Liver cirrhosis patient diet recommendation method, system and equipment based on machine learning, medium and computer program

Abstract

The invention relates to a liver cirrhosis patient diet recommendation method, system and device based on machine learning, a medium and a computer program.The recommendation method comprises the steps that features are extracted through a graph convolutional network, a multi-dimensional patient feature map is constructed, and a multi-task deep reinforcement learning framework is established based on the feature map; the nutritional requirements of the patient are quantitatively analyzed, the complication risk is dynamically evaluated, and the diet preference mode is identified. Data distribution offset is monitored through an online incremental learning module, model parameters are updated through an adversarial domain adaptation algorithm, and dynamic feature drift compensation is achieved. And fusing the output information of the multi-task framework and the compensated data, generating a hierarchical diet decision tree, carrying out Monte Carlo tree search on a food material knowledge graph based on a patient state vector, and balancing nutrition and experience in combination with a reward mechanism to obtain a diet recommendation decision. Compared with the prior art, a personalized diet scheme can be provided for the patients with cirrhosis, and the treatment effect and the life quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical big data application technology, and in particular to a method, system, device, medium, and computer program for recommending diet for patients with cirrhosis based on machine learning. Background Art

[0002] In the field of healthcare technology, especially in nutritional support services for patients with specific diseases (such as critically ill patients and oral cancer patients), existing technologies have made certain progress.

[0003] For example, the reference patent CN118262872A proposes a nutritional support service system and method for critically ill patients. Through modules such as physiological data analysis, electrolyte balance regulation, and appetite management, combined with data fusion technology, principal component analysis, and bioelectrical impedance signal processing technology, it aims to improve data processing accuracy and the ability to predict changes in nutritional needs.

[0004] However, this program mainly focuses on critically ill patients, and has shortcomings in long-term dietary management and personalized demand response for patients with chronic diseases such as cirrhosis, especially in dynamically monitoring changes in patient status and adaptability to dietary preferences, failing to provide more precise personalized services.

[0005] Another reference patent, CN118248284A, focuses on the nutritional management of oral cancer patients after surgery. By constructing a malnutrition risk prediction model, a diet transition time prediction model, and a comprehensive diet assessment strategy, combined with user portraits and knowledge graph technology, it provides personalized nutritional guidance and diet recommendations.

[0006] However, this program is mainly aimed at patients with oral cancer, and its comprehensive and dynamic support for dietary management of patients with cirrhosis still needs to be improved in terms of processing multimodal data streams, real-time monitoring of changes in patients' physiological indicators, and long-term prediction of nutritional needs through deep learning models.

[0007] The main defects of the existing technology include:

[0008] Limitations of target population: Existing systems are mostly designed for acute illnesses or postoperative recovery scenarios. They lack a continuous optimization mechanism for long-term nutritional management of patients with chronic diseases such as cirrhosis and are unable to adapt to dynamic changes in the disease course.

[0009] Insufficient multimodal data integration: Traditional methods fail to effectively mine the topological correlation features between cross-modal data when integrating heterogeneous information such as physiological indicator time series data and unstructured text, resulting in insufficient comprehensiveness in the representation of patient status.

[0010] Lack of dynamic adaptability: Static models are unable to cope with the dynamic drift of patient data distribution (such as changes in physiological indicators caused by disease progression). They lack online incremental learning and feature drift compensation mechanisms, leading to the accumulation of recommendation biases.

[0011] Simplified decision-making mechanism: Existing recommendation strategies mostly focus on a single goal (such as achieving nutritional standards), fail to balance the conflicting demands of nutritional optimization and improved dining experience, and lack dynamic decision-making capabilities under complex constraints.

[0012] These issues limit the effectiveness of existing technologies in dietary management for patients with cirrhosis. Addressing these technical bottlenecks, we urgently need to propose an improved solution that integrates dynamic modeling of multimodal data with reinforcement learning decision-making to enhance the accuracy and long-term adaptability of personalized dietary recommendations. Summary of the Invention

[0013] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and to provide a diet recommendation method, system, device, medium, and computer program for patients with cirrhosis based on machine learning, which can provide patients with a more personalized and real-time responsive diet management plan, in order to improve the nutritional status of patients, reduce the risk of complications, and optimize the dietary experience.

[0014] The purpose of the present invention can be achieved by the following technical solutions:

[0015] A first aspect of the present invention provides a method for recommending a diet for patients with cirrhosis based on machine learning, comprising the following steps:

[0016] S100: Constructs a multi-dimensional patient feature map, integrating a multimodal dataset consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation texts in real time through an API interface, and uses a graph convolutional network to extract topological correlation features between multimodal data;

[0017] S200: Based on the multidimensional patient feature map obtained in step S100 and the extracted topological correlation features, a multi-task deep reinforcement learning framework is established, which includes a nutritional needs prediction branch, a complication risk warning branch, and a dietary preference adaptation branch, for analyzing the patient's nutritional needs, complication risks, and dietary preferences;

[0018] S300: During the operation of the multi-task deep reinforcement learning framework, a dynamic feature drift compensation mechanism is implemented, patient data distribution deviation is monitored in real time through an online incremental learning module, and model parameters are updated using an adversarial domain adaptation algorithm;

[0019] S400: Based on the patient's nutritional needs information, complication risk information, and dietary preference information output by the multi-task deep reinforcement learning framework in step S200, and the valid data after dynamic feature drift compensation in step S300, a hierarchical dietary decision tree is generated, and a Monte Carlo tree search is performed on the food knowledge graph based on the current patient state vector. The reward mechanism of reinforcement learning is combined to balance the optimization of nutritional indicators and the improvement of dietary experience, and dietary recommendation decision information is obtained.

[0020] Furthermore, step S100 specifically includes the following sub-steps:

[0021] S101: Access a multimodal dataset consisting of patient physiological indicators, dietary behavior, clinical tests, and doctor-patient conversations in real time through an API interface. After data cleaning, time alignment, and entity association processing, a unified multimodal dataset with a unified timestamp is formed.

