Chronic heart failure nutrition and health assessment system based on machine learning

Abstract

The invention discloses a chronic heart failure nutrition health assessment system based on machine learning, which belongs to the field of health monitoring and comprises an acquisition processing module, a feature fusion module, a diet recognition module, a metabolism analysis module, a risk assessment module, a health prediction module, a nutrition recommendation module, an interpretation feedback module, a simulation prediction module and a feedback monitoring module. According to the method, the nutrition behavior and physiological status of the patient can be comprehensively described, the accuracy and efficiency of nutrition evaluation are remarkably improved, the nutrition intervention target spot is accurately identified, and the understanding and compliance of the patient are improved.

Description

Technical Field

[0001] The present invention relates to the field of health monitoring, and in particular to a chronic heart failure nutritional health assessment system based on machine learning. Background Art

[0002] Chronic heart failure (CHF) is one of the cardiovascular diseases with the highest hospitalization rate worldwide. Its morbidity and mortality continue to rise with the aging population and the trend of chronic disease comorbidity. The disease is not only manifested as cardiac systolic or diastolic dysfunction, but also accompanied by systemic metabolic reconstruction, including pathological processes such as energy metabolism disorder, electrolyte imbalance, nutrient intake imbalance, and intestinal flora disorder. Therefore, reasonable nutritional intervention can not only reduce the burden on the heart and control water and sodium retention, but also delay disease progression and reduce the frequency of acute attacks by improving the metabolic environment. However, there are still many bottlenecks in the current nutritional assessment and intervention practices. On the one hand, traditional nutritional management methods are highly dependent on the experience of nutritionists, and it is difficult to fully integrate patients' dietary behavior, physiological state and individual differences. On the other hand, the fragmentation of data sources and the lack of unified standards are also problems that need to be solved urgently. For example, the electronic medical record system provides structured physiological indicators and medication records, dietary images and voice self-descriptions are often unstructured data, and wearable devices collect high-frequency, continuous time series data. These data with different modalities are difficult to efficiently integrate under the same analysis framework. In addition, current nutrition recommendation systems generally lack the description of causal mechanisms, and have limited support for the interpretability and operability of recommendation results, making them unable to cope with personalized and dynamic chronic disease management tasks.

[0003] The existing chronic heart failure nutrition and health assessment system cannot fully characterize the patient's nutritional behavior and physiological status, reduces the accuracy and efficiency of nutritional assessment, and cannot accurately identify nutritional intervention targets. Therefore, we propose a chronic heart failure nutrition and health assessment system based on machine learning. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and propose a chronic heart failure nutritional health assessment system based on machine learning.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The machine learning-based chronic heart failure nutrition and health assessment system includes an acquisition and processing module, a feature fusion module, a diet recognition module, a metabolic analysis module, a risk assessment module, a health prediction module, a nutrition recommendation module, an explanation feedback module, a simulation prediction module, and a feedback monitoring module;

[0007] The acquisition and processing module is used to collect structured data, unstructured data, and time-series data, and preprocess various types of data through format standardization, missing value completion, and semantic unification;

[0008] The feature fusion module synchronously aligns various types of preprocessed data to generate cross-modal unified features;

[0009] The diet recognition module recognizes the actual diet content of the patient and constructs a 3D nutrition map;

[0010] The metabolic analysis module is used to identify potential causal relationships between nutritional imbalance and metabolic abnormalities;

[0011] The risk assessment module evaluates the patient's nutritional risk level based on the cross-modal unified features and metabolic analysis results, and outputs a corresponding risk level report while highlighting high-risk nutritional indicators;

[0012] The health prediction module is used to identify the causal impact of nutritional factors on cardiac function indicators;

[0013] The nutrition recommendation module generates a diet plan that takes into account nutritional needs, taste preferences, and ingredient availability based on the risk level report and causal analysis results;

[0014] The explanation and feedback module dynamically adjusts the depth of explanation according to the patient's health literacy and cognitive level;

[0015] The simulation prediction module is used to construct a digital twin of the patient individual for multi-scale simulation;

[0016] The feedback monitoring module extracts new information from wearable devices, diet record APPs, and regular test data, continuously updates the risk assessment and recommendation plans, and records the patient's actual implementation of the recommendation plan.

