degree, and rule fulfillment completeness, ensuring consistency between the output and the same version of the system under the same input criteria.

[0087] The system also outputs a graph evidence chain. The graph evidence chain is an ordered set of graph nodes and relationships corresponding to the recommendation results, organized in the order of the reasoning path, and includes node identifiers, relationship identifiers, source identifiers, credibility parameters, and version identifiers, enabling the basis for the recommendation to be verified.

[0088] S7. For candidate dietary plans that fail to pass the safety gating conditions, identify the dominant nutrient that causes the upper bound of nutritional risk to be positive, and use the dominant nutrient as a constraint to perform a map replacement search to repair the candidate dietary plan. After repair, return to step S4 or step S5.

[0089] For candidate solutions that fail the gating process, the system determines the dominant nutrient. :

[0090]

[0091] The system calculates the upper limit contribution of each ingredient to the dominant nutrient:

[0092]

[0093] And determine the ingredients to be replaced:

[0094]

[0095] The system retrieves alternative sets along food substitutability relationships. For each alternative candidate, perform a replacement and recalculate. , choose to The replacement result with the largest decrease is used as the repair result, and the repaired solution is returned to the candidate generation or risk gating step for review. The system records the dominant nutrient identifier, the identifier of the ingredient to be replaced, the identifier of the substitute ingredient, and the before and after repair. Version identifier.

[0096] like Figure 4 As shown, the system first performs interval nutrient summation on each scheme m in the candidate dietary scheme set to obtain the nutrient upper bound vector. Combined with individualized nutrient threshold vectors Calculate the standardized excess value and obtain the upper bound of the nutritional risk of the scheme, thereby performing a safety gating judgment; when When the solution directly enters the sorting and interpretation module, it outputs the recommendation results and the graph evidence chain. Time system to determine dominant nutrients The algorithm calculates the upper bound contribution of each ingredient in the proposed solution to the dominant nutrient and identifies the ingredients to be replaced. Then, it retrieves the set of alternative ingredients along the substitutability relationships in the food knowledge graph and performs a "replacement-recalculation" process on the candidate solutions. —Recalculation "Iterative repair, select to make The replacement result with the largest decrease is returned to the gating link for review until it passes the gating and is output, forming a closed-loop recommendation process that is traceable, interpretable, and meets individualized nutritional safety constraints.

[0097] S8. Obtain feedback data related to the dietary recommendation results, and update the food knowledge graph and ranking process based on the feedback data. Apply safety and non-deterioration constraints during the update process to ensure that the unsafe recommendation rate after the update is not higher than the unsafe recommendation rate before the update.

[0098] The system acquires feedback data, which includes actual intake data, follow-up test indicators, and timestamps. Based on the feedback data, the system generates update candidates for reliability parameters, source fluctuation upper bound coefficients, or ranking parameters, and calculates the unsafe recommendation rate before and after the update. and Apply safety and non-deterioration constraints:

[0099]

[0100] When the constraint is met, an update is submitted and a new version identifier is generated; when the constraint is not met, a collaborative rollback is triggered, reverting the graph and sorting process to the version before the update, and entering a cooldown period to limit the update frequency, while recording the rollback reason and rollback timestamp.

[0101] like Figure 2As shown, the present invention also provides a knowledge graph-based intelligent dietary recommendation system for kidney disease patients, used to implement the knowledge graph-based intelligent dietary recommendation method for kidney disease patients as described above, including:

[0102] The health data parsing module is used to acquire patient health data and output patient state vectors. This module completes field mapping, unit normalization, timestamp consistency verification, missing data verification, and quality tag generation, and outputs patient state vectors. Additional data version identifiers and follow-up window boundaries. This module provides consistent input to the threshold calculation module and candidate generation module, and provides the original record reference identifier and verification results to the audit log module.

[0103] The knowledge graph management module is used to build and store the food knowledge graph and maintain the source identifiers, credibility parameters, and version identifiers of nodes and relationships. This module provides ontology constraint checks to ensure the legality of node and relationship types; it provides a graph query interface supporting constraints by food category, rule predicate constraints, substitutable relationship retrieval, and evidence chain backtracking; and it provides version governance capabilities, generating version identifiers and fixing snapshots for graph changes, ensuring that the same recommendation request is calculated under a fixed version caliber.

[0104] The threshold calculation module generates individualized nutritional threshold vectors based on patient state vectors. This module performs baseline threshold generation, follow-up trend strength calculation, threshold tightening calculation, and threshold version identification management, and outputs the individualized nutritional threshold vector. The bound trigger information and timestamp are used for auditing and review.

[0105] The candidate generation module is used to generate a set of candidate dietary plans based on a food knowledge graph. This module maintains dietary structure templates and permanently identifies template versions; based on... Trigger the rule set and solidify the rule version identifier; perform candidate retrieval by slot and ensure that the candidates have the fields required for upper bound summary; perform constraint search to generate combinations and call some upper bound pruning logic to reduce search branches that must exceed the threshold; when outputting the candidate set, output the candidate generation version identifier and reference version identifier set at the same time.