[0022] S102: Based on the multimodal unified data set formed in step S101, a patient feature map is constructed that includes multiple types of nodes, including basic patient information, physiological indicators, dietary behavior, test reports, and doctor-patient conversations. The internal correlation relationships of the data are represented by defining association edges between the nodes.

[0023] S103: For the patient feature map constructed in step S102, a graph convolutional network or a relational graph convolutional network is used to perform node neighborhood aggregation operations to extract topological correlation features that reflect the interactive characteristics of multimodal data.

[0024] Furthermore, step S200 specifically includes the following sub-steps:

[0025] S201: Building a multi-task deep reinforcement learning framework based on the multi-dimensional patient feature map and topological correlation features extracted in step S100;

[0026] S202: Integrating three branches, namely, nutritional needs prediction, complication risk warning, and dietary preference adaptation, into the multi-task deep reinforcement learning framework to respectively implement functions of quantitative analysis of patient nutritional needs, dynamic assessment of complication risk, and dietary preference pattern recognition;

[0027] S203: Jointly model the outputs of the three branches in S202, perform cross-branch fusion processing on the nutritional demand prediction results, complication risk assessment indicators, and dietary preference identification features, and output a comprehensive analysis result including the patient's real-time status.

[0028] Furthermore, in step S300, the operation process of the three branches of nutritional demand prediction, complication risk warning, and dietary preference adaptation includes:

[0029] The process of nutritional requirement prediction: Based on the physiological indicator nodes and clinical test nodes in the patient characteristic map constructed in step S102, the gated recurrent unit network of the nutritional requirement prediction branch in step S202 is called, the liver metabolic dynamics model is integrated to generate a nutrient requirement baseline, and the prediction results are input into the cross-branch fusion module in step S203;

[0030] The complication risk warning process includes: constructing a risk feature tensor using the topological correlation features extracted in step S103, and calculating the cross-period risk correlation weights through the multi-head attention mechanism of the complication risk warning branch in step S202;

[0031] The process of dietary preference pattern recognition: parse the unstructured doctor-patient dialogue text integrated in step S101, extract dietary behavior trigger words through semantic role labeling of the dietary preference adaptation branch in step S202, and realize dynamic coupling of preference characteristics and nutritional restriction conditions in step S203 based on the neighborhood node propagation mechanism of the graph convolutional network in step S103.

[0032] Furthermore, step S300 specifically includes the following sub-steps:

[0033] S301: During the execution of the multi-task deep reinforcement learning framework, the online incremental learning module is used to monitor the patient data distribution in real time to identify any deviations.

[0034] S302: When the online incremental learning module detects that the patient data distribution is offset, the adversarial domain adaptation algorithm is used to update the model parameters, thereby realizing a dynamic feature drift compensation mechanism.

[0035] Furthermore, step S400 specifically includes the following sub-steps:

[0036] S401: Based on the comprehensive analysis results of the patient's real-time status output in step S203 and the valid data after dynamic feature drift compensation in step S300, a node weight distribution function of a hierarchical dietary decision tree is constructed, wherein the root node of the decision tree is associated with the nutritional restriction conditions in the patient feature map constructed in step S102, and the child nodes dynamically adjust the food association path based on the topological association features extracted in step S103;

[0037] S402: In the hierarchical dietary decision tree generated in step S401, using the current patient state vector as the initial search starting point, combined with the nutritional requirements, risk probabilities, and preference weights output by each branch in step S202, a Monte Carlo tree search is performed on the food knowledge graph. The exploration and exploitation balance coefficient in the search strategy is dynamically adjusted using the model parameters updated in step S300;

[0038] S403: Based on the candidate diet plans generated by the Monte Carlo tree search in step S402, the reward mechanism of the multi-task deep reinforcement learning framework in step S200 is called, and the nutritional optimization target output by the cross-branch fusion module in step S203 and the dietary preference characteristics calibrated in step S300 are synchronously integrated. The Pareto frontier of nutritional indicators and dietary experience in the recommended plan is iteratively optimized through the gradient ascent algorithm to generate the final diet recommendation decision information.

[0039] A second aspect of the present invention provides a diet recommendation system for patients with cirrhosis based on machine learning, comprising:

[0040] Multimodal feature map construction module: used to construct a multidimensional patient feature map. It integrates multimodal datasets consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation text in real time through API interfaces, and uses graph convolutional networks to extract topological correlation features between multimodal data.

[0041] Multi-task reinforcement learning modeling module: This module is used to establish a multi-task deep reinforcement learning framework based on the obtained multi-dimensional patient feature map and extracted topological correlation features, including a nutritional needs prediction branch, a complication risk warning branch, and a dietary preference adaptation branch. This framework is used to analyze patients' nutritional needs, complication risks, and dietary preferences.

[0042] Dynamic feature drift compensation module: used to implement a dynamic feature drift compensation mechanism during the operation of the multi-task deep reinforcement learning framework, monitor the patient data distribution offset in real time through the online incremental learning module, and update the model parameters using the adversarial domain adaptation algorithm;

[0043] Hierarchical dietary decision generation module: It is used to generate a hierarchical dietary decision tree based on the patient's nutritional needs information, complication risk information, dietary preference information, and valid data after dynamic feature drift compensation output by the multi-task deep reinforcement learning framework. It performs Monte Carlo tree search on the food knowledge graph based on the current patient state vector, and combines the reinforcement learning reward mechanism to balance the optimization of nutritional indicators and the improvement of dietary experience to obtain dietary recommendation decision information.

[0044] A third aspect of the present invention provides an electronic device comprising a memory and a processor, wherein the processor is configured to execute a program in the memory, thereby implementing the above-mentioned method for recommending a diet for patients with cirrhosis based on machine learning.

[0045] A fourth aspect of the present invention provides a storage medium comprising computer-executable instructions, which, when executed by a computer processor, is used to execute the above-mentioned method for recommending a diet for patients with cirrhosis based on machine learning.

