A hybrid model driven nutrition risk prediction method and system for precise nutrition intervention

Abstract

This invention discloses a hybrid model-driven nutritional risk prediction method and system for precision nutritional intervention. The method constructs a large-scale patient typing model containing multiple typical patient types and their associated assessment parameter packages, integrating a standardized first screening model with a personalized second screening model. In application, the first screening model is first used to standardize and initially screen patient data. Then, the large-scale patient typing model extracts patient feature inputs based on the patient data and initial screening results for typing and obtains the corresponding assessment parameter packages. The second screening model is then personalized based on these parameter packages, supplemented with key data, and combined with personalized thresholds to calculate the final risk level and generate a comprehensive assessment report. This invention achieves dynamic individualized adaptation of assessment strategies, improves the accuracy and clinical interpretability of risk identification, and optimizes workflow and resource allocation.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and digital health, specifically to a nutritional risk assessment method, and more particularly to a hybrid model-driven nutritional risk prediction method and system for precision nutritional intervention. Background Technology

[0002] Nutritional status is closely related to clinical outcomes in hospitalized patients. Malnutrition significantly increases the incidence of complications, prolongs hospital stays, and affects treatment effectiveness. Therefore, timely and accurate nutritional risk screening and assessment of patients is a prerequisite for implementing effective clinical nutritional support.

[0003] Currently, standardized nutritional risk screening tools, such as the Nutritional Risk Screening 2002 (NRS2002) and the Universal Screening Tool for Malnutrition (MUST), are widely used in clinical practice. These tools, based on a fixed set of scoring rules (such as age, disease severity, weight change, and food intake), provide healthcare professionals with a rapid and unified assessment framework, greatly improving screening efficiency and ensuring, to some extent, the consistency of assessments. However, this "one-size-fits-all" standardized assessment model has significant limitations. First, standardized assessments cannot fully reflect the individual heterogeneity of patients. For example, patients with different primary diseases (such as tumors, post-surgery, and chronic organ failure) exhibit significant differences in the mechanisms of malnutrition, key risk factors, and responses to nutritional interventions, but existing scoring systems struggle to fully quantify and incorporate these differences into the assessment. Second, the assessment results are disconnected from subsequent precise nutritional intervention decisions. Standardized scoring can only indicate "the existence of risk," but it cannot provide direct, individualized decision-making basis for "how to intervene" (such as choosing which nutritional support pathway or specific nutrient ratios), still heavily relying on the personal experience of clinicians.

[0004] In recent years, with the development of medical big data and artificial intelligence technologies, researchers have begun to explore the use of machine learning models for more refined nutritional risk prediction. These data-driven methods can uncover the complex nonlinear relationships between multidimensional patient characteristics (including numerous laboratory indicators, imaging features, electronic medical record texts, etc.) and nutritional outcomes, theoretically providing more personalized risk assessments. However, purely data-driven models also face challenges in practical applications: First, models typically operate as a "black box," lacking interpretability in their decision-making process, making it difficult to gain the trust of clinicians and integrate with existing, evidence-based standard workflows; second, model training and deployment often require a large amount of high-quality labeled data, which is difficult to achieve under regulations protecting patient privacy and data security, leading to overfitting and poor generalization ability when training on data from a single institution.

[0005] In summary, existing technologies suffer from a core contradiction: standardized, interpretable screening tools lack personalized precision, while highly personalized data-driven models are difficult to integrate with standardized clinical pathways and present privacy and generalization challenges. How to construct a new paradigm for nutritional risk assessment that inherits the efficiency and interpretability advantages of standardized assessments, deeply integrates personalized data insights, and enables continuous and safe evolution has become a critical technical problem urgently needing to be solved in the fields of clinical nutrition and medical informatics. Summary of the Invention