[0017] As a further solution of the present invention, the structured data specifically includes electronic medical records, laboratory test results, drug usage information, and diet questionnaire scale scores;

[0018] The unstructured data specifically includes diet images, voice descriptions, and free-text medical records or nursing records;

[0019] The time-series data specifically includes wearable device monitoring data, nutritional intake logs, daily weight change data, and drug administration time records.

[0020] As a further solution of the present invention, the specific steps for the feature fusion module to generate cross-modal unified features are as follows:

[0021] S1.1: Embed the structured data through a fully connected neural network, encode the unstructured data through CNN and Transformer respectively, then use bidirectional LSTM to perform sequence modeling on the time series data, and collect the feature data generated after different data encoding processes for various types of data;

[0022] S1.2: Calculate the attention weights of various types of feature data at each moment based on the time attention mechanism, perform weighted fusion on various types of feature data based on the attention weights at each moment, then map various types of feature data to the same dimension, and calculate the spatial attention weights of various types of feature data in the current state. According to the calculated spatial attention weights, fuse the information of each modality;

[0023] S1.3: Through the modality-time joint splicing operation, splice various types of feature data processed by the time attention mechanism and the spatial attention mechanism into an overall feature, map the spliced overall feature to the same shared space, and then perform normalization processing on the spliced overall feature to generate the final cross-modal unified feature.

[0024] As a further solution of the present invention, the specific steps for the diet recognition module to recognize the actual diet content of the patient and construct a 3D nutrition map are as follows:

[0025] S2.1: The diet recognition module encodes the cross-modal unified feature through a CNN feature extractor into a feature map with semantic information, constructs multiple sliding windows of size N×N on the feature map, predicts the object existence probability and position offset of multiple anchor boxes at each sliding window position, and selects the region with a high existence probability as the candidate box;

[0026] S2.2: Use ROI Align to accurately crop and resample each candidate box into a unified size feature matrix, then judge the specific food category of each candidate box, further correct the position of the candidate box, and then retain the one with the highest classification probability among multiple candidate boxes of the same type of object, and eliminate other candidate boxes with an overlap degree higher than the set threshold. Then, according to the image depth information in the candidate box, calculate the actual weight of the food in the candidate box, and use the density conversion table to calculate the corresponding food quality;

[0027] S2.3: Calculate the corresponding nutrient content according to the weight of various foods, use speech-to-text technology to convert the patient's speech into text, then use entity recognition technology to extract the food names and intakes in each text data, and fuse the extracted data groups with the image recognition results;

[0028] S2.4: Record the nutrient intake of the patient at each moment of each day to construct the corresponding dynamic time series, and map the nutrient-time intensity information to a three-dimensional space of the nutrient dimension, time dimension, and intake intensity to form the corresponding 3D heat map, so as to reflect the structure, frequency, and risk pattern of the patient's daily nutritional intake behavior in real time.

[0029] As a further solution of the present invention, the specific steps for the metabolic analysis module to identify the potential causal relationship between nutritional imbalance and metabolic abnormality are as follows:

[0030] S3.1: Through Z-score standardization processing, convert the original metabolite concentration data into a unified numerical range, and calculate the standardized scores of various metabolites. If the standardized score of any metabolite is higher than the preset threshold, it is determined that the metabolite is abnormal;

[0031] S3.2: Extract metabolite and nutrient information from public databases such as HMDB, KEGG, ChEBI, PubChem, and USDA, and calculate the embedding distance of synonymous entity pairs in different sources through the word embedding method. If the embedding distance is less than the preset threshold, the two entities are considered equivalent and merged into a unified entity node;