[0046] A fifth aspect of the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned method for recommending a diet for patients with cirrhosis based on machine learning.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1) The present invention constructs a multi-dimensional patient feature map and uses a graph convolutional network to extract topological correlation features between multimodal data, providing a more comprehensive and accurate basis for dietary recommendations for patients with cirrhosis. It integrates a variety of data such as physiological indicators, dietary behavior, clinical tests, and doctor-patient dialogues, breaking the limitations of traditional data utilization and fully tapping the potential correlations in the data. The multi-task deep reinforcement learning framework established on this basis includes multiple branches such as nutritional needs prediction, complication risk warning, and dietary preference adaptation. It can simultaneously conduct in-depth analysis of patients' nutritional needs, complication risks, and dietary preferences, achieving a comprehensive consideration of multiple objectives. Compared with single-task models, it can better meet the complex dietary recommendation needs of patients with cirrhosis. In addition, the introduction of the dynamic feature drift compensation mechanism, using the online incremental learning module and the adversarial domain adaptation algorithm, effectively solves the problem of patient data distribution offset, ensuring that the model can continuously and stably output accurate dietary recommendation decision information during long-term use, thereby improving the robustness and adaptability of the entire system.

[0049] 2) From the perspective of practical application, the present invention provides a personalized, scientific and reasonable diet recommendation plan for patients with cirrhosis. Through hierarchical diet decision trees and Monte Carlo tree search, combined with the reward mechanism of reinforcement learning, it can recommend ingredients and diet plans that meet both the nutritional needs and physical conditions of patients and their dietary preferences on the basis of balancing the optimization of nutritional indicators and the improvement of dietary experience, which will help patients better adhere to healthy eating habits, improve treatment effects and quality of life. At the same time, the method can dynamically adjust the diet recommendation plan according to the patient's real-time status, respond to changes in the patient's condition in a timely manner, and provide strong support for the long-term diet management of patients with cirrhosis. For medical staff, the invention also provides an effective auxiliary tool to help them formulate diet plans for patients more efficiently, save time and energy, improve the quality and efficiency of medical services, and have significant social benefits and promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the process of the diet recommendation method for patients with cirrhosis based on machine learning in the present invention.

[0051] Figure 2 Schematic diagram of the process of step 1 in the present invention;

[0052] Figure 3Schematic diagram of the process of step 2 in the present invention;

[0053] Figure 4 Schematic diagram of the process of step 4 in the present invention. DETAILED DESCRIPTION

[0054] Overall, this application constructs a multi-dimensional patient feature map, integrates multimodal data sets such as patient physiological indicators, dietary behaviors, clinical tests, and doctor-patient conversations in real time, and uses graph convolutional networks to extract topological correlation features between data, and then establishes a multi-task deep reinforcement learning framework to analyze patients' nutritional needs, complication risks, and dietary preferences.

[0055] On this basis, this application implements a dynamic feature drift compensation mechanism, which monitors and adjusts model parameters in real time through online incremental learning to ensure the accuracy and timeliness of recommendations.

[0056] Finally, this application generates a hierarchical dietary decision tree, combines the Monte Carlo tree search and reinforcement learning reward mechanism, balances the optimization of nutritional indicators and the improvement of dietary experience, and obtains dietary recommendation decision information.

[0057] This application also proposes a diet recommendation system for patients with cirrhosis based on machine learning. The system includes a multimodal feature map construction module, a multi-task reinforcement learning modeling module, a dynamic feature drift compensation module and a hierarchical diet decision generation module. The modules work together to realize intelligent recommendation of personalized diet for patients with cirrhosis.

[0058] In addition, the present application also proposes an electronic device and a storage medium capable of executing the above-mentioned machine learning-based diet recommendation method for patients with cirrhosis, as well as a computer program product comprising a computer program for executing the method.

[0059] The use of these devices and media can enable the machine learning-based diet recommendation method for patients with cirrhosis of the present application to be effectively implemented in actual scenarios, providing patients with cirrhosis with more personalized and intelligent diet recommendation services.

[0060] The solution of this application can significantly improve the accuracy of dietary recommendations, and help patients with cirrhosis manage their disease and improve their quality of life.

[0061] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, circuit structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0062] Example 1

[0063] This embodiment 1 provides a method for recommending diet for patients with cirrhosis based on machine learning. Figure 1 , including the following steps:

[0064] S100: Construct a multi-dimensional patient feature map, integrate multimodal data sets consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation texts in real time through API interfaces, and use graph convolutional networks to extract topological correlation features between multimodal data.

[0065] In step S100, the following sub-steps are specifically included (see Figure 2 ):

[0066] S101: Access a multimodal dataset consisting of patient physiological indicators, dietary behavior, clinical tests, and doctor-patient conversations in real time through an API interface. After data cleaning, time alignment, and entity association processing, a multimodal unified dataset with a unified timestamp is formed.

[0067] Specifically, multi-source heterogeneous data such as patient physiological indicators, dietary behavior, clinical tests, and doctor-patient conversations are accessed in real time through API interfaces to form a multimodal dataset. Since raw data often have noise, missing values, duplicate values, inconsistencies, and other problems, data cleaning is required. Data cleaning techniques are used, such as deleting irrelevant data, smoothing noisy data, filling missing values, and correcting erroneous data, to improve data quality. At the same time, in order to achieve alignment of multimodal data in the time dimension, time information is marked according to timestamps or sampling rates, and time alignment is performed using strategies such as interpolation, alignment to the nearest timestamp, or a unified time grid. In addition, entity association processing is also required. Entity alignment technology is used to identify the same entities in different data sources through pre-trained language models, similarity metrics, and clustering algorithms, and to associate them to construct a multimodal unified dataset with a unified timestamp.

[0068] S102: Based on the multimodal unified data set formed in step S101, a patient feature map of multiple types of nodes including basic patient information, physiological indicators, dietary behavior, test reports, and doctor-patient dialogues is constructed, and the intrinsic correlation relationship of the data is represented by defining the association edges between the nodes.