[0006] This invention aims to overcome the inherent contradiction between standardized nutritional assessment tools and personalized data-driven models in the prior art, and provides a hybrid model-driven nutritional risk prediction method and system for precision nutritional intervention. By constructing an intelligent bridge of "patient subtyping-parameter configuration", it integrates standardized screening and personalized assessment, and outputs accurate, interpretable and intervention-guiding dynamic risk assessment results.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a hybrid model-driven method for predicting nutritional risks for precision nutritional intervention, the method comprising the following steps: S100. Based on the patient population's historical clinical characteristics data, nutritional risk scores, and nutritional intervention outcome data, a large-scale patient classification model is constructed using unsupervised clustering algorithms and supervised association analysis. The large-scale patient classification model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package at least includes feature weight configurations, data collection templates, and risk assessment thresholds for the corresponding typical patient type. S201. Obtain basic clinical data of the target patients and use a pre-deployed first screening model based on preset standardized rules to obtain a standardized nutritional risk score and the corresponding preliminary risk level. S202. Input the basic clinical data and preliminary risk level of the target patient into the patient classification model to determine the target patient type; obtain the assessment parameter package associated with the target patient type from the patient classification model; S203. Based on the feature weight configuration in the assessment parameter package, a pre-deployed second screening model is configured in a personalized manner; the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package; the second screening model calculates and outputs a personalized nutritional risk score and the corresponding final risk level based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment threshold in the assessment parameter package. S300. Integrate the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

[0008] Optionally, in step S100 of the above method, the supervised association analysis specifically includes: analyzing the nutritional intervention outcomes of historical patients associated with each typical patient type, configuring one or more recommended nutritional intervention programs for each typical patient type, and incorporating the recommended nutritional intervention programs as components of the assessment parameter package.

[0009] Optionally, in step S202 of the above method, determining the target patient type to which the target patient belongs specifically includes: calculating the similarity between the basic clinical data and preliminary risk level and the feature representation of each typical patient type, and selecting one or more types with the highest similarity as the target patient type.

[0010] Optionally, in step S203 of the above method, the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package, specifically including: Compare the existing clinical data of the target patient with the data items defined in the data collection template to identify data item gaps; Based on the data item gaps, a data acquisition request is generated and sent.

[0011] Optionally, the preset standardized rules used by the first screening model are the scoring rules in NRS2002 or the MUST nutrition risk screening tool.

[0012] Optionally, in the above method, step S203, calculating a personalized nutritional risk score based on the weighted clinical characteristic data and the supplementary clinical data specifically includes: A trained risk prediction sub-model is invoked for processing; the structure or hyperparameters of the risk prediction sub-model are defined by the evaluation parameter package.

[0013] Optionally, the above method further includes step S400: feeding the comprehensive nutritional risk assessment report generated in step S300, as well as the actual nutritional intervention plan and final outcome data of the target patient, into the historical data of the patient group as new training samples for periodically updating the large patient classification model.

[0014] Optionally, the periodic updating of the patient subtyping model described above is performed using a federated learning framework, specifically including: At each local node, the locally maintained large-scale patient subtyping model is incrementally trained using the newly added training samples. Upload the updated parameters of the trained model to the central node; At the central node, model update parameters from multiple local nodes are securely aggregated to generate a global model update. The global model update is distributed to each local node.

[0015] Optionally, the assessment parameter package further includes a dynamic scoring rule template, used in step S203 to dynamically generate or adjust temporary scoring rules for the target patient based on the collected supplementary clinical data; the second screening model, based on the dynamic scoring rule template, performs secondary processing and scoring fusion on the weighted clinical feature data and the supplementary clinical data to calculate the personalized nutritional risk score.