[0032] S3.3: Extract semantic relationships from text data through NLP methods, construct a set of relationship categories based on the annotation relationships in each database, record the standardized entities and the extracted relationships as triples in the graph in a structured manner, and then encode the entities and relationships in the graph into low-dimensional vectors to establish a metabolite-nutrient association graph;

[0033] S3.4: Based on the identified abnormal metabolites, use the metabolite-nutrient association knowledge graph to retrieve all nutrient paths that cause the current metabolic abnormality through the path search algorithm, and then verify whether the nutrients inferred in the metabolite-nutrient association graph show excess or deficiency in the patient's diet, and calculate the corresponding intake deviation rate;

[0034] S3.5: If both the metabolite standardized score and the intake deviation rate are higher or lower than the preset threshold, it indicates that the direction of nutritional intake abnormality is consistent with the metabolite abnormality, and calculate the comprehensive score of the causal path between various abnormal metabolites and the corresponding nutrients;

[0035] S3.6: Integrate the path score and the consistency test result, calculate the causal association strength and sort and output, and use the output high-score causal pairs as the key targets for nutritional intervention recommendations, and recommend focusing on regulating the intake of the nutrients to correct the corresponding metabolite abnormalities.

[0036] As a further solution of the present invention, the specific steps for the health prediction module to identify the causal impact of nutritional factors on cardiac function indicators are as follows:

[0037] S4.1: Collect the generated cross-modal unified features and metabolic analysis results, construct a structured data set. Among them, each input data in the structured data is mapped to a high-dimensional space by using a linear transformation to obtain an input embedding vector;

[0038] S4.2: Calculate the Query, Key, and Value matrices of each input embedding vector. According to the Query matrix and the Key matrix, calculate the self-attention weights between the input embedding vectors at different time steps through different attention heads, normalize the attention scores through the Softmax function, calculate the attention-weighted sum of each input embedding vector, and then splice the weighted sum results of each attention head to generate output features;

[0039] S4.3: Perform a non-linear transformation on the output features at each time step through the FFN network, perform layer normalization on the output features after the non-linear transformation, and then add the original global feature vector and the output features to generate a corresponding high-dimensional feature matrix, calculate the individual causal effect of each nutrient factor on the target cardiac function index, and establish a causal forest model;

[0040] S4.5: According to the individual effect estimates of each patient output by the causal forest model, identify the nutrient factors that play a major role in the prediction, and use SHAP to calculate the attention-weighted importance contribution value of each nutrient factor. The larger the importance contribution value, the more significant the influence of the nutrient factor on the predicted cardiac function index of the current patient;

[0041] S4.6: Based on the importance contribution values of each nutrient factor, identify the key nutrient factors that have both significant causal effects and high explanatory contribution degrees, and visually output the key factors, causal effect directions, and recommended intervention intensities of each patient.

[0042] As a further solution of the present invention, the specific steps for the explanation feedback module to dynamically adjust the explanation depth are as follows:

[0043] S5.1: Construct a discrete hierarchical space of explanation granularity. The higher the number of layers in the hierarchical space, the deeper the semantics and the stronger the professionalism. Then evaluate the patient's health literacy score and the patient's cognitive processing speed score, establish the patient's cognitive characteristics, and set the fitness function of the explanation level and the cognitive vector;

[0044] S5.2: Randomly generate multiple initial solutions from the discrete hierarchical space of explanation granularity to form an explanation population. Each group of individuals in the population represents any selected explanation level. For the explanation level corresponding to each group of individuals, calculate its matching degree with the patient's cognitive characteristics through the fitness function. The higher the fitness, the more suitable the level is for the patient;

[0045] S5.3: Select the individual with the highest fitness as the optimal interpretation level found in the current search. Then calculate the search weights of the remaining individuals and the individual with the highest fitness. If the search weight of the remaining individuals is less than 1, update the selected interpretation level around the individual with the highest fitness based on the mechanism of surrounding the prey. If the search weight is greater than 1, randomly select an interpretation level based on the mechanism of searching for the prey, and update the selected interpretation level based on the selected interpretation level;