[0069] Specifically, based on the multimodal unified data set formed in step S101, a patient characteristic graph is constructed. The patient characteristic graph is a knowledge graph that contains multiple types of nodes such as patient basic information, physiological indicators, dietary behavior, test reports, doctor-patient dialogues, etc. By defining the association edges between nodes, such as the association between patients and physiological indicators, the association between physiological indicators and test reports, etc., the inherent association relationship of the data is characterized. During the construction process, various entities are first identified and extracted as nodes, and then the association relationship between nodes is determined based on medical knowledge, semantic information in the data, etc., and organized and stored in a graph structure. Knowledge graph construction technology, such as graph databases, is used to achieve structured representation of data, providing a basis for subsequent feature extraction and analysis.

[0070] S103: For the patient feature map constructed in step S102, a graph convolutional network or a relational graph convolutional network is used to perform node neighborhood aggregation operations to extract topological correlation features that reflect the interactive characteristics of multimodal data.

[0071] Specifically, for the patient feature map constructed in step S102, a graph convolutional network (GCN) or a relational graph convolutional network (RGCN) is used to perform node neighborhood aggregation operations to extract topological correlation features. The core idea of the graph convolutional network is to learn the low-dimensional embedding representation of the node by aggregating the information of the node and its neighboring nodes. In the patient feature map, the features of each node not only contain its own information, but also interact with the information of the surrounding nodes. Through the graph convolution operation, the topological structure and feature interaction between nodes can be captured, and topological correlation features reflecting the interactive characteristics of multimodal data can be extracted. These features can more comprehensively characterize the patient's status and characteristics, and provide richer information for subsequent analysis and modeling.

[0072] S200: Based on the multidimensional patient feature map obtained in step S100 and the extracted topological correlation features, a multi-task deep reinforcement learning framework is established, which includes a nutritional needs prediction branch, a complication risk warning branch, and a dietary preference adaptation branch, for analyzing the patient's nutritional needs, complication risks, and dietary preferences;

[0073] In step S200, the following sub-steps are specifically included (see Figure 3 ):

[0074] S201: Based on the multi-dimensional patient feature map and topological correlation features extracted in step S100, a multi-task deep reinforcement learning framework is constructed.

[0075] Specifically, the multi-task deep reinforcement learning framework is constructed based on the multi-dimensional patient feature map and its topological correlation features extracted in step S100. Multi-task learning is a machine learning paradigm that can handle multiple tasks simultaneously. It improves the generalization ability and learning efficiency of the model through shared representation learning and cross-task knowledge transfer. In reinforcement learning, the intelligent agent learns the optimal strategy by interacting with the environment to maximize the cumulative reward. In this step, the deep learning network is used as an approximator of the policy function or value function of reinforcement learning, which can handle complex high-dimensional state space and action space.

[0076] Specifically, the multi-task deep reinforcement learning framework consists of multiple neural network modules, including shared layers and task-specific layers. The shared layers are used to extract common features across different tasks, while the task-specific layers are used to handle the unique requirements of each task. In this way, the model can share knowledge across different tasks while maintaining targeted optimization for each task. Furthermore, the multi-task reinforcement learning framework requires the design of effective reward mechanisms and exploration strategies to balance competition and collaboration between tasks, ensuring that the model achieves good performance on multiple tasks.

[0077] S202: Integrate three branches of nutritional demand prediction, complication risk warning, and dietary preference adaptation into the multi-task deep reinforcement learning framework to respectively realize the functions of quantitative analysis of patient nutritional needs, dynamic assessment of complication risk, and dietary preference pattern recognition.

[0078] Specifically, the multi-task deep reinforcement learning framework we constructed integrates three branches: nutritional needs prediction, complication risk warning, and dietary preference adaptation. These branches enable quantitative analysis of patient nutritional needs, dynamic assessment of complication risk, and recognition of dietary preference patterns. Each branch is optimized for a specific task, processing different types of input data and outputting corresponding results.

[0079] Specifically, taking nutritional needs prediction as an example, this branch may need to analyze the patient's physiological indicators, eating behavior and other information to predict their needs for various nutrients. The complication risk warning branch needs to comprehensively consider factors such as the patient's medical history, physiological indicators and current treatment status to assess the possibility of complications. The dietary preference adaptation branch focuses on analyzing data such as the patient's historical eating behavior and doctor-patient conversations to identify the patient's dietary preference patterns.

[0080] Specifically, to improve model performance, each branch can adopt different deep learning structures, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to process time series data, and convolutional neural networks (CNNs) to process images or sequence data. At the same time, technologies such as attention mechanisms can be used to enhance the model's ability to capture key features.

[0081] S203: Jointly model the outputs of the three branches in S202, perform cross-branch fusion processing on the nutritional demand prediction results, complication risk assessment indicators, and dietary preference identification features, and output a comprehensive analysis result including the patient's real-time status.

[0082] The outputs of the three branches in S202 are jointly modeled, integrating the nutritional needs prediction results, complication risk assessment indicators, and dietary preference identification features across branches to output a comprehensive analysis of the patient's real-time status. The goal of joint modeling is to form a comprehensive and accurate description of the patient's status by integrating the outputs of multiple tasks.

[0083] Specifically, joint modeling can be achieved in a variety of ways, such as splicing the output features of each branch and then further processing them through a fully connected layer or other fusion modules. During the fusion process, attention mechanisms or weighted summation methods can be used to dynamically adjust the importance weights of the outputs of each branch to ensure that the final result can fully reflect the patient's status in different aspects. In this way, the comprehensive analysis results obtained not only include the patient's nutritional needs information, but also take into account their complication risks and dietary preferences, providing a comprehensive basis for subsequent decisions such as dietary recommendations.

