Nutrition management strategy recommendation method and system for ICU survival stroke postoperative rehabilitation patient

Abstract

The invention discloses a nutrition management strategy recommendation method and system for an ICU survival stroke postoperative rehabilitation patient, and relates to the technical field of medical health and artificial intelligence, and the method comprises the steps: obtaining basic physiological indexes, existing examination data and medicine use records of the ICU survival stroke postoperative rehabilitation patient according to a medical record; after data preprocessing, obtaining corresponding character feature representation through a pre-training model, calling a domain knowledge graph to perform knowledge enhancement on the feature representation, and constructing multi-dimensional nutrition and health features; using a graph convolutional neural network to perform learning representation on the multi-dimensional nutrition and health feature corresponding graph structure, extracting a nutrition and health portrait of the patient, and performing nutrition and health state evaluation of the patient; and extracting a personalized nutrition management strategy for recommendation according to the evaluation result of the nutrition health state of the patient. According to the invention, accurate evaluation of the nutrition health state of the rehabilitation patient is realized by combining mass rehabilitation nutrition management knowledge according to existing inspection data, and an accurate and personalized nutrition management scheme is provided.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical health and artificial intelligence, and more specifically, to a method and system for recommending a nutritional management strategy for postoperative rehabilitation patients surviving stroke in the ICU. Background Art

[0002] Stroke is an acute cerebrovascular disease with a high incidence and high disability rate. Postoperative patients often suffer from severe functional disorders and metabolic disorders. For postoperative stroke patients surviving in the ICU, nutritional management during the rehabilitation period is crucial, directly affecting the rehabilitation effect, quality of life, and long-term prognosis of the patients. However, nutritional management for postoperative stroke patients faces many challenges: (1) Dysphagia: Stroke patients often suffer from dysphagia (such as swallowing disorders or swallowing weakness), resulting in difficulty in eating. Usually, an intestinal tube or nasogastric tube needs to be indwelled for nutritional supplementation, increasing the risk of malnutrition and aspiration pneumonia. (2) Metabolic disorders: The energy consumption and protein catabolism of stroke patients increase significantly, easily leading to negative nitrogen balance and muscle atrophy. (3) Large individual differences: Factors such as the patient's age, underlying diseases, stroke type, and severity vary greatly, and a unified nutritional management plan is difficult to meet individual needs. (4) Dynamic changes during the rehabilitation stage: During the rehabilitation process of patients, nutritional needs change dynamically with the condition, and traditional static nutritional management methods are difficult to adapt to. (5) High risk of complications: Malnutrition will further increase the risk of complications such as infection, pressure sores, and deep vein thrombosis, prolong the hospital stay, and increase medical costs.

[0003] The nutritional management method for postoperative rehabilitation patients surviving stroke in the ICU often relies on the empirical judgment of medical staff, resulting in problems such as low efficiency and insufficient accuracy. At the same time, due to large individual differences and complex nutritional needs of patients, traditional methods are difficult to provide precise and personalized nutritional support, lack the ability of dynamic adjustment, and are difficult to meet the diverse needs of patients. And in the health status assessment of patients, it often also relies on preset detection indicators, which increases the economic pressure of postoperative stroke rehabilitation patients to a certain extent. Therefore, how to evaluate the health status of rehabilitation patients based on existing examination data combined with knowledge enhancement means, and provide the most appropriate nutritional support strategy for postoperative stroke patients surviving in the ICU to improve the rehabilitation effect is an urgent problem to be solved. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for recommending a nutritional management strategy for postoperative rehabilitation patients surviving stroke in the ICU, which can accurately evaluate the nutritional and health status of rehabilitation patients based on existing examination data combined with a large amount of rehabilitation nutrition management knowledge, and provide a precise and personalized nutritional management plan to meet the diverse needs of postoperative stroke patients.

[0005] The first aspect of the present invention provides a method for recommending a nutritional management strategy for ICU surviving stroke patients after surgery, including the following steps: Obtain the basic physiological indicators, existing examination data, and drug use records of ICU surviving stroke patients after surgery according to the medical records; Use the basic physiological indicators to obtain basic feature representations to provide baseline data for the nutritional health assessment of patients, generate examination feature representations and drug feature representations using the existing examination data and drug use records, call the knowledge graph in the relevant field to enhance the feature representations, and construct multi-dimensional nutritional health features by combining the knowledge-enhanced feature representations with the baseline data; Construct a database based on the historical nutritional management examples of stroke patients after surgery, introduce a graph learning strategy, use a graph convolutional neural network to learn and represent the graph structure corresponding to the multi-dimensional nutritional health features, extract the nutritional health portraits of patients, and evaluate the nutritional health status of patients; Extract personalized nutritional management strategies from the database according to the evaluation results of the patients' nutritional health status for recommendation, monitor the rehabilitation effects and changes in the nutritional health portraits of patients, and dynamically adjust the nutritional management strategies.