[0046] S5.4: Repeat the fitness calculation, optimal individual selection, and individual position update until the optimal fitness in the population no longer improves after multiple iterations. Take the interpretation level represented by the individual with the highest fitness in the current population as the personalized interpretation depth to be adopted in this recommendation, and automatically adjust the currently used terms, logical chains, and visualization methods.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] The present invention extracts the patient's diet information through image recognition and speech transcription technologies, uses CNN for food detection and classification, combines depth estimation and nutritional conversion tables to calculate the intake of various nutrients, and constructs a dynamic nutritional intake time series and 3D heat map by integrating text information. By standardizing metabolites and constructing a metabolite-nutrient knowledge graph with multiple databases, identifying the causal paths of nutritional and metabolic abnormalities, and then using causal forests and SHAP-Transformers to evaluate the individual causal effects and importance of nutritional factors on cardiac function indicators, extracting key nutritional factors, providing intervention suggestions, and then adaptively selecting the interpretation level based on the patient's cognitive characteristics to achieve multi-level and personalized result presentation and nutritional management support, which can comprehensively depict the patient's nutritional behavior and physiological state, significantly improve the accuracy and efficiency of nutritional assessment, achieve accurate identification of nutritional intervention targets, and improve patient understanding and compliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0050] Figure 1 It is a system block diagram of a chronic heart failure nutritional health assessment system based on machine learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0052] Example 1, refer toFigure 1 , a nutrition and health assessment system for chronic heart failure based on machine learning, including a data collection and processing module, a feature fusion module, a diet recognition module, a metabolic analysis module, a risk assessment module, a health prediction module, a nutrition recommendation module, an explanation and feedback module, a simulation prediction module, and a feedback monitoring module.

[0053] The data collection and processing module is used to collect structured data, unstructured data, and time-series data, and preprocess various types of data through format standardization, missing value completion, and semantic unification; the feature fusion module synchronously aligns various types of preprocessed data to generate cross-modal unified features.

[0054] Specifically, the structured data is embedded and represented through a fully connected neural network, the unstructured data is encoded through CNN and Transformer respectively, then the time-series data is modeled through a bidirectional LSTM, and the feature data generated after different data encoding processes for various types of data is collected. The attention weights of various types of feature data at each moment are calculated based on the time attention mechanism, and various types of feature data are weighted and fused based on the attention weights at each moment. Then, various types of feature data are mapped to the same dimension, and the spatial attention weights of various types of feature data in the current state are calculated. According to the calculated spatial attention weights, the information of each modality is weighted and fused. Through the modality-time joint splicing operation, various types of feature data processed by the time attention mechanism and the spatial attention mechanism are spliced into an overall feature, and the spliced overall feature is mapped into the same shared space, and then the spliced overall feature is normalized to generate the final cross-modal unified feature.

[0055] It should be further noted that the structured data specifically includes electronic medical records, laboratory test results, drug usage information, and diet questionnaire scale scores; the unstructured data specifically includes diet images, voice descriptions, and free-text medical records or nursing records; the time-series data specifically includes wearable device monitoring data, nutrition intake logs, daily weight change data, and drug administration time records.

[0056] The diet recognition module identifies the actual diet content of the patient and constructs a 3D nutrition map.

[0057] Specifically, the diet recognition module encodes the cross-modal unified features through a CNN feature extractor into a feature map with semantic information, constructs multiple sliding windows of size N×N on the feature map, predicts the object existence probabilities and position offsets of multiple anchor boxes at each sliding window position, selects the regions with high existence probabilities as candidate boxes, uses ROI Align to precisely crop and resample each candidate box into a unified-size feature matrix, then determines the specific food categories of each candidate box and further corrects the positions of the candidate boxes. After that, it retains the one with the highest classification probability among multiple candidate boxes of the same type of object, eliminates other candidate boxes with an overlap degree higher than the set threshold, calculates the actual portion of the food within the candidate box based on the image depth information in the candidate box, calculates the corresponding food mass using a density conversion table, calculates the nutrient content corresponding to each type of food based on its weight, converts the patient's speech into text using speech-to-text technology, then extracts the food names and intakes in each text data using entity recognition technology, fuses the extracted data with the image recognition results, records the nutrient intake of the patient at each moment of each day to construct a corresponding dynamic time series, and maps the nutrient-time intensity information to a three-dimensional space of nutrient dimension, time dimension, and intake intensity to form a corresponding 3D heat map to reflect the structure, frequency, and risk patterns of the patient's daily nutritional intake behavior in real time.