[0084] S300: During the operation of the multi-task deep reinforcement learning framework, a dynamic feature drift compensation mechanism is implemented, patient data distribution deviation is monitored in real time through an online incremental learning module, and model parameters are updated using an adversarial domain adaptation algorithm;

[0085] In step S300, the operation process of the three branches of nutritional demand prediction, complication risk warning, and dietary preference adaptation includes:

[0086] The process of nutritional requirement prediction: Based on the physiological indicator nodes and clinical test nodes in the patient characteristic map constructed in step S102, the gated recurrent unit network of the nutritional requirement prediction branch in step S202 is called, the liver metabolic dynamics model is integrated to generate a nutrient requirement baseline, and the prediction results are input into the cross-branch fusion module in step S203;

[0087] The complication risk warning process includes: constructing a risk feature tensor using the topological correlation features extracted in step S103, and calculating the cross-period risk correlation weights through the multi-head attention mechanism of the complication risk warning branch in step S202;

[0088] The process of dietary preference pattern recognition: parse the unstructured doctor-patient dialogue text integrated in step S101, extract dietary behavior trigger words through semantic role labeling of the dietary preference adaptation branch in step S202, and realize dynamic coupling of preference characteristics and nutritional restriction conditions in step S203 based on the neighborhood node propagation mechanism of the graph convolutional network in step S103.

[0089] Step S300 specifically includes the following sub-steps:

[0090] S301: During the execution of the multi-task deep reinforcement learning framework, the online incremental learning module is used to monitor the patient data distribution in real time to identify any deviations.

[0091] Specifically, during the execution of the multi-task deep reinforcement learning framework, an online incremental learning module monitors patient data distribution in real time to identify any drift. Online incremental learning is a learning method that processes data streams and continuously updates models. It incrementally updates the model by continuously receiving new data samples, ensuring that the model can adapt to changes in the data distribution. In the medical field, patient data such as physiological indicators and dietary behaviors may change over time, and these changes can lead to degradation in model performance. Therefore, the online incremental learning module is crucial for promptly detecting data distribution drift. By monitoring statistical characteristics of the data, such as the mean and variance, or by utilizing specialized detection algorithms such as window-based statistical tests, data distribution drift can be effectively identified.

[0092] S302: When the online incremental learning module detects that the patient data distribution is offset, the adversarial domain adaptation algorithm is used to update the model parameters, thereby realizing a dynamic feature drift compensation mechanism.

[0093] Specifically, when the online incremental learning module detects that the distribution of patient data has shifted, the adversarial domain adaptation algorithm is used to update the model parameters, thereby realizing a dynamic feature drift compensation mechanism. The adversarial domain adaptation algorithm is an effective method for dealing with domain shift. Its basic idea is to learn a mapping to map the features of the source domain (original data distribution) and the target domain (shifted data distribution) to a shared feature space, in which the feature distributions of the source domain and the target domain are as close as possible. In the multi-task deep reinforcement learning framework, through the adversarial domain adaptation algorithm, the model can automatically adjust the parameters according to the shifted data, thereby adapting to the new data distribution and improving the generalization ability and stability of the model. This dynamic compensation mechanism can ensure that the model continues to output accurate prediction results during long-term operation, which is of great significance for application scenarios that require long-term tracking and management, such as dietary recommendations for patients with cirrhosis.

[0094] S400: Based on the patient's nutritional needs information, complication risk information, and dietary preference information output by the multi-task deep reinforcement learning framework in step S200, and the valid data after dynamic feature drift compensation in step S300, a hierarchical dietary decision tree is generated, and a Monte Carlo tree search is performed on the food knowledge graph based on the current patient state vector. The reward mechanism of reinforcement learning is combined to balance the optimization of nutritional indicators and the improvement of dietary experience, and dietary recommendation decision information is obtained.

[0095] In step S400, the following sub-steps are specifically included (see Figure 4 ):

[0096] S401: Based on the comprehensive analysis results of the patient's real-time status output in step S203 and the valid data after dynamic feature drift compensation in step S300, a node weight distribution function of a hierarchical dietary decision tree is constructed, wherein the root node of the decision tree is associated with the nutritional restriction conditions in the patient feature map constructed in step S102, and the child nodes dynamically adjust the food association path based on the topological association features extracted in step S103;

[0097] Specifically, based on the comprehensive analysis results of the patient's real-time status output in step S203 and the valid data after dynamic feature drift compensation in step S300, a node weight distribution function of a hierarchical diet decision tree is constructed. The root node of the decision tree is associated with the nutritional restrictions in the patient feature map constructed in step S102, and the child nodes dynamically adjust the food association path through the topological association features extracted in step S103. The purpose of constructing a hierarchical diet decision tree is to provide a structured decision framework for diet recommendations, so that the recommended plan can comprehensively consider multiple factors such as the patient's nutritional needs, complication risks, and dietary preferences. The node weight distribution function determines the weight of each node according to the patient's real-time status and feature importance, thereby affecting the branch selection of the decision tree. The root node is associated with the patient's nutritional restrictions to ensure that the recommended diet plan meets the patient's nutritional needs; the child nodes use topological association features to dynamically adjust the food association path so that the recommended ingredients can better meet the patient's comprehensive needs.

[0098] S402: In the hierarchical dietary decision tree generated in step S401, using the current patient state vector as the initial search starting point, combined with the nutritional requirements, risk probabilities, and preference weights output by each branch in step S202, a Monte Carlo tree search is performed on the food knowledge graph. The exploration and exploitation balance coefficient in the search strategy is dynamically adjusted using the model parameters updated in step S300;

[0099] Specifically, in the hierarchical dietary decision tree generated in step S401, the current patient state vector is used as the initial search starting point, and the nutritional requirements, risk probabilities, and preference weights output by each branch in step S202 are combined to perform a Monte Carlo tree search (MCTS) on the food knowledge graph. The exploration and development balance coefficient in the search strategy is dynamically adjusted through the model parameters updated in step S300. Monte Carlo tree search is a heuristic search algorithm that combines random simulation and tree search, and is suitable for complex state spaces and decision-making processes. It continuously iterates through four stages of selection, exploration, simulation, and backpropagation to gradually build and optimize the search tree. In this step, MCTS uses the patient's state vector and the output information of each branch to explore different food combinations in the food knowledge graph and evaluate their potential nutritional and preference values. At the same time, by dynamically adjusting the exploration and development balance coefficient, it is ensured that the search process can not only widely explore possible food combinations, but also make full use of existing information for effective development.