[0016] In a second aspect, the present invention provides a hybrid model-driven nutrition risk prediction system for precision nutrition intervention, the system comprising: The large model building module is used to construct a large patient classification model based on the patient population's historical clinical characteristic data, nutritional risk scores, and nutritional intervention outcome data, through unsupervised clustering algorithms and supervised association analysis. The large patient classification model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package includes at least the feature weight configuration, data collection template, and risk assessment threshold for the corresponding typical patient type. The standardized initial screening module is used to acquire basic clinical data of target patients and use a pre-deployed first screening model based on preset standardized rules to obtain a standardized nutritional risk score and corresponding preliminary risk level. The patient subtyping and parameter matching module is used to input the basic clinical data and preliminary risk level of the target patient into the patient subtyping model to determine the target patient type to which the target patient belongs; and to obtain the assessment parameter package associated with the target patient type from the patient subtyping model. The personalized assessment module is used to configure a pre-deployed second screening model based on the feature weight configuration in the assessment parameter package; the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package; the second screening model calculates and outputs a personalized nutritional risk score and the corresponding final risk level based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment threshold in the assessment parameter package. The comprehensive report generation module is used to integrate the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects: (1) In this invention, by dynamically configuring feature weights, data collection templates, and risk assessment thresholds based on patient type assessment parameter packages, the system can automatically focus on key indicators such as anastomosis status and drainage fluid properties for specific patient groups, such as those in the early postoperative period after gastrointestinal surgery, and apply a dedicated questionnaire. This allows the final risk level to more accurately reflect the individual's true risk under specific pathophysiological conditions, effectively avoiding misjudgments caused by general standards. Overall, it achieves dynamic individualized adaptation of assessment strategies, significantly improving the accuracy of risk discrimination.

[0018] (2) In this invention, the pre-deployed first screening model performs rapid initial screening based on preset standardized rules, ensuring the breadth and efficiency of the screening. Only when the initial risk level reaches a threshold is a deeper, personalized assessment requiring the loading of an assessment parameter package triggered. This design forms an efficient funnel-shaped workflow, avoiding the resource consumption of comprehensive in-depth assessments of all patients. Simultaneously, supplementary data collection based on the data collection templates in the parameter package achieves targeted and minimized data collection, reducing the burden on medical staff. Overall, a tiered assessment process is constructed, optimizing clinical work efficiency and resource allocation.

[0019] (3) In this invention, the comprehensive report output by the system simultaneously includes a standardized nutritional risk score based on general rules and a personalized analysis based on a large patient subtyping model. The report can clearly explain the decision-making logic, for example: the patient is classified as 'tumor cachexia', and according to this type parameter, the weights of its 'inflammatory markers' and 'muscle loss' features are relatively high, so the risk level is increased. This structured knowledge, derived from unsupervised clustering and supervised association analysis, makes the assessment process transparent and the results traceable, greatly enhancing clinicians' trust in intelligent assisted decision-making. Overall, it provides multi-level, structured explanatory evidence, enhancing clinical trust and acceptance.

[0020] (4) In this invention, by associating nutritional intervention outcome data during the model building phase, the system possesses the ability to learn from historical efficacy. When new assessment reports and actual intervention outcomes are fed back to the model, it can drive the iterative updating of the patient subtyping model and its assessment parameter package. For example, when a certain type of patient responds well to a specific nutritional regimen, the system can automatically optimize the risk threshold or intervention recommendation for that type, thereby achieving dynamic alignment and co-evolution between assessment strategies and clinical practice effects. Overall, a continuous evolution mechanism based on real-world evidence is established.

[0021] (5) In this invention, core knowledge is encapsulated in a structured patient subtyping model and parameter package, rather than relying on raw patient data. This modular expression makes it possible to achieve cross-institutional knowledge sharing and collaborative optimization by securely exchanging or aggregating these model parameters while strictly protecting data privacy, effectively overcoming the bottleneck of model generalization ability caused by the scarcity of data in a single institution. Overall, a modular knowledge expression suitable for cross-institutional collaboration is formed, which helps model generalization and privacy protection. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of the operation of each model of the method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the specific process of step S100 in the method provided in the embodiment of the present invention; Figure 4 This is a schematic flowchart of steps S201 to S203 in the method provided in the embodiments of the present invention; Figure 5 This is a flowchart illustrating step S400 in the method provided in this embodiment of the invention; Figure 6 This is a functional block diagram of the system provided in the embodiments of the present invention. Detailed Implementation