[0058] The metabolic analysis module is used to identify potential causal relationships between nutritional imbalance and metabolic abnormalities.

[0059] Specifically, through Z-score normalization, the original metabolite concentration data is converted into a unified numerical range, and the standardized scores of various metabolites are calculated. If the standardized score of any metabolite is higher than the preset threshold, it is determined that the metabolite is abnormal. Metabolite and nutrient information is extracted from public databases such as HMDB, KEGG, ChEBI, PubChem, and USDA, and the embedding distances of synonymous entity pairs from different sources are calculated by the word embedding method. If the embedding distance is less than the preset threshold, the two entities are considered equivalent and merged into a unified entity node. Semantic relationships are extracted from the text data by the NLP method, and based on the annotation relationships in each database, a set of relationship categories is constructed. The standardized entities and the extracted relationships are recorded as triples in the graph in a structured manner. Then, the entities and relationships in the graph are encoded into low-dimensional vectors to establish a metabolite-nutrient association graph. Based on the identified abnormal metabolites, using the metabolite-nutrient association knowledge graph, all nutrient paths leading to the current metabolic abnormality are retrieved through the path search algorithm. Then, it is verified whether the nutrients inferred in the metabolite-nutrient association graph show excess or deficiency in the patient's diet, and the corresponding intake deviation rate is calculated. If both the metabolite standardized score and the intake deviation rate are higher or lower than the preset threshold, it indicates that the direction of nutritional intake abnormality is consistent with the metabolite abnormality. The comprehensive scores of the causal paths between various abnormal metabolites and the corresponding nutrients are calculated, the path scores and the results of the consistency test are integrated, the causal association strength is calculated and sorted for output, and the high-score causal pairs output are used as the key targets for nutritional intervention recommendations, and it is recommended to focus on regulating the intake of this nutrient to correct the corresponding metabolite abnormality.

[0060] The risk assessment module evaluates the patient's nutritional risk level based on the cross-modal unified features and metabolic analysis results, and at the same time outputs a corresponding risk level report and highlights the high-risk nutritional indicators.

[0061] Example 2, referring to Figure 1 , a chronic heart failure nutritional health assessment system based on machine learning, including a data acquisition and processing module, a feature fusion module, a diet recognition module, a metabolic analysis module, a risk assessment module, a health prediction module, a nutrition recommendation module, an explanation and feedback module, a simulation prediction module, and a feedback monitoring module.

[0062] The health prediction module is used to identify the causal impact of nutritional factors on cardiac function indicators.

[0063] Specifically, collect the generated cross-modal unified features and metabolic analysis results to construct a structured dataset. Among them, each input data in the structured data is mapped to a high-dimensional space by using a linear transformation to obtain an input embedding vector. Calculate the Query, Key, and Value matrices of each input embedding vector. According to the Query matrix and the Key matrix, calculate the self-attention weights between the input embedding vectors at different time steps through different attention heads. Normalize the attention scores through the Softmax function, and calculate the attention-weighted sum of each input embedding vector. Then concatenate the weighted sum results of each attention head to generate an output feature. Perform a non-linear transformation on the output feature of each time step through an FFN network, perform layer normalization on the non-linearly transformed output feature, and then add the original global feature vector to the output feature to generate a corresponding high-dimensional feature matrix. Calculate the individual causal effect of each nutrient factor on the target cardiac function index, establish a causal forest model, and estimate the individualized effect of each patient according to the output of the causal forest model. Identify the nutrient factors that play a major role in the prediction, and use SHAP to calculate the attention-weighted importance contribution value of each nutrient factor. The larger the importance contribution value, the more significant the influence of the nutrient factor on the predicted cardiac function index of the current patient. Based on the importance contribution values of each nutrient factor, identify the key nutrient factors that have both significant causal effects and high explanatory contribution degrees, and visually output the key factors, causal effect directions, and recommended intervention intensities of each patient.