[0106] The upper bound summary module is used to perform interval nutrient summarization on candidate dietary plans and output a nutrient upper bound vector. This module permanently references the specified values ​​during computation. , , The version of the map must be used to ensure consistency in recalculation; for the same scheme and the same version combination, result caching can be enabled to reduce the overhead of repeated calculations in the repair loop.

[0107] The risk gating module is used to calculate the upper bound of nutritional risk and select a set of safe dietary plans based on safety gating conditions. This module maintains the standardized factor S(n) and its version identifier to ensure the stability of the risk measurement caliber; the gating decision output is written to the audit log and provides the set of safe plans to the sorting and interpretation module.

[0108] The repair retrieval module is used to identify the dominant nutrients for candidate dietary plans that have not passed the gating process and perform a map substitution retrieval to complete the repair. This module identifies the dominant nutrients. Calculate the contribution value And locate the ingredients to be replaced Retrieve the set of alternatives along the substitutability relation and evaluate the replacement. The replacement result with the largest risk reduction is selected; the repair result is sent back to the upper bound summary module and the risk gating module for review, forming a closed loop.

[0109] The sorting and interpretation module sorts the set of safe dietary plans, outputs dietary recommendations, and generates a graph-based evidence chain. This module uses a deterministic sorting key to ensure stable output. The evidence chain export calls the backtracking interface of the graph management module to ensure that the version referenced in the evidence chain is consistent with the version referenced by the risk gating system, avoiding inconsistencies between recommendations and interpretations.

[0110] The feedback safety update module is used to acquire feedback data and update the food knowledge graph and ranking process under the constraint of safety and no degradation. This module generates update candidates based on the credibility parameter, the upper bound coefficient of source fluctuation, and the ranking parameter; performs safety and no degradation constraint verification using the unsafe recommendation rate as a safety indicator; submits the update and generates a new version identifier when the verification passes; and triggers the collaborative rollback module and the cooling control module when the verification fails.

[0111] In one specific embodiment, the feedback security update module includes a collaborative rollback module and a cooling control module. The collaborative rollback module is used to roll back the food knowledge graph and sorting process to the corresponding version before the update when the security non-deterioration constraint is violated. The cooling control module is used to enter a cooling period after the rollback and limit the update frequency, while recording the rollback reason and rollback timestamp.

[0112] In one specific embodiment, the system further includes an audit log module for recording candidate dietary plan identifiers, nutritional upper bound vectors, individualized nutritional threshold vectors, nutritional risk upper bounds, safety gating judgment results, dominant nutrient identifiers, graph evidence chains, and corresponding version identifiers. The audit logs must include at least: candidate plan identifiers, nutritional upper bound vectors, individualized nutritional threshold vectors, nutritional risk upper bounds, gating judgment results, dominant nutrient identifiers, graph evidence chain summaries, referenced graph version identifiers and threshold version identifiers, risk values ​​before and after repair, updated acceptance indicators, and rollback records, ensuring the recommendation process is recalculated and accountability is traceable.

[0113] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent dietary recommendation method for kidney disease patients, characterized in that, Specifically, the following steps are included: S1. Obtain patient health data related to nutritional management of patients with kidney disease and perform structured parsing to obtain patient state vectors; S2. Construct and store a food knowledge graph, which includes food nodes, nutrient component nodes, patient status nodes, and dietary rule nodes. The food nodes are associated with nutrient attribute range values ​​and credibility parameters. The nodes and relationships of the food knowledge graph are associated with source identifiers and version identifiers. S3. Generate individualized nutrition threshold vectors based on patient state vectors; S4. Generate a set of candidate dietary plans based on the food knowledge graph, and perform interval nutrition summarization for each candidate dietary plan to obtain the nutrition upper bound vector. S5. Calculate the upper bound of nutritional risk based on the nutritional upper bound vector and the individualized nutritional threshold vector, and use the nutritional risk upper bound not being greater than zero as a safety gate condition to screen and obtain a set of safe dietary plans. S6. Sort the set of safe diet plans and output the diet recommendation results, and output the graph evidence chain corresponding to the diet recommendation results; S7. For candidate dietary plans that fail to pass the safety gating conditions, identify the dominant nutrient that causes the upper bound of nutritional risk to be positive, and use the dominant nutrient as a constraint to perform a map replacement search to repair the candidate dietary plan. After repair, return to step S4 or step S5. S8. Obtain feedback data related to the dietary recommendation results, and update the food knowledge graph and ranking process based on the feedback data. Apply safety and non-deterioration constraints during the update process to ensure that the unsafe recommendation rate after the update is not higher than the unsafe recommendation rate before the update.

2. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 1, characterized in that, The patient's health data includes information on kidney disease staging, kidney function indicators, blood biochemistry indicators, body weight indicators, medication information, and dietary intake records; the kidney function indicators include estimated glomerular filtration rate values; and the blood biochemistry indicators include serum sodium, serum potassium, and serum phosphorus values.

3. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 1, characterized in that, The individualized nutrition threshold vector consists of the upper limit of daily intake of sodium, potassium, phosphorus and protein, and is determined by the baseline threshold and follow-up trend.

4. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 3, characterized in that, The follow-up trend is calculated from the changes in test indicators within a continuous follow-up window; when the follow-up trend indicates an increase in serum potassium levels, the upper limit of daily potassium intake is tightened; when the follow-up trend indicates an increase in serum phosphorus levels, the upper limit of daily phosphorus intake is tightened; when the follow-up trend indicates a decrease in estimated glomerular filtration rate, the upper limit of daily protein intake is tightened.

5. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 1, characterized in that, The food knowledge graph uses ontology to constrain node types and relation types. The relation types include the relationship of food containing nutrients, the relationship of nutrient association test indicators, the relationship of patient status triggering dietary rules, the relationship of food satisfying dietary rules, the relationship of food substitutability, and the relationship of cooking methods affecting nutrient retention. The graph evidence chain consists of nodes and relationships corresponding to the dietary recommendation results arranged in the order of the reasoning path. The graph evidence chain includes node identifiers, relationship identifiers, source identifiers, credibility parameters, and version identifiers.

6. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 1, characterized in that, The interval nutrition summary includes linearly superimposing the interval values ​​of nutritional attributes of each ingredient in the candidate dietary plan by portion size, and superimposing the source fluctuation upper bound coefficient and the cooking retention upper bound coefficient to obtain the nutrition upper bound vector; for any nutrient n, the upper bound intake value of the candidate dietary plan m. satisfy: ， in, This refers to the quantity of ingredient i, in grams. The content of nutrient n units in food i. This is the upper bound coefficient for source fluctuation. The upper bound coefficient is retained for cooking.

7. The method for intelligent dietary recommendation for kidney disease patients based on knowledge graphs according to claim 1, characterized in that, The upper bound of nutritional risk is calculated as follows: For each nutrient n, a standardized excess value is calculated. ， in, To standardize excess values, This represents the upper limit of daily intake of nutrient n. Standardized factors; upper limit of nutritional risk satisfy ， and with As a safety gating condition; the dominant nutrient is to enable The nutrient with the highest value is selected; the map replacement retrieval and repair of candidate dietary plans includes: calculating the upper bound contribution value of each ingredient to the dominant nutrient in the candidate dietary plan. ， in, Let i be the quantity of ingredient i. The content of nutrient n units in food i. This is the upper bound coefficient for source fluctuation. To preserve the upper bound coefficient for cooking, the ingredient with the largest contribution value is selected as the ingredient to be replaced; a set of alternative ingredients is obtained by searching along the food substitutability relationship; for each alternative ingredient, replacement is performed and recalculation is performed. , choose to The replacement result with the largest decrease is taken as the repair result, and the repaired candidate dietary plan is returned to step S4 or step S5.

8. A knowledge graph-based intelligent dietary recommendation system for kidney disease patients, characterized in that, A method for implementing a knowledge graph-based intelligent dietary recommendation method for kidney disease patients as described in any one of claims 1-7 includes: The health data parsing module is used to acquire patient health data and output patient state vectors; The knowledge graph management module is used to build and store food knowledge graphs and maintain the source identifiers, credibility parameters and version identifiers of nodes and relationships; The threshold calculation module is used to generate individualized nutrition threshold vectors based on the patient state vector; The candidate generation module is used to generate a set of candidate dietary plans based on a food knowledge graph. The upper bound summary module is used to perform interval nutrient summary on candidate dietary plans and output the nutrient upper bound vector. The risk gating module is used to calculate the upper bound of nutritional risk and screen a set of safe dietary plans based on safety gating conditions. The repair retrieval module is used to identify the dominant nutrients for candidate dietary plans that have not passed the gating and perform a map substitution retrieval to complete the repair. The sorting and interpretation module is used to sort the set of safe diet plans, output dietary recommendations, and output a graph evidence chain. The feedback security update module is used to acquire feedback data and update the food knowledge graph and sorting process under the constraint of security and no degradation.

9. A knowledge graph-based intelligent dietary recommendation system for kidney disease patients according to claim 8, characterized in that, It also includes an audit log module, which records candidate dietary plan identifiers, nutritional upper bound vectors, individualized nutritional threshold vectors, nutritional risk upper bounds, safety gating judgment results, dominant nutrient identifiers, graph evidence chains, and corresponding version identifiers.

10. A knowledge graph-based intelligent dietary recommendation system for kidney disease patients according to claim 8, characterized in that, The feedback security update module includes a collaborative rollback module and a cooling control module. The collaborative rollback module is used to roll back the food knowledge graph and sorting process to the corresponding version before the update when the security non-deterioration constraint is violated. The cooling control module is used to enter a cooling period after the rollback and limit the update frequency, while recording the rollback reason and rollback timestamp.

Images