[0006] In this solution, obtaining the basic physiological indicators, existing examination data, and drug use records of ICU surviving stroke patients after surgery according to the medical records specifically includes: Obtain the identity information of ICU surviving stroke patients after surgery, generate a unique identifier according to the identity information, construct a retrieval tag by combining the unique identifier with a preset time range, and use the retrieval tag to obtain the corresponding medical record data; Use the basic physiological indicators, existing examination data, and drug use records as key information to generate structured fields and unstructured descriptions, retrieve based on the structured fields and unstructured descriptions in the medical record data, use similarity calculation to obtain specific types of data, and generate subsets of medical record data corresponding to each key information; Perform data cleaning and data desensitization on the medical record data in different subsets of medical record data, remove duplicate data and patient identity information and handle missing values, then convert the data from different sources into a unified format through data standardization processing and data integration, and perform data association to integrate into a complete patient record, realizing the preprocessing of patient medical record data.

[0007] In this solution, obtaining the text representations corresponding to the basic physiological indicators, existing examination data, and drug use records of patients specifically includes: Obtain the preprocessed medical record data subset corresponding to the patient's basic physiological indicators, existing examination data, and drug use records. Initialize the RoBerta pre-trained model, BERT pre-trained model, and CNN pre-trained model, and import each medical record data subset into the RoBerta pre-trained model, BERT pre-trained model, and CNN pre-trained model for word embedding; Obtain the text representations generated by the medical record data subset in different pre-trained models, import the text representations into the fully connected layer to obtain the corresponding implicit representations, introduce the attention mechanism, obtain the attention scores of each implicit representation, and obtain the attention weights after normalization processing; Use the attention weights to fuse the text representations generated by different pre-trained models to construct a deep text representation, and respectively obtain the basic feature representation, examination feature representation, and drug feature representation corresponding to the basic physiological indicators, existing examination data, and drug use records.

[0008] In this solution, call the relevant domain knowledge graph to enhance the features of the feature representations, and use the knowledge-enhanced feature representations combined with the baseline data to construct multi-dimensional nutrition and health features. Specifically: According to the professional knowledge base and corpus in the field of post-stroke rehabilitation, link to the corresponding domain knowledge graph, use the obtained basic feature representation, examination feature representation, and drug feature representation to locate in the domain knowledge graph, read and mark the entity nodes in the domain knowledge graph, and obtain the domain knowledge graph embedding representation; Based on the domain knowledge graph embedding representation, perform random walks in the domain knowledge graph embedding. Starting from the patient as the starting entity and ending with the rehabilitation nutrition and health status as the ending entity, sample the paths containing the marked entity nodes from the starting entity to the ending entity in the domain knowledge graph; In the obtained path set, splice the entities corresponding to the patient's abnormal indicators, use the Bi-LSTM network to encode the spliced paths to obtain the corresponding embedding matrix, use the attention mechanism to generate a weight vector for the embedding matrix, and obtain the weight of each path; Preset a weight threshold, compare the weight of each path with the weight threshold, filter the paths greater than the weight threshold, and read the feature representations and feature representation combinations involved in the filtered paths; Perform feature extraction through the read feature representations and feature representation combinations to achieve knowledge enhancement of feature selection, and combine the baseline data of the patient's nutrition and health assessment with the extracted features to construct the patient's multi-dimensional nutrition and health features.

[0009] In this solution, a database is constructed based on the historical nutrition management cases of stroke patients after rehabilitation. A graph learning strategy is introduced, and a graph convolutional neural network is used to learn and represent the corresponding graph structure of multi-dimensional nutritional health features, and the nutritional health portrait of the patient is extracted. Specifically: Retrieve the historical nutrition management cases of stroke patients after rehabilitation according to the medical record data. After preprocessing the historical nutrition management cases, integrate and store them in the database. Introduce a graph learning strategy to perform global graph representation on the constructed database. Consider patients and nutritional health status as nodes, and construct an edge structure according to the interaction relationship between the nodes; Take the multi-dimensional nutritional health features of the patient as additional features of the patient node. Use the additional features to calculate similarity to obtain similar patients. Based on the similar patients, obtain neighborhood nodes, and construct a patient subgraph and a nutritional health status subgraph respectively according to different types of neighborhood nodes; Based on the graph convolutional neural network, learn the global graph and the patient subgraph respectively. In the global graph, use the message propagation and neighbor aggregation mechanisms of the graph convolutional neural network to obtain the global embedding representation of the patient and the global embedding representation of the nutritional health status; Obtain the local embedding representation of the patient in the patient subgraph. Perform adaptive weighting on the global embedding representation of the patient and the local embedding representation of the patient, splice the weighted embedding representations, and obtain the final embedding representation of the patient to construct the nutritional health portrait of the patient.