[0064] The nutrition recommendation module generates a diet plan that takes into account nutritional needs, taste preferences, and ingredient availability based on the risk level report and causal analysis results; the explanation feedback module dynamically adjusts the depth of explanation according to the patient's health literacy and cognitive level.

[0065] Specifically, a discrete hierarchical space of interpretation granularity is constructed, where the higher the number of layers in the hierarchical space, the deeper the semantics and the stronger the professionalism. Then, the patient's health literacy score and the patient's cognitive processing speed score are evaluated, and the patient's cognitive characteristics are established. An adaptation function between the interpretation level and the cognitive vector is set. Multiple initial solutions are randomly generated from the discrete hierarchical space of interpretation granularity to form an interpretation population. Each group of individuals in the population represents any selected interpretation level at present. For the interpretation level corresponding to each group of individuals, the matching degree between it and the patient's cognitive characteristics is calculated through the adaptation function. The higher the adaptation degree, the more suitable the level is for the patient. The individual with the highest adaptation degree is selected as the optimal interpretation level found in the current search. Then, the search weights of the remaining individuals and the individual with the highest adaptation degree are calculated. If the search weight of the remaining individuals is less than 1, based on the mechanism of surrounding the prey, the selected interpretation level is updated around the individual with the highest adaptation degree. If the search weight is greater than 1, based on the mechanism of searching for the prey, an interpretation level is randomly selected, and the selected interpretation level is updated based on the selected interpretation level. The adaptation degree calculation, the selection of the optimal individual, and the update of the individual position are repeated until the optimal adaptation degree in the population no longer improves after multiple iterations. The interpretation level represented by the individual with the highest adaptation degree in the current population is used as the personalized interpretation depth to be adopted in this recommendation, and the currently used terms, logical chains, and visualization methods are automatically adjusted.

[0066] The simulation prediction module is used to construct a digital twin of the patient individual for multi-scale simulation; the feedback monitoring module extracts new information from wearable devices, diet record APPs, and regular detection data, continuously updates the risk assessment and recommendation plan, and records the actual implementation of the recommendation plan by the patient.

Claims

1. A machine learning-based nutritional health assessment system for chronic heart failure, characterized by: It includes acquisition and processing module, feature fusion module, diet recognition module, metabolic analysis module, risk assessment module, health prediction module, nutrition recommendation module, explanation and feedback module, simulation prediction module and feedback monitoring module; The acquisition and processing module is used to collect structured data, unstructured data and time series data, and pre-process various types of data through format standardization, missing value completion and semantic unification; The feature fusion module synchronously aligns various types of preprocessed data to generate cross-modal unified features; The diet recognition module recognizes the actual diet content of the patient and constructs a 3D nutritional map; The metabolic analysis module is used to identify potential causal relationships between nutritional imbalance and metabolic abnormalities; The risk assessment module assesses the patient's nutritional risk level based on cross-modal unified features and metabolic analysis results, outputs a corresponding risk level report, and highlights high-risk nutritional indicators; The health prediction module is used to identify the causal effects of nutritional factors on cardiac function indicators; The nutrition recommendation module generates a diet plan that takes into account nutritional needs, taste preferences and food availability based on the risk level report and causal analysis results; The explanation feedback module dynamically adjusts the explanation depth according to the patient's health literacy and cognitive level; The simulation prediction module is used to construct a digital twin of an individual patient and perform multi-scale simulation; The feedback monitoring module extracts new information from wearable devices, diet record apps, and regular test data, and continuously updates risk assessments and recommendations, while also recording patients’ actual implementation of the recommendations.