[0023] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] like Figures 1 to 5 As shown, embodiments of the present invention provide a hybrid model-driven method for predicting nutritional risks for precision nutritional intervention. The method includes the following steps: S100. Based on historical clinical characteristic data, nutritional risk scores, and nutritional intervention outcome data of the patient population, a large-scale patient subtyping model is constructed using unsupervised clustering algorithms and supervised association analysis. Specifically, the supervised association analysis includes analyzing the nutritional intervention outcomes of historical patients associated with each typical patient type, and configuring one or more recommended nutritional intervention plans for each typical patient type. The large-scale patient subtyping model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package includes at least the feature weight configuration for the corresponding typical patient type, a data collection template, a risk assessment threshold, and the recommended nutritional intervention plan. Optionally, the assessment parameter package also includes a dynamic scoring rule template for subsequently dynamically generating temporary scoring rules.

[0025] S201. Obtain basic clinical data of the target patient, and use a pre-deployed first screening model based on preset standardized rules to obtain a standardized nutritional risk score and corresponding preliminary risk level. The preset standardized rules are scoring rules from NRS2002 or the MUST nutritional risk screening tool.

[0026] S202. Extract the basic clinical data and preliminary risk level of the target patient and input them into the patient classification model. The patient classification model calculates the similarity between the basic clinical data and preliminary risk level and the feature representations of each typical patient type, and selects one or more types with the highest similarity as the target patient type to which the target patient belongs. From the patient classification model, it obtains the assessment parameter package associated with the target patient type.

[0027] S203. Based on the feature weight configuration in the assessment parameter package, a pre-deployed second screening model is configured in a personalized manner. The second screening model, according to the data collection template in the assessment parameter package, identifies data gaps by comparing the target patient's existing clinical data with the data items defined in the data collection template, and generates and sends a data collection request based on these gaps to collect supplementary clinical data from the target patient. Based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment threshold in the assessment parameter package, the second screening model calculates and outputs a personalized nutritional risk score (hereinafter referred to as "personalized score") and the corresponding final risk level. Specifically, calculating the personalized score includes: calling a trained risk prediction sub-model for processing; the structure or hyperparameters of the risk prediction sub-model are defined by the assessment parameter package. If the assessment parameter package contains a dynamic scoring rule template, the second screening model also dynamically generates or adjusts temporary scoring rules for the target patient based on the collected supplementary clinical data, and performs secondary processing and score fusion on the weighted data to calculate the personalized score.

[0028] S300. Integrate the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

[0029] S400: The comprehensive nutritional risk assessment report generated in step S300, along with the actual nutritional intervention plan and final outcome data of the target patient, are fed back as new training samples to the historical data of the patient population. The patient subtyping model is periodically updated using a federated learning framework, specifically including: at each local node, incremental training of the locally maintained patient subtyping model is performed using the newly added training samples. The trained model update parameters are uploaded to the central node. At the central node, the model update parameters from multiple local nodes are securely aggregated to generate a global model update. The global model update is distributed to each local node.

[0030] It is worth noting that the operation of this invention relies on three key types of data, which together form the cornerstone of the system's learning and application. First, there is clinical characteristic data used for rapid initial screening. For example, when a patient is admitted after gastric cancer surgery, the system automatically extracts basic information from the medical record, such as age (65 years old), body mass index (BMI 18.5), disease diagnosis and severity (subtotal gastrectomy). This data strictly follows the rules of standardized tools such as NRS2002 and can be directly used for rapid scoring to form preliminary results of risk screening.