[0010] In this solution, the nutritional health portrait of the patient is used to evaluate the nutritional health status of the patient. Specifically: Use the graph convolutional neural network to obtain the global embedding representation of the nutritional health status and the local embedding representation of the nutritional health status from the global graph and the nutritional health status subgraph respectively. Perform adaptive weighting and splicing on the global embedding representation of the nutritional health status and the local embedding representation of the nutritional health status to obtain the final embedding representation of the nutritional health status; Calculate the inner product of the final embedding representation of the patient corresponding to the nutritional health portrait of the patient and the final embedding representation of the nutritional health status. According to the inner product, screen the nutritional health status nodes that meet the preset threshold, and take the average of the qualified nutritional health status nodes to output the evaluation result of the patient's nutritional health status.

[0011] In this solution, according to the evaluation result of the patient's nutritional health status, extract personalized nutrition management strategies in the database for recommendation, monitor the rehabilitation effect and the change of the nutritional health portrait of the patient, and dynamically adjust the nutrition management strategy. Specifically: Among the similar patients screened based on the multi-dimensional nutritional health features of the patient, use the evaluation result of the patient's nutritional health status for secondary fine screening, and obtain the corresponding nutrition management strategy according to the result of the secondary screening as the nutrition management strategy to be recommended; Obtain the dietary habit characteristics and postoperative rehabilitation characteristics of the patient, obtain the similarity between the patient and similar patients according to the dietary habit characteristics and postoperative rehabilitation characteristics, set different behavior weights for different dietary habit characteristics and postoperative rehabilitation characteristics, and accumulate the different behavior weights of similar patients to generate the adaptation degree between the similar patients and the to-be-recommended nutrition management strategy; Multiply and accumulate the similarity between the patient and similar patients and the adaptation degree of the similar patients to the to-be-recommended nutrition management strategy to obtain the recommendation weight of the nutrition management strategy, and select the nutrition management strategy that meets the preset requirements for recommendation to the patient; Adapting the similar patients through monitoring the patient's rehabilitation effect and nutritional health portrait, and updating the patient's nutrition management strategy according to the adapted similar patients.

[0012] The second aspect of the present invention provides a nutrition management strategy recommendation system for ICU surviving stroke patients after surgery, and the system includes: a patient data acquisition unit, a patient data pre-training characterization unit, a patient nutritional health assessment unit, a nutrition management recommendation unit, and a nutrition management feedback unit; The patient data acquisition unit obtains the basic physiological indicators, existing examination data, and drug use records of ICU surviving stroke patients after surgery according to the medical record files, and performs data preprocessing; The patient data pre-training characterization unit generates text characterizations of the patient's basic physiological indicators, existing examination data, and drug use records, and calls the relevant domain knowledge graph to perform knowledge enhancement on the examination feature characterizations and drug feature characterizations obtained through the text characterizations; The patient nutritional health assessment unit uses the knowledge-enhanced examination feature characterizations and drug feature characterizations to combine with the baseline data to construct multi-dimensional nutritional health features, constructs a database according to the historical nutrition management instances of stroke patients after surgery, introduces a graph learning strategy, and uses a graph convolutional neural network to learn and represent the corresponding graph structure of the multi-dimensional nutritional health features, and extracts the patient's nutritional health portrait to realize the patient's nutritional health assessment; The nutrition management recommendation unit extracts personalized nutrition management strategies from the database for recommendation according to the evaluation results of the patient's nutritional health status, and sends and displays the obtained personalized nutrition management strategies to medical staff and patients in a preset manner; The nutrition management feedback unit monitors the changes in the patient's rehabilitation effect and nutritional health portrait, and generates feedback information according to the changes in the rehabilitation effect and nutritional health portrait to dynamically adjust the nutrition management strategy.

[0013] Compared with the prior art, the beneficial effects of the present disclosure are: The present invention provides a convenient and comprehensive nutritional health assessment method for ICU survivors after stroke surgery, forms a nutritional health evaluation tool, and characterizes the patient's body condition based on existing examination data combined with a large amount of rehabilitation nutrition management knowledge, providing a data basis for subsequent nutritional health management strategy recommendations.

[0014] The present invention uses a graph convolutional neural network to achieve in-depth analysis and intelligent recommendation of multi-dimensional nutritional health features and nutritional management plans, and promotes the improvement of the neurological and living abilities of stroke rehabilitation patients by formulating individualized nutritional management, so as to achieve the goal of improving the quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments or exemplifications of the present invention, the following will briefly introduce the drawings required for use in the embodiments or exemplifications. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the drawings shown.

[0016] Figure 1 Shows the flowchart of the nutritional management strategy recommendation method for ICU survivors after stroke surgery according to the present invention; Figure 2 Shows the flowchart of enhancing knowledge of feature representation by invoking a knowledge graph in related fields; Figure 3 Shows the flowchart of extracting and recommending nutritional management strategies from a database according to the patient's nutritional health status assessment results; Figure 4 Shows the block diagram of the nutritional management strategy recommendation system for ICU survivors after stroke surgery according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to better understand the above-mentioned objects, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0019] Figure 1 Shows the flowchart of the nutritional management strategy recommendation method for ICU survivors after stroke surgery according to the present invention.