[0100] S403: Based on the candidate diet plans generated by the Monte Carlo tree search in step S402, the reward mechanism of the multi-task deep reinforcement learning framework in step S200 is called, and the nutritional optimization target output by the cross-branch fusion module in step S203 and the dietary preference characteristics calibrated in step S300 are synchronously integrated. The Pareto frontier of nutritional indicators and dietary experience in the recommended plan is iteratively optimized through the gradient ascent algorithm to generate the final diet recommendation decision information.

[0101] Specifically, for the candidate diets generated through the Monte Carlo tree search in step S402, the reward mechanism of the multi-task deep reinforcement learning framework in step S200 is invoked. The nutritional optimization objective output by the cross-branch fusion module in step S203 is synchronously integrated with the dietary preference features calibrated in step S300. A gradient ascent algorithm is then used to iteratively optimize the Pareto frontier of nutritional indicators and dietary experience in the recommended diets to generate the final dietary recommendation decision information. The reward mechanism is a core component of reinforcement learning, used to measure the pros and cons of different diets. In this step, the reward mechanism comprehensively considers the nutritional optimization objective and dietary preference features, ensuring that the optimized diet achieves the optimal balance between nutritional indicators and dietary experience. The gradient ascent algorithm is used to maximize the reward function and iteratively adjust the solution parameters to gradually approach the Pareto frontier, the boundary where no better solution can be found in the multi-objective optimization. The resulting dietary recommendation decision information not only meets the patient's nutritional needs and dietary preferences, but also takes into account factors such as the risk of complications, providing patients with a personalized, scientifically sound dietary plan.

[0102] Example 2

[0103] The machine learning-based diet recommendation system for patients with cirrhosis in this embodiment includes:

[0104] Multimodal feature map construction module: used to construct a multidimensional patient feature map. It integrates multimodal datasets consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation text in real time through API interfaces, and uses graph convolutional networks to extract topological correlation features between multimodal data.

[0105] Multi-task reinforcement learning modeling module: This module is used to establish a multi-task deep reinforcement learning framework based on the obtained multi-dimensional patient feature map and extracted topological correlation features, including a nutritional needs prediction branch, a complication risk warning branch, and a dietary preference adaptation branch. This framework is used to analyze patients' nutritional needs, complication risks, and dietary preferences.

[0106] Dynamic feature drift compensation module: used to implement a dynamic feature drift compensation mechanism during the operation of the multi-task deep reinforcement learning framework, monitor the patient data distribution offset in real time through the online incremental learning module, and update the model parameters using the adversarial domain adaptation algorithm;

[0107] Hierarchical dietary decision generation module: It is used to generate a hierarchical dietary decision tree based on the patient's nutritional needs information, complication risk information, dietary preference information, and valid data after dynamic feature drift compensation output by the multi-task deep reinforcement learning framework. It performs Monte Carlo tree search on the food knowledge graph based on the current patient state vector, and combines the reinforcement learning reward mechanism to balance the optimization of nutritional indicators and the improvement of dietary experience to obtain dietary recommendation decision information.

[0108] This dietary recommendation system for patients with cirrhosis consists of the following key modules. The multimodal feature map construction module utilizes the Python programming language and the open-source Apache NiFi tool to implement data collection and preprocessing, Apache Spark for large-scale data processing, and the Neo4j graph database for storing and managing patient feature maps. The multi-task reinforcement learning modeling module utilizes the Python programming language and the TensorFlow and Keras deep learning frameworks to build a multi-task deep reinforcement learning framework. The nutritional requirement prediction branch incorporates a gated recurrent unit (GRU) network, the complication risk warning branch utilizes a multi-head attention mechanism, and the dietary preference adaptation branch incorporates semantic role labeling technology. The dynamic feature drift compensation module utilizes the Python programming language and a deep learning model optimized using the adaptive moment estimation method, utilizing the open-source Ray library for large-scale distributed computing. The hierarchical dietary decision generation module utilizes the Python programming language and scikit-learn to construct a hierarchical dietary decision tree. Monte Carlo tree search is implemented using the Monte Carlo Tree Search (MCTS) algorithm, and an adaptive reward function is also introduced.

[0109] This module is developed in the Python programming language and uses the Flask or Django framework to build a RESTful API to connect with the hospital information system. In terms of data collection, the Apache NiFi tool is used to achieve real-time collection of multi-source data such as patient physiological indicators, dietary behaviors, clinical test results, and doctor-patient conversation texts. In the data preprocessing stage, data processing libraries such as Pandas and NumPy are used to complete data cleaning, time alignment, and entity association processing. The graph convolutional network (GCN) part is implemented using the TensorFlow or PyTorch deep learning framework to construct a multi-layer neural network structure and extract topological correlation features between multimodal data. Finally, the processed data is stored in the Neo4j graph database to construct a patient feature map containing multiple types of nodes such as patient basic information, physiological indicators, and dietary behaviors.

[0110] Example 3

[0111] This embodiment provides an electronic device for implementing the above-mentioned method for recommending a diet for patients with cirrhosis based on machine learning. The electronic device includes a memory, a processor, and a communication interface for data interaction with external devices. The memory is used to store program instructions, patient data (such as physiological indicators, dietary behavior, clinical test results, multimodal data such as doctor-patient conversation texts), model parameters (including parameters of a multi-task deep reinforcement learning framework, parameters of a graph convolutional network, etc.), a knowledge base related to diet recommendations (such as food knowledge graphs, nutritional requirement standards, etc.), and temporary files generated during the operation of the system. The processor is used to execute program instructions in the memory to run the various modules of the above-mentioned diet recommendation system for patients with cirrhosis based on machine learning. The communication interface connects to external devices such as hospital information systems and mobile terminals through wired or wireless networks to achieve data transmission and interaction.