2. The machine learning-based chronic heart failure nutrition and health assessment system according to claim 1, characterized in that: The specific steps of the feature fusion module to generate cross-modal unified features are as follows: S1.1: Structured data is embedded and represented through a fully connected neural network, unstructured data is encoded through CNN and Transformer respectively, and then bidirectional LSTM is used to perform sequence modeling on time series data, and feature data generated by various types of data after different data encoding processes are collected; S1.2: Based on the temporal attention mechanism, the attention weights of various feature data at each moment are calculated, and the various feature data are weighted and fused based on the attention weights at each moment. After that, the various feature data are mapped to the same dimension, and the spatial attention weights of various feature data in the current state are calculated. According to the calculated spatial attention weights, the information of each modality is weighted and fused; S1.3: Through the modal-temporal joint splicing operation, various feature data processed by the temporal attention mechanism and the spatial attention mechanism are spliced ​​into overall features, and the overall features generated by the splicing are mapped to the same shared space, and then the spliced ​​overall features are normalized to generate the final cross-modal unified features.

3. The machine learning-based chronic heart failure nutrition and health assessment system according to claim 2, characterized in that: The specific steps of the diet recognition module identifying the actual diet content of the patient and constructing a 3D nutritional map are as follows: S2.1: The diet recognition module encodes the cross-modal unified features into a feature map with semantic information through the CNN feature extractor, and constructs multiple groups of N×N sliding windows on the feature map. For each sliding window position, it predicts the object existence probability and position offset of multiple anchor boxes, and selects the area with high existence probability as the candidate box; S2.2: Use ROI Align to accurately crop and resample each candidate frame into a uniform-sized feature matrix, then determine the specific food category of each candidate frame, and further correct the position of the candidate frame. Then, retain the one with the highest classification probability among multiple candidate frames of the same object, and remove other candidate frames whose overlap with it exceeds the set threshold. Then, calculate the actual weight of the food in the candidate frame based on the image depth information in the candidate frame, and use the density conversion table to calculate the corresponding food mass; S2.3: Calculate the corresponding nutrient content of each type of food based on its weight, use speech transcription technology to convert the patient's speech into text, then use entity recognition technology to extract the food name and intake from each text data, and fuse the extracted data sets with the image recognition results; S2.4: Record the patient's nutrient intake at every moment of the day to construct a corresponding dynamic time series, and map the nutrient-time intensity information to the three-dimensional space of nutrient dimension, time dimension, and intake intensity to form a corresponding 3D heat map to reflect the structure, frequency, and risk pattern of the patient's daily nutritional intake behavior in real time.

4. The machine learning-based chronic heart failure nutrition and health assessment system according to claim 3, characterized in that: The specific steps of the metabolic analysis module to identify the potential causal relationship between nutritional imbalance and metabolic abnormalities are as follows: S3.1: Through Z-score normalization, the original metabolite concentration data is converted into a unified numerical range, and the standardized scores of various metabolites are calculated. If the standardized score of any metabolite is higher than the preset threshold, the metabolite is judged to be abnormal; S3.2: Extract metabolite and nutrient information from public databases such as HMDB, KEGG, ChEBI, PubChem, and USDA, and calculate the embedding distance of synonymous entity pairs from different sources using the word embedding method. If the embedding distance is less than the preset threshold, the two entities are considered equivalent and merged into a unified entity node; S3.3: Extract semantic relations from text data through NLP methods, construct a set of relation categories based on the annotated relations in each database, record the standardized entities and extracted relations as triplets in the graph in a structured manner, and then encode the entities and relations in the graph into low-dimensional vectors to establish a metabolite-nutrient association map; S3.4: Based on the abnormal metabolites identified, the metabolite-nutrient association knowledge graph is used to retrieve all nutrient pathways that lead to the current metabolic abnormality through a path search algorithm. Then, it is verified whether the nutrients inferred from the metabolite-nutrient association graph are excessive or deficient in the patient's diet, and the corresponding intake deviation rate is calculated. S3.5: If the metabolite standardized score and intake deviation rate are both higher or lower than the preset threshold, it means that the abnormal nutritional intake is consistent with the abnormal metabolite direction, and the comprehensive score of the causal path between each abnormal metabolite and the corresponding nutrient is calculated; S3.6: Integrate the pathway score and consistency test results, calculate the causal association strength and rank the output, and use the high-scoring causal pairs as the key targets for nutritional intervention recommendations, and recommend focusing on regulating the intake of this nutrient to correct the corresponding metabolite abnormalities.