[0020] Such asFigure 1 As shown in Figure 1 , in the first embodiment of the present invention, a method for recommending a nutritional management strategy for ICU surviving stroke patients after surgery is provided, including: S102, obtaining the basic physiological indicators, existing examination data, and drug use records of ICU surviving stroke patients after surgery according to the medical records; S104, using the basic physiological indicators to obtain basic feature representations to provide baseline data for the nutritional health assessment of patients, generating examination feature representations and drug feature representations using the existing examination data and drug use records, and invoking a knowledge graph in the relevant field to enhance the feature representations. Using the knowledge-enhanced feature representations combined with the baseline data to construct multi-dimensional nutritional health features; S106, constructing a database based on the historical nutritional management cases of stroke patients after surgery, introducing a graph learning strategy, using a graph convolutional neural network to learn and represent the graph structure corresponding to the multi-dimensional nutritional health features, extracting the nutritional health portraits of patients, and evaluating the nutritional health status of patients; S108, extracting personalized nutritional management strategies from the database according to the evaluation results of the patients' nutritional health status for recommendation, monitoring the rehabilitation effects and changes in the nutritional health portraits of patients, and dynamically adjusting the nutritional management strategies.

[0021] It should be noted that the identity information of ICU surviving stroke patients after surgery is obtained, a unique identifier is generated according to the identity information, a retrieval label is constructed by combining the unique identifier with a preset time range, and the corresponding medical record data is obtained using the retrieval label; the basic physiological indicators usually include height, weight, blood pressure, heart rate, body temperature, and body fat measurement indicators, usually including skinfold thickness, upper arm muscle circumference, calf circumference, etc. The existing examination data includes laboratory examination and imaging examination results, extracting the laboratory examination results of the patient's blood routine, urine routine, biochemical indicators (such as blood glucose, blood lipid, liver function, etc.), interleukin 6, prealbumin, blood ammonia, etc., and obtaining the patient's imaging examination reports, such as CT, MRI, X-ray, etc. The drug use records include prescription drugs, drug dosages, drug use times, etc. Extract the patient's drug prescriptions and drug use records. For over-the-counter drugs or drugs taken by the patient himself / herself, they can be supplemented through the patient's self-report or nursing records.

[0022] Generate structured fields and unstructured descriptions using basic physiological indicators, existing examination data, and medication records as key information. The unstructured descriptions include chart styles, corresponding positions in examination reports, etc. Retrieve in the medical record data based on the structured fields and unstructured descriptions, use similarity calculation to obtain specific types of data, and generate subsets of medical record data corresponding to each key information; perform data cleaning and data desensitization on the medical record data in different subsets of medical record data, remove duplicate data and patient identity information and handle missing values, then convert data from different sources into a unified format through data standardization processing and data integration, and perform data association to integrate into a complete patient record, realizing the preprocessing of patient medical record data.

[0023] It should be noted that different deep learning models are good at classifying different types of text data. Due to the multi-modal nature of medical record data, fusing the text representations generated by different deep learning models can leverage the recognition advantages of each model and effectively improve the model performance. Obtain subsets of preprocessed medical record data corresponding to the patient's basic physiological indicators, existing examination data, and medication records, initialize the RoBerta pre-trained model, BERT pre-trained model, and CNN pre-trained model, and import each subset of medical record data into the RoBerta pre-trained model, BERT pre-trained model, and CNN pre-trained model for word embedding; obtain the text representations generated by the subset of medical record data in different pre-trained models, import the text representations into the fully connected layer to obtain the corresponding implicit representations, introduce the attention mechanism, obtain the attention scores of each implicit representation, and obtain the attention weights after normalization processing; use the attention weights to fuse the text representations generated by different pre-trained models to construct a deep text representation, and respectively obtain the basic feature representation, examination feature representation, and medication feature representation corresponding to the basic physiological indicators, existing examination data, and medication records.

[0024] Figure 2 The flowchart showing the knowledge enhancement of feature representations by invoking the knowledge graph in the relevant field is shown.

[0025] According to the embodiments of the present invention, the knowledge graph in the relevant field is invoked to perform knowledge enhancement on the feature representations, and the knowledge-enhanced feature representations are combined with the baseline data to construct multi-dimensional nutritional health features, specifically: S202, link the knowledge graph in the corresponding field according to the professional knowledge base and corpus in the field of post-stroke rehabilitation, use the obtained basic feature representation, examination feature representation, and medication feature representation to locate in the knowledge graph in the field, read and mark the entity nodes in the knowledge graph in the field, and obtain the knowledge graph embedding representation; S204. Based on the domain knowledge graph embedding representation, perform random walks in the domain knowledge graph embedding, starting from the patient as the starting entity and ending with the rehabilitation nutritional health status as the ending entity, and sample the paths containing the labeled entity nodes from the starting entity to the ending entity in the domain knowledge graph; S206. Concatenate the entities corresponding to the abnormal indicators of the patient in the obtained path set, use the Bi-LSTM network to encode the concatenated path, obtain the corresponding embedding matrix, use the attention mechanism to generate a weight vector for the embedding matrix, and obtain the weight of each path; S208. Preset a weight threshold, compare the weight of each path with the weight threshold, filter the paths greater than the weight threshold, and read the feature representations and combinations of feature representations involved in the filtered paths; S210. Perform feature extraction through the read feature representations and combinations of feature representations to achieve knowledge enhancement of feature selection, combine the baseline data of the patient's nutritional health assessment with the extracted features, and construct the multi-dimensional nutritional health features of the patient.