[0112] When executing the program, the processor first obtains the patient's multimodal data in real time through the communication interface, and calls the relevant programs of the multimodal feature map construction module, uses the graph convolutional network to process the data, and constructs a multidimensional patient feature map. Next, the processor calls the multi-task reinforcement learning modeling module to analyze the patient's nutritional needs, complication risks and dietary preferences. During the analysis process, the dynamic feature drift compensation module monitors the changes in data distribution in real time. Once an offset is detected, the adversarial domain adaptation algorithm is immediately used to update the model parameters. Finally, the processor calls the hierarchical diet decision generation module, combines the patient's real-time status and the updated model parameters, performs a Monte Carlo tree search on the food knowledge graph, and uses the reinforcement learning reward mechanism to optimize the search results to generate the final diet recommendation decision information. The electronic device can be a server deployed in a hospital data center or a smart terminal device used in the patient's home, providing convenient and personalized diet recommendation services for patients with cirrhosis.

[0113] Example 4

[0114] The present embodiment relates to a storage medium comprising computer executable instructions, which, when executed by a computer processor, are used to execute the above-mentioned method for recommending a diet for patients with cirrhosis based on machine learning. The storage medium can be a USB flash drive, an SD card, a solid-state drive, a mechanical hard disk, an optical disc, a random access memory (RAM), a read-only memory (ROM), or other types of storage devices. The program stored in the storage medium includes multiple modules, such as a multimodal feature map construction module, a multi-task reinforcement learning modeling module, a dynamic feature drift compensation module, and a hierarchical diet decision generation module. When the program is loaded into a computer processor, the processor executes these instructions to collect multimodal data such as the patient's physiological indicators, dietary behavior, clinical test results, and doctor-patient conversation texts in real time. It then uses a graph convolutional network to extract the complex relationship features between multimodal data and construct a multidimensional patient feature map. Simultaneously, based on the constructed map and features, a multi-task deep reinforcement learning framework is established to analyze the patient's nutritional needs, complication risks, and dietary preferences. During model operation, the online incremental learning module monitors changes in data distribution and uses an adversarial domain adaptation algorithm to update model parameters to ensure the accuracy and stability of the recommended solutions. Finally, combining the patient's real-time status and model output results, a Monte Carlo tree search is performed on the food knowledge graph, and a reinforcement learning reward mechanism is used to optimize the search results to generate personalized dietary recommendations. This storage medium can be widely used in hospital information systems, mobile medical applications, or home health management systems to provide scientific and accurate dietary guidance services for patients with cirrhosis.

[0115] Example 5

[0116] This embodiment introduces a computer program product primarily for implementing the aforementioned machine learning-based dietary recommendation method for patients with cirrhosis. The program modules include a multimodal feature map construction module, which collects and integrates multimodal data such as patients' physiological indicators, dietary behaviors, clinical test results, and textual conversations between doctors and patients in real time. It then uses a graph convolutional network to extract topological correlation features between these data. The multitask reinforcement learning modeling module, the core component of the program, builds a multitask deep reinforcement learning framework based on the constructed multidimensional patient feature map, encompassing nutritional needs prediction, complication risk warning, and dietary preference adaptation. The dynamic feature drift compensation module monitors changes in data distribution in real time during model execution. When drift is detected, it automatically updates model parameters using an adversarial domain adaptation algorithm to ensure the accuracy and timeliness of recommendations. The hierarchical dietary decision generation module combines the patient's real-time status and model outputs to perform a Monte Carlo tree search on the ingredient knowledge graph. It then optimizes the search results using a reinforcement learning reward mechanism, ultimately generating personalized dietary recommendations. This computer program product has good compatibility and scalability, and can be widely used in hospital information systems, mobile medical applications, and home health management systems. It can provide scientific and accurate dietary guidance for patients with cirrhosis, assist patients in effective diet management, and improve treatment effects and quality of life.

[0117] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for recommending diet for patients with cirrhosis based on machine learning, characterized in that: The following steps are involved: S100: Constructs a multi-dimensional patient feature map, integrating a multimodal dataset consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation texts in real time through an API interface, and uses a graph convolutional network to extract topological correlation features between multimodal data; S2 00: Based on the multi-dimensional patient feature map obtained in step S100 and the extracted topological correlation features, a multi-task deep reinforcement learning framework is established, which includes a nutritional demand prediction branch, a complication risk warning branch, and a dietary preference adaptation branch, to analyze the patient's nutritional needs, complication risk, and dietary preferences; S300: During the operation of the multi-task deep reinforcement learning framework, a dynamic feature drift compensation mechanism is implemented, patient data distribution deviation is monitored in real time through an online incremental learning module, and model parameters are updated using an adversarial domain adaptation algorithm; S400: Based on the patient's nutritional needs information, complication risk information, and dietary preference information output by the multi-task deep reinforcement learning framework in step S200, and the valid data after dynamic feature drift compensation in step S300, a hierarchical dietary decision tree is generated, and a Monte Carlo tree search is performed on the food knowledge graph based on the current patient state vector. The reward mechanism of reinforcement learning is combined to balance the optimization of nutritional indicators and the improvement of dietary experience, and dietary recommendation decision information is obtained.

2. The method for recommending a diet for patients with cirrhosis based on machine learning according to claim 1, characterized in that: Step S100 specifically includes the following sub-steps: S101: Access a multimodal dataset consisting of patient physiological indicators, dietary behavior, clinical tests, and doctor-patient conversations in real time through an API interface. After data cleaning, time alignment, and entity association processing, a unified multimodal dataset with a unified timestamp is formed. S102: Based on the multimodal unified data set formed in step S101, a patient feature map is constructed that includes multiple types of nodes, including basic patient information, physiological indicators, dietary behavior, test reports, and doctor-patient conversations. The internal correlation relationships of the data are represented by defining association edges between the nodes. S103: For the patient feature map constructed in step S102, a graph convolutional network or a relational graph convolutional network is used to perform node neighborhood aggregation operations to extract topological correlation features that reflect the interactive characteristics of multimodal data.