5. The machine learning-based chronic heart failure nutrition and health assessment system according to claim 1, characterized in that: The specific steps of the health prediction module for identifying the causal effect of nutritional factors on cardiac function indicators are as follows: S4.1: Collect the generated cross-modal unified features and metabolic analysis results to construct a structured data set, wherein each input data in the structured data is mapped to a high-dimensional space by using a linear transformation to obtain an input embedding vector; S4.2: Calculate the query, key, and value matrices of each input embedding vector. Based on the Query matrix and the Key matrix, calculate the self-attention weights between the input embedding vectors at different time steps through different attention heads, normalize the attention scores through the Softmax function, and calculate the weighted sum of the attention of each input embedding vector. Then concatenate the weighted sum results of each attention head to generate the output features. S4.3: Perform nonlinear transformation on the output features of each time step through the FFN network, perform layer normalization on the output features after nonlinear transformation, add the original global feature vector to the output features, generate the corresponding high-dimensional feature matrix, calculate the individual causal effect of each nutritional factor on the target cardiac function index, and establish a causal forest model; S4.5: Based on the causal forest model, output the individualized effect estimate for each patient, identify the nutritional factors that play a major role in the prediction, and use SHAP to calculate the attention-weighted importance contribution value of each nutritional factor. The larger the importance contribution value, the more significant the impact of the nutritional factor on the current patient's predicted cardiac function index; S4.6: Based on the importance contribution value of each nutritional factor, identify the key nutritional factors that have both significant causal effects and high explanatory contributions, and visualize the key factors, causal effect directions and recommended intervention intensity for each patient.

6. The machine learning-based chronic heart failure nutrition and health assessment system according to claim 1, characterized in that: The specific steps of dynamically adjusting the interpretation depth by the interpretation feedback module are as follows: S5.1: Construct a discrete hierarchical space of explanation granularity, where the higher the number of levels in the hierarchical space, the deeper the semantics and the stronger the professionalism. Then evaluate the patient's health literacy score and the patient's cognitive processing speed score, establish the patient's cognitive characteristics, and set the fitness function between the explanation level and the cognitive vector; S5.2: Randomly generate multiple initial solutions from the discrete hierarchical space of explanation granularity to form an explanation population. Each group of individuals in the population represents any currently selected explanation level. For the explanation level corresponding to each group of individuals, the degree of matching with the patient's cognitive characteristics is calculated through the fitness function. The higher the fitness, the more suitable the level is for the patient. S5.3: Select the individual with the highest fitness as the optimal explanation level found in the current search, and then calculate the search weights of the remaining individuals and the individual with the highest fitness. If the search weights of the remaining individuals are less than 1, then based on the surrounding prey mechanism, the selected explanation level is updated around the individual with the highest fitness. If the search weight is greater than 1, then based on the search prey mechanism, the explanation level is randomly selected, and based on the selected explanation level, the selected explanation level is updated; S5.4: Repeat the fitness calculation, optimal individual selection and individual position update until the optimal fitness in the population no longer improves after multiple rounds of iterations. The explanation level represented by the individual with the highest fitness in the current population is used as the personalized explanation depth to be used in this recommendation, and the currently used terms, logical chains and visualization methods are automatically adjusted.

Images