[0026] It should be noted that each path sampled based on random walks is a path set starting from the patient as the starting entity and ending with the rehabilitation nutritional health status as the ending entity. Only the information between the starting entity and the ending entity is concerned in the path set, and the impact of the patient's abnormal medical indicators on the recommendation result of the nutritional management strategy is ignored. The patient's abnormal medical indicators need to be focused on, as they determine to a certain extent the nutritional needs in the patient's nutritional management strategy. Therefore, concatenate the entities corresponding to the abnormal indicators of the patient in the obtained path set to improve the accuracy of the recommendation. In addition, compare the weight of each path with the weight threshold, filter the paths greater than the weight threshold, and read the feature representations and combinations of feature representations involved in the filtered paths, which realizes the feature optimization of the patient's existing examination data. In addition, based on the relevant domain knowledge graph, obtain the commonly used combinations of feature representations for nutritional health assessment, construct new evaluation feature indicators, and realize the knowledge enhancement of feature representations in nutritional health assessment to improve the evaluation performance of nutritional health assessment.

[0027] Retrieve the historical nutritional management instances of stroke patients after surgery according to the medical record data, preprocess the historical nutritional management instances and integrate them into the database, introduce a graph learning strategy to perform a global graph representation on the constructed database, and define the global as a set of nodes, is a set of edge structures. The patient and nutritional health status are used as nodes, and edge structures are constructed according to the interaction relationships between the nodes. The multidimensional nutritional health characteristics of the patient are used as additional features of the patient node. Similar patients are obtained by calculating similarity using the additional features. Neighborhood nodes are obtained based on the similar patients. Patient subgraphs and nutritional health status subgraphs are constructed respectively according to different types of domain nodes. Based on the graph convolutional neural network, the global graph and patient subgraphs are learned respectively. In the global graph, the message propagation and neighbor aggregation mechanisms of the graph convolutional neural network are used to obtain the global embedding representations of the patient and nutritional health status; where the patient node is represented by the neighbor node aggregation mechanism to obtain the global embedding representation of the patient Specifically: ; Among them, represents the activation function, represents the parameter matrix of feature transformation, is used as the initial feature vector of the patient and nutritional health status nodes, represents the weight matrix; similarly, the global embedding representation of the nutritional health status ; In the patient subgraph, the local embedding representation of the patient is obtained. The global embedding representation of the patient and the local embedding representation of the patient are adaptively weighted, and the weighted embedding representations are concatenated to obtain the final embedding representation of the patient to construct the nutritional health portrait of the patient. The global embedding representation of the nutritional health status and the local embedding representation of the nutritional health status are obtained respectively from the global graph and the nutritional health status subgraph using the graph convolutional neural network. The global embedding representation of the nutritional health status and the local embedding representation of the nutritional health status are adaptively weighted and concatenated to obtain the final embedding representation of the nutritional health status; the inner product of the final embedding representation of the patient corresponding to the patient's nutritional health portrait and the final embedding representation of the nutritional health status is calculated, and the nutritional health status nodes that meet the preset threshold are screened according to the inner product, and the nutritional health status nodes that meet the requirements are averaged to output the evaluation result of the patient's nutritional health status.

[0028] Figure 3 shows a flowchart of extracting and recommending nutritional management strategies from the database according to the evaluation result of the patient's nutritional health status.

[0029] According to an embodiment of the present invention, personalized nutritional management strategies are extracted and recommended from the database according to the evaluation result of the patient's nutritional health status, and the rehabilitation effect and changes in the nutritional health portrait of the patient are monitored, and the nutritional management strategies are dynamically adjusted. Specifically: S302. Among the similar patients screened based on the multi-dimensional nutritional health characteristics of the patient, use the evaluation results of the patient's nutritional health status for secondary fine screening, and obtain the corresponding nutritional management strategy as the nutritional management strategy to be recommended according to the results of the secondary screening; S304. Obtain the eating habit characteristics and postoperative rehabilitation characteristics of the patient, obtain the similarity between the patient and the similar patients according to the eating habit characteristics and postoperative rehabilitation characteristics, set different behavior weights for different eating habit characteristics and postoperative rehabilitation characteristics, and accumulate the different behavior weights of the similar patients to generate the adaptation degree between the similar patients and the nutritional management strategy to be recommended; S306. Multiply and accumulate the similarity between the patient and the similar patients by the adaptation degree of the similar patients to the nutritional management strategy to be recommended to obtain the recommendation weight of the nutritional management strategy, and select the nutritional management strategy that meets the preset requirements for the patient for recommendation; S308. Adaptively change the similar patients by monitoring the patient's rehabilitation effect and nutritional health portrait, and update the patient's nutritional management strategy according to the adaptively changed similar patients.