[0031] Secondly, there is supplementary clinical data used for in-depth assessment. After initial screening indicates risk, the system will guide the collection of more in-depth information for further subtyping and evaluation. Taking the above patient as an example, this may include postoperative inflammatory markers (such as C-reactive protein), anastomotic condition assessment, drainage fluid properties and volume, body composition analysis data, etc. This data goes beyond the scope of the standard questionnaire and can finely characterize the patient's specific pathophysiological state such as "high inflammatory response" or "anastomotic leakage risk," providing "detailed evidence" for personalized judgment.

[0032] Finally, there is the nutritional intervention outcome data used for model learning and optimization. The core of system construction and iteration lies in learning "what type of patient, what intervention, and what outcome." For example, historical data records several cases of "early post-gastrointestinal surgery" patients with similar characteristics to those mentioned above. They received different nutritional support regimens (such as routine enteral nutrition or specific formula-based immunonutrition), and their complication rates, length of hospital stay, and improvements in nutritional indicators were recorded. By analyzing this data, the system can summarize the most effective assessment weights and intervention recommendations for this "early post-gastrointestinal surgery" patient group.

[0033] Based on the accumulation of the above data, the system constructs a "large-scale patient subtyping model" through step S100. This process can be simply understood as follows: 1) Data preparation and preprocessing: Integrating the clinical characteristics of historical patients, supplementary collected data, and corresponding outcome data, and performing cleaning and standardization; 2) Unsupervised clustering: Automatically discovering inherent patterns in the data through algorithms, and classifying patients into multiple typical types with common characteristics, such as "high inflammatory response type" and "sarcopenia-dominated type"; 3) Supervised association analysis: For each type, analyzing the response of its patient population to different nutritional interventions, thereby configuring personalized assessment parameters (such as key feature weights, risk thresholds, and recommended intervention plans) for the corresponding typical patient types; 4) Model integration: Encapsulating all types and their assessment parameters into a callable intelligent model.

[0034] In application, the system first uses standardized rules (such as NRS2002) to perform initial screening of target patients (S201). If the initial screening indicates risk, a deeper process is initiated: the system matches patient data with a large patient subtyping model to determine its type and loads the corresponding assessment parameter package (S202); subsequently, supplementary data collection is guided by the assessment parameter package, and a personalized model is used for weighted calculation and threshold judgment to obtain the final risk level (S203). Finally, all results are summarized to generate a comprehensive report (S300). Through this process, the present invention achieves a closed loop from rapid universal screening to precise individualized assessment.

[0035] As in step S100, this embodiment specifically achieves this through four stages: data preparation and preprocessing, unsupervised clustering (discovering typical patient types), supervised association analysis (configuring intervention plans for each type), and model integration and validation.

[0036] Specifically, in data preparation and preprocessing, the following data of historical patients are extracted from systems such as Hospital Information System (HIS), Laboratory Information System (LIS), and Electronic Medical Record (EMR): (1) Clinical characteristic data: including demographic information, diagnostic information, vital signs, laboratory test results, imaging reports, medication records, etc.; (2) Nutritional risk score: using standardized tools (such as NRS2002, MUST) to retrospectively score historical patients, or extracting existing scores from medical records; (3) Nutritional intervention outcome data: recording the actual nutritional intervention plan received by the patient (such as enteral nutrition formula, parenteral nutrition composition, oral nutritional supplements, etc.) and the clinical outcome after intervention (such as complications, length of hospital stay, readmission rate, mortality rate, changes in nutritional indicators, etc.).

[0037] The collected clinical characteristic data and nutritional risk scores are transformed into a high-quality, structured historical characteristic matrix of the patient population that can be processed by machine learning algorithms. .