3. The method for recommending diet for patients with cirrhosis based on machine learning according to claim 2, characterized in that: Step S200 specifically includes the following sub-steps: S201: Building a multi-task deep reinforcement learning framework based on the multi-dimensional patient feature map and topological correlation features extracted in step S100; S202: Integrating three branches, namely, nutritional needs prediction, complication risk warning, and dietary preference adaptation, into the multi-task deep reinforcement learning framework to respectively implement functions of quantitative analysis of patient nutritional needs, dynamic assessment of complication risk, and dietary preference pattern recognition; S203: Jointly model the outputs of the three branches in S202, perform cross-branch fusion processing on the nutritional demand prediction results, complication risk assessment indicators, and dietary preference identification features, and output a comprehensive analysis result including the patient's real-time status.

4. The method for recommending a diet for patients with cirrhosis based on machine learning according to claim 3, characterized in that: In step S300, the operation process of the three branches of nutritional demand prediction, complication risk warning, and dietary preference adaptation includes: The process of nutritional requirement prediction: Based on the physiological indicator nodes and clinical test nodes in the patient characteristic map constructed in step S102, the gated recurrent unit network of the nutritional requirement prediction branch in step S202 is called, the liver metabolic dynamics model is integrated to generate a nutrient requirement baseline, and the prediction results are input into the cross-branch fusion module in step S203; The complication risk warning process includes: constructing a risk feature tensor using the topological correlation features extracted in step S103, and calculating the cross-period risk correlation weights through the multi-head attention mechanism of the complication risk warning branch in step S202; The process of dietary preference pattern recognition: parse the unstructured doctor-patient dialogue text integrated in step S101, extract dietary behavior trigger words through semantic role labeling of the dietary preference adaptation branch in step S202, and realize dynamic coupling of preference characteristics and nutritional restriction conditions in step S203 based on the neighborhood node propagation mechanism of the graph convolutional network in step S103.

5. The method for recommending diet for patients with cirrhosis based on machine learning according to claim 1, characterized in that: Step S300 specifically includes the following sub-steps: S301: During the execution of the multi-task deep reinforcement learning framework, the online incremental learning module is used to monitor the patient data distribution in real time to identify any deviations. S302: When the online incremental learning module detects that the patient data distribution is offset, the adversarial domain adaptation algorithm is used to update the model parameters, thereby realizing a dynamic feature drift compensation mechanism.

6. The method for recommending diet for patients with cirrhosis based on machine learning according to claim 3, characterized in that: Step S400 specifically includes the following sub-steps: S401: Based on the comprehensive analysis results of the patient's real-time status output in step S203 and the valid data after dynamic feature drift compensation in step S300, a node weight distribution function of a hierarchical dietary decision tree is constructed, wherein the root node of the decision tree is associated with the nutritional restriction conditions in the patient feature map constructed in step S102, and the child nodes dynamically adjust the food association path based on the topological association features extracted in step S103; S402: In the hierarchical dietary decision tree generated in step S401, using the current patient state vector as the initial search starting point, combined with the nutritional requirements, risk probabilities, and preference weights output by each branch in step S202, a Monte Carlo tree search is performed on the food knowledge graph. The exploration and exploitation balance coefficient in the search strategy is dynamically adjusted using the model parameters updated in step S300; S403: Based on the candidate diet plans generated by the Monte Carlo tree search in step S402, the reward mechanism of the multi-task deep reinforcement learning framework in step S200 is called, and the nutritional optimization target output by the cross-branch fusion module in step S203 and the dietary preference characteristics calibrated in step S300 are synchronously integrated. The Pareto frontier of nutritional indicators and dietary experience in the recommended plan is iteratively optimized through the gradient ascent algorithm to generate the final diet recommendation decision information.

7. A diet recommendation system for patients with cirrhosis based on machine learning, characterized in that: include: Multimodal feature map construction module: used to construct a multidimensional patient feature map. It integrates multimodal datasets consisting of patient physiological indicator data streams, dietary behavior time series data, clinical test structured data, and unstructured doctor-patient conversation text in real time through API interfaces, and uses graph convolutional networks to extract topological correlation features between multimodal data. Multi-task reinforcement learning modeling module: This module is used to establish a multi-task deep reinforcement learning framework based on the obtained multi-dimensional patient feature map and extracted topological correlation features, including a nutritional needs prediction branch, a complication risk warning branch, and a dietary preference adaptation branch. This framework is used to analyze patients' nutritional needs, complication risks, and dietary preferences. Dynamic feature drift compensation module: used to implement a dynamic feature drift compensation mechanism during the operation of the multi-task deep reinforcement learning framework, monitor the patient data distribution offset in real time through the online incremental learning module, and update the model parameters using the adversarial domain adaptation algorithm; Hierarchical dietary decision generation module: It is used to generate a hierarchical dietary decision tree based on the patient's nutritional needs information, complication risk information, dietary preference information, and valid data after dynamic feature drift compensation output by the multi-task deep reinforcement learning framework. It performs Monte Carlo tree search on the food knowledge graph based on the current patient state vector, and combines the reinforcement learning reward mechanism to balance the optimization of nutritional indicators and the improvement of dietary experience to obtain dietary recommendation decision information.

8. An electronic device comprising a memory and a processor, characterized in that: The processor is used to execute the program in the memory to implement the diet recommendation method for patients with cirrhosis based on machine learning as described in any one of claims 1 to 6.

9. A storage medium containing computer-executable instructions, characterized in that: The storage medium of the computer-executable instructions, when executed by a computer processor, is used to execute the diet recommendation method for patients with cirrhosis based on machine learning as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the diet recommendation method for patients with cirrhosis based on machine learning as described in any one of claims 1 to 6.

Images