[0030] It should be noted that the types and amounts of food consumed daily are recorded through questionnaires or diet diaries, and the daily intake of calories, protein, fat, carbohydrates, vitamins and minerals is calculated to understand the patient's dietary preferences (such as vegetarian, low-fat diet) and taboos (such as allergic foods) to evaluate the patient's dietary habits. The postoperative rehabilitation characteristics of the patient are evaluated by obtaining the patient's swallowing function, metabolic function, gastrointestinal function and mobility. In order to obtain the degree of adaptation of the patient to the nutritional management strategy, different behavioral weights are set for different dietary habit characteristics and postoperative rehabilitation characteristics. The initial weight is preset by obtaining the Pearson correlation coefficient between different dietary habit characteristics and postoperative rehabilitation characteristics and nutritional needs. The quantitative coefficient is generated according to whether the nutritional management strategy contains the corresponding management strategy and the feedback of similar patients on the management strategy. The quantitative coefficient is used to weight the preset initial weight to generate the behavioral weight, and the behavioral weight is accumulated to obtain the degree of adaptation of the patient to the nutritional management strategy. To calculate the degree of adaptation of the target patient to the recommended nutritional management strategy, the similarity between the target patient and similar patients should be used as the weight, and the degree of adaptation of similar patients to the nutritional management strategy should be multiplied and accumulated as the final result. Monitor the patient's rehabilitation effect and nutritional health portrait through regular follow-up (such as once a week). Rehabilitation assessment includes NIHSS scale score, brief mental state examination scale, Kubota water drinking test, daily living ability assessment, etc. Re-screen similar patients based on the patient's rehabilitation effect and changes in nutritional health portrait, and use the re-acquired patient nutritional health status assessment results for secondary fine screening to achieve the update of the recommended nutritional management strategy. At the same time, evaluate the effect of the nutritional management strategy through data analysis combined with real-time feedback and suggestions from patients, obtain the adjustment direction of the patient's nutritional management strategy based on the evaluation effect, and adjust the plan in time based on the adjustment direction. In addition, analyze the patient's nutritional health status based on the evaluation results of the nutritional management strategy, and predict the patient's nutritional risk and rehabilitation trend.

[0031] Figure 4 A block diagram of the nutritional management strategy recommendation system for ICU survivors undergoing postoperative rehabilitation is shown.

[0032] The second aspect of the present invention provides a nutritional management strategy recommendation system 4 for ICU survivors undergoing postoperative rehabilitation of stroke, the system comprising: a patient data collection unit 401, a patient data pre-training representation unit 402, a patient nutritional health assessment unit 403, a nutritional management recommendation unit 404 and a nutritional management feedback unit 405; The patient data collection unit 401 obtains the basic physiological indicators, existing examination data and drug use records of ICU survivors of stroke surgery and rehabilitation according to the medical records, and performs data preprocessing; The patient data pre-training representation unit 402 generates text representations of the patient's basic physiological indicators, existing examination data, and medication usage records, and invokes a knowledge graph in the relevant field to enhance the knowledge of the examination feature representation and medication feature representation obtained through the text representation; The patient nutritional health assessment unit 403 uses the knowledge-enhanced examination feature representation and medication feature representation to combine with the baseline data to construct multi-dimensional nutritional health features, constructs a database based on the historical nutritional management instances of stroke patients after surgery, introduces a graph learning strategy, and uses a graph convolutional neural network to learn and represent the graph structure corresponding to the multi-dimensional nutritional health features, and extracts the nutritional health portrait of the patient to achieve the nutritional health assessment of the patient; The nutritional management recommendation unit 404 extracts personalized nutritional management strategies from the database according to the evaluation results of the patient's nutritional health status for recommendation, and sends and displays the obtained personalized nutritional management strategies to medical staff and patients in a preset manner; The nutritional management feedback unit 405 monitors the rehabilitation effect and the change of the nutritional health portrait of the patient, and generates feedback information according to the change of the rehabilitation effect and the nutritional health portrait to dynamically adjust the nutritional management strategy.

[0033] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for the nutritional management strategy recommendation method for ICU surviving stroke patients after surgery. When the program for the nutritional management strategy recommendation method for ICU surviving stroke patients after surgery is executed by a processor, the steps of the nutritional management strategy recommendation method for ICU surviving stroke patients after surgery are implemented.

[0034] In several embodiments provided in the present application, it should be understood that the disclosed method can be implemented in other ways. The above-described embodiments are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed with each other can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical or other forms.