[0038] First, it is necessary to process the historical characteristic matrix of the patient population. The missing values ​​in the feature matrix are determined. Specifically, based on the feature similarity among patients, the KNN algorithm is used to provide a baseline estimate for each missing value. For feature j of patient i, the estimated imputation value for the missing feature in the feature matrix is ​​determined by the following formula: , For patients The The estimated imputed values ​​for each feature, Index the target patients for whom missing values ​​are to be filled. This is the index of a specific feature (column) in the data matrix. For target patients of The set of the most similar patients (derived by calculating distances based on other non-missing features). For in set The index of the middle loop, for example configured as 5 or 10, represents a similar patient; For similar patients The The actual observed values ​​of each feature Similarity weights reflect patient... With patients similarity, For example, it is configured as the reciprocal of the distance, in the form of the reciprocal of the Euclidean distance: , That is, the patient and The Euclidean distance across all non-missing features is used; the closer the distance, the greater the weight. A very small positive number (e.g., 10⁻⁶) is used to avoid division by zero errors. This estimate is then used as the initial value for multiple imputation. The uncertainty of missing data is simulated by iteratively fitting a regression model, generating multiple complete datasets for subsequent ensemble analysis, thereby more robustly handling complex missing patterns in medical data.

[0039] Then, the historical feature matrix of the patient population needs to be processed. Outliers. Specifically, the Isolation Forest algorithm is used to automatically identify and handle outliers. This algorithm measures the degree of anomalousness of a sample by constructing multiple isolation trees. The outlier score is determined by the following formula: , in, For patient samples The anomaly score, the closer the value is to 1, the higher the probability of an anomaly. Given a patient sample's feature vector, samples with anomaly scores exceeding a threshold are classified as abnormal. This threshold is a hyperparameter; for example, a threshold of 0.6 is used to classify anything exceeding this threshold as abnormal. Simultaneously, the abnormal feature values ​​are replaced with the median of that feature to avoid interfering with subsequent analyses. For the number of patient samples, Let x be the average path length of sample x in the isolated forest. This is a standardized constant for the path length, typically 1. , Here, represents the Euler constant. After handling missing and outlier values, the resulting historical characteristic matrix of the patient population undergoes data standardization and coding. Specifically, for continuous clinical characteristics, robust standardization is employed to eliminate dimensions and resist the influence of potential outliers. The standardization calculation formula is as follows: , in, For patients The The standardized values ​​of each feature For patients The Original values ​​of each feature. For all patients The median of the features, For all patients The interquartile range of each feature, i.e. For the morphological features, numerical values ​​are generated using either label encoding or one-hot encoding, depending on their characteristics.

[0040] Meaningful derived features are extracted from the raw data, such as disease categories from diagnostic codes, slopes of change from time-series laboratory indicators, and key symptom entities from text records using natural language processing techniques. To address the curse of dimensionality that high-dimensional features may bring and to improve subsequent clustering results, the UMAP algorithm is used for feature dimensionality reduction. UMAP constructs a feature map in a high-dimensional space and optimizes its low-dimensional representation, where the core weights are determined by the following formula: ), in, In high-dimensional space, patient samples With patient samples The connection weights between them represent similarity. For patient samples With patient samples The original distance between them (such as Euclidean distance). For patient samples The distance to its nearest neighbor is used to ensure that each point is connected to at least one other point. Let be a scale parameter, determined through optimization, such that ,in This is the set number of nearest neighbors, used to control the balance between local and global structure. It can reduce the feature matrix to a preset dimension (e.g., 10 dimensions) while maintaining the global and local structure of the data, resulting in a dimensionality-reduced historical feature matrix of the patient population. Each row represents a patient, and each column represents a feature after dimensionality reduction.

[0041] Specifically, in unsupervised clustering, the dimensionality-reduced historical feature matrix is... As input, a hierarchical density clustering algorithm is used for patient classification. This algorithm can automatically discover patient groups with arbitrary shapes and densities without pre-specifying the number of clusters, and can effectively distinguish core patients from noise points (atypical patients).

[0042] First, the inter-patient distance is calculated based on the input historical feature matrix. Specifically, a distance metric based on local density is constructed, and the distance metric is determined by the following formula: , in, Sample (patient) and The reachability between them is the core distance metric of this algorithm. Two patient samples, To obtain from the sample The distance to its k-th nearest neighbor defines the local density of sample $a$, with a smaller distance indicating a higher density. For the sample and The original distance between them This is the number of nearest neighbors considered when defining local density. and Quantify local density to identify different density clusters in patients.

[0043] Then, a minimum spanning tree is constructed based on the reach distance, from clusters of different densities, and based on the stability of the clusters, through... The persistence of clusters within a hierarchical structure is measured, and the most stable and significant clusters are selected. The stability of patient clusters is determined by the following formula: , in, For clusters The stability score indicates that the cluster is more persistent and reliable within the clustering hierarchy; a higher score indicates that the cluster is more persistent and reliable. This represents a patient cluster, or a typical patient type. For clusters One patient sample (point). In the clustering hierarchy, points The distance threshold when forming the current cluster. In the clustering hierarchy, points The distance threshold when leaving the current cluster. In a hierarchical clustering tree, samples Join / Leave Cluster The distance threshold at that time was used to extract the patient clusters with the highest statistical significance.

[0044] Ultimately, the algorithm automatically outputs K stable "typical patient types" and their corresponding patient type labels. For each type, the center vector of its feature prototypes is calculated, and each type is assigned a clinically understandable name based on its feature importance score. The feature importance score is determined by the following formula:

[0045] in, For the first Of the patient types, the first Importance score of each feature Index for patient types, Index of features For type All patients in the middle The mean of each feature, Let $j$ be the global mean of the $j$-th feature of all patients. For all patients The global standard deviation of each feature. The higher the feature importance score, the more typical and distinctive the feature is in this type. Each type is given a clinically understandable name. For example, the cluster with features of "high inflammatory markers, high metabolic rate, and low albumin" is named "high inflammatory response type"; the cluster with features of "low muscle mass, poor functional status, and significantly decreased grip strength" is named "sarcopenia-dominated type".

[0046] In unsupervised clustering, on the one hand, a complete hierarchical clustering tree is constructed based on the historical feature matrix and local density measurement, exploring the intrinsic structure of the data without presupposition. The tree encompasses all possible subtypes from fine-grained to coarse-grained. On the other hand, based on clinical knowledge, the most statistically significant patient subtypes are selected by evaluating the stability of each cluster in the hierarchical structure.

[0047] Specifically, in supervised association analysis, the patient type labels obtained from unsupervised clustering are... and the original historical feature matrix obtained in data preparation and preprocessing. Nutritional intervention outcome data The aim is to correlate these factors to identify the optimal assessment and intervention strategies for each patient type and to construct an assessment parameter package.

[0048] First, from the original historical feature matrix Nutritional intervention outcome data In, for each patient type ( This study aims to identify and recommend optimal nutritional support strategies, thereby uncovering recommended interventions. Specifically, to avoid confounding bias in observational data, a causal forest model combined with inverse probability weighting is used to estimate the net effect of each intervention on specific patient types. For each type... By maximizing the average treatment effect To determine the optimal intervention plan It is determined by the following formula: , in, For the first Recommended optimal intervention for each patient type To maximize the value of the following expression The value of , Index of possible intervention options (e.g., options A, B, C). Intervention Program For type The estimated average treatment effect for patients was derived from models such as causal forests. Positive values ​​indicate benefit, and the larger the value, the greater the benefit.

[0049] Then, in order to quantify the different clinical indicators for each patient type ( The importance of risk prediction is determined by configuring feature weights. Specifically, for each type... Training a gradient boosting tree model to predict nutritional risk outcomes And calculate the SHAP value for each feature to quantify its contribution to risk prediction for this type of patient: , in, For type Prediction model and patients ,feature The SHAP value represents the contribution of this feature to the prediction result. To calculate the feature index of contribution, For the first Predictive models trained for each patient type (such as gradient boosting trees). For a specific patient sample, For feature set The middle does not contain features A subset, The total number of features, To use feature sets Add features The predicted value of the time model, To use only the feature set The predicted values ​​of the model are then used. Subsequently, the SHAP values ​​are regularized using Lasso regression to obtain a sparse, interpretable feature weight vector. .

[0050] Furthermore, through risk assessment threshold optimization, for each patient type ( Establish personalized risk level classification standards. Specifically, for each type... Based on the predicted risk scores and actual outcomes of its patient population, the optimal risk threshold was determined by maximizing the Youden index. , , , in, The Youden index is a comprehensive indicator for evaluating the efficacy of diagnostic tests, with a value ranging from -1 to 1. A higher value indicates stronger diagnostic ability. Candidate thresholds for classifying high-risk / medium-risk / low-risk To be at the threshold Sensitivity (true positive rate) under [the specified conditions]. To be at the threshold Specificity (true negative rate) under [condition]. For the first The personalized optimal risk threshold determined for each patient type, i.e., the Youden index. The biggest one This value allows for personalized risk level classification.

[0051] Finally, dynamic scoring rules are generated to capture and formalize the complex judgment logic in clinical practice. Specifically, a Bayesian rule list approach is used to automatically learn concise and interpretable clinical rules from the data. These rules are represented in "IF-THEN" form (e.g., IF...). AND (THEN risk score +2), and store it in the assessment parameter package as a dynamic scoring rule template.

[0052] Finally, a complete assessment parameter package is constructed for each typical patient type, including: a. Feature weight configuration: The weight vector obtained by joint optimization of SHAP values ​​and Lasso regression. .

[0053] b. Data collection template: Define a list of data items that must be collected and are recommended to be collected for assessing this type of patient, based on feature weights and clinical importance.

[0054] c. Risk assessment threshold: A personalized threshold determined based on the principle of maximizing the Youden index. .

[0055] d. Recommended nutritional intervention plan: The plan with the best average treatment effect for this type of patient, as assessed by the causal forest model. .

[0056] e. Dynamic scoring rule template: A set of interpretable temporary scoring adjustment rules learned from a Bayesian rule list. .

[0057] Specifically, in the model integration and validation process, the clustering model, feature prototypes of each type, and all evaluation parameter packages obtained from the association analysis, generated in the above steps, are integrated and encapsulated into a unified, callable software object, namely, the large-scale patient subtyping model. This model provides a standardized prediction interface: inputting supplementary clinical data of a new patient, after the same preprocessing and dimensionality reduction, the model can output its patient type, matching probability, and automatically retrieve the corresponding evaluation parameter package.

[0058] Example: Constructing an assessment parameter package for the "frail" type. Suppose that a "frail elderly" patient group was obtained through clustering. Supervised association analysis revealed: Feature weighting: The weights were determined by combining SHAP value analysis with Lasso regression as follows: age 0.28, grip strength 0.31, activities of daily living (ADL) score 0.25, and BMI 0.16.

[0059] Data collection template: Define required data collection items: age, grip strength test, ADL score, mini nutrition assessment (MNA) questionnaire; suggested data collection items: gait test, body composition analysis.

[0060] Risk assessment threshold: Based on the outcome data of this group, the Youden Index determined its high-risk threshold to be 4.2 points (the general threshold is 5 points).

[0061] Recommended intervention: Causal forest model analysis showed that the "oral nutritional supplementation combined with protein supplementation and resistance training" program had the most significant net effect on improving sarcopenia and reducing readmission rate in this type of patient.

[0062] Dynamic scoring rule: The Bayesian rule is learned to determine the rule: IF ADL score < 60 THEN Risk level is increased by one level.

[0063] It is worth noting that in step S100, firstly, an unsupervised clustering algorithm automatically