[0035] Those of ordinary skill in the art will understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments; and the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0036] Alternatively, if the above integrated unit of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, ROM, RAM, magnetic disks, or optical discs.

[0037] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation, characterized in that: The following steps are involved: Obtain basic physiological indicators, existing examination data, and drug use records of postoperative stroke survivors in the ICU based on medical records; Use basic physiological indicators to obtain basic feature representations to provide baseline data for patient nutrition and health assessment, use the existing examination data and drug use records to generate examination feature representations and drug feature representations, call related field knowledge graphs to enhance the feature representations, and use the knowledge-enhanced feature representations combined with baseline data to construct multidimensional nutrition and health features; A database was built based on historical nutritional management examples of stroke postoperative rehabilitation patients. A graph learning strategy was introduced. A graph convolutional neural network was used to learn and represent the graph structure corresponding to multi-dimensional nutritional health characteristics, extract the nutritional health portrait of the patient, and evaluate the nutritional health status of the patient. Based on the patient's nutritional health status assessment results, personalized nutritional management strategies are extracted from the database for recommendation, the patient's rehabilitation effect and changes in nutritional health profile are monitored, and the nutritional management strategy is dynamically adjusted.

2. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 1, characterized in that: The basic physiological indicators, existing examination data and drug use records of ICU survivors of stroke postoperative rehabilitation were obtained according to the medical records, specifically: Obtain the identity information of ICU survivors of stroke and postoperative rehabilitation, generate a unique identifier based on the identity information, construct a search tag by combining the unique identifier with a preset time range, and use the search tag to obtain the corresponding medical record data; Using basic physiological indicators, existing examination data and drug use records as key information to generate structured fields and unstructured descriptions, searching the medical record data based on the structured fields and unstructured descriptions, using similarity calculation to obtain data of a specific type, and generating a subset of medical record data corresponding to each key information; The medical record data in different medical record data subsets are cleaned and desensitized to remove duplicate data and patient identity information and process missing values. Then, data from different sources are converted into a unified format through data standardization and data integration, and data association is performed to integrate them into complete patient records to achieve preprocessing of patient medical record data.

3. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 1, characterized in that: Obtain the text representations of the patient's basic physiological indicators, existing examination data, and drug use records, specifically: Obtain a preprocessed subset of medical record data corresponding to the patient's basic physiological indicators, existing examination data, and drug use records, initialize the RoBerta pre-training model, the BERT pre-training model, and the CNN pre-training model, and import each subset of medical record data into the RoBerta pre-training model, the BERT pre-training model, and the CNN pre-training model for word embedding; Obtain text representations generated by different pre-trained models for subsets of medical record data, import the text representations into the fully connected layer to obtain the corresponding implicit representations, introduce the attention mechanism, obtain the attention scores of each implicit representation, and obtain the attention weights after normalization; The attention weights are used to fuse the text representations generated by different pre-trained models to construct a deep text representation, and the basic feature representations, examination feature representations and drug feature representations corresponding to the basic physiological indicators, existing examination data and drug use records are obtained respectively.

4. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 1, characterized in that: The knowledge graph of related fields is called to enhance the feature representation, and the feature representation after knowledge enhancement is combined with the baseline data to construct multi-dimensional nutrition and health features, specifically: According to the professional knowledge base and corpus in the field of post-stroke rehabilitation, the corresponding domain knowledge graph is linked, the basic feature representation, examination feature representation and drug feature representation obtained are used to locate in the domain knowledge graph, the entity nodes in the domain knowledge graph are read and marked, and the domain knowledge graph embedding representation is obtained; Based on the domain knowledge graph embedding representation, a random walk is performed in the domain knowledge graph embedding, with the patient as the starting entity and the rehabilitation nutrition health status as the ending entity, and a path containing marked entity nodes from the starting entity to the ending entity in the domain knowledge graph is sampled; In the acquired path set, entities corresponding to abnormal indicators of the patient are spliced, a Bi-LSTM network is used to encode the spliced ​​paths, a corresponding embedding matrix is ​​acquired, and a weight vector is generated for the embedding matrix using an attention mechanism to acquire the weight of each path; Preset a weight threshold, compare the weight of each path with the weight threshold, filter out paths that are greater than the weight threshold, and read the feature representations and feature representation combinations involved in the filtered paths; Feature extraction is performed by reading feature representations and feature representation combinations to achieve knowledge enhancement of feature selection. The baseline data of the patient's nutritional health assessment is combined with the extracted features to construct the patient's multidimensional nutritional health characteristics.

5. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 1, characterized in that: Based on the historical nutritional management examples of stroke postoperative rehabilitation patients, a database was constructed, and a graph learning strategy was introduced. A graph convolutional neural network was used to learn and represent the corresponding graph structure of multi-dimensional nutritional health characteristics, and to extract the nutritional health portrait of the patient, specifically: The historical nutritional management examples of stroke postoperative rehabilitation patients were retrieved based on medical record data, and the historical nutritional management examples were pre-processed and integrated into the database. The graph learning strategy was introduced to represent the constructed database as a global graph, with patients and nutritional health status as nodes, and edge structures were constructed based on the interactive relationships between nodes. The multidimensional nutritional health characteristics of the patient are used as additional features of the patient node, and similarity calculation is performed using the additional features to obtain similar patients, and neighborhood nodes are obtained based on the similar patients, and patient subgraphs and nutritional health status subgraphs are constructed according to different types of domain nodes; Based on the graph convolutional neural network, the global graph and the patient subgraph are learned respectively, and the message propagation and neighbor aggregation mechanism of the graph convolutional neural network are used in the global graph to obtain the global embedding representation of the patient and the global embedding representation of the nutritional health status; The local embedded representation of the patient is obtained in the patient subgraph, the global embedded representation of the patient and the local embedded representation of the patient are adaptively weighted, the weighted embedded representations are spliced, and the final embedded representation of the patient is obtained to construct the nutritional health portrait of the patient.

6. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 5, characterized in that: Use the patient's nutritional health profile to assess the patient's nutritional health status, specifically: Use graph convolutional neural network to obtain global embedding representation of nutritional health status and local embedding representation of nutritional health status from global graph and nutritional health status subgraph respectively, perform adaptive weighted splicing on global embedding representation of nutritional health status and local embedding representation of nutritional health status, and obtain final embedding representation of nutritional health status; Calculate the inner product of the patient's final embedded representation corresponding to the patient's nutritional health portrait and the final embedded representation of the nutritional health status, filter the nutritional health status nodes that meet the preset threshold according to the inner product, average the nutritional health status nodes that meet the requirements, and output the patient's nutritional health status assessment result.

7. The method for recommending nutritional management strategies for ICU survivors of stroke surgery and rehabilitation according to claim 1, characterized in that: According to the evaluation results of the patient's nutritional health status, the personalized nutritional management strategy is extracted from the database for recommendation, the patient's rehabilitation effect and nutritional health profile changes are monitored, and the nutritional management strategy is dynamically adjusted, specifically: Among similar patients screened based on the multidimensional nutritional health characteristics of the patients, secondary fine screening is performed using the patient nutritional health status assessment results, and the corresponding nutritional management strategy is obtained as the nutritional management strategy to be recommended based on the results of the secondary screening; Obtain the patient's dietary habit characteristics and postoperative rehabilitation characteristics, obtain the similarity between the patient and similar patients based on the dietary habit characteristics and postoperative rehabilitation characteristics, set different behavioral weights for different dietary habit characteristics and postoperative rehabilitation characteristics, accumulate different behavioral weights of similar patients, and generate the degree of adaptation between similar patients and the nutritional management strategy to be recommended; The recommendation weight of the nutrition management strategy is obtained by multiplying the similarity between the patient and similar patients and the degree of adaptation of similar patients to the recommended nutrition management strategy, and then selecting the nutrition management strategy that meets the preset requirements for recommendation to the patient; By monitoring the patient's rehabilitation effects and nutritional health portraits, adaptive changes are made to similar patients, and the patient's nutritional management strategy is updated based on similar patients after adaptive changes.

8. A nutritional management strategy recommendation system for ICU survivors of stroke post-operative rehabilitation patients, used to implement the nutritional management strategy recommendation method for ICU survivors of stroke post-operative rehabilitation patients as described in any one of claims 1 to 7, the system comprising: Patient data collection unit, patient data pre-training representation unit, patient nutrition and health assessment unit, nutrition management recommendation unit and nutrition management feedback unit; The patient data collection unit obtains basic physiological indicators, existing examination data and drug use records of ICU survivors of stroke surgery and rehabilitation according to medical records, and performs data preprocessing; The patient data pre-training representation unit generates text representations of the patient's basic physiological indicators, existing examination data, and drug use records, and calls the relevant field knowledge graph to perform knowledge enhancement on the examination feature representation and drug feature representation obtained through the text representation; The patient nutrition and health assessment unit uses the knowledge-enhanced examination feature representation and drug feature representation combined with baseline data to construct a multidimensional nutrition and health feature, builds a database based on historical nutrition management examples of post-stroke rehabilitation patients, introduces a graph learning strategy, uses a graph convolutional neural network to learn and represent the graph structure corresponding to the multidimensional nutrition and health feature, and extracts the patient's nutrition and health portrait to evaluate the patient's nutrition and health status; The nutrition management recommendation unit extracts personalized nutrition management strategies from the database according to the patient's nutrition and health status assessment results and recommends them, and sends and displays the obtained personalized nutrition management strategies to medical staff and patients in a preset manner; The nutrition management feedback unit monitors the patient's rehabilitation effect and changes in the nutrition and health portrait, and generates feedback information according to the changes in the rehabilitation effect and the nutrition and health portrait to dynamically adjust the nutrition management strategy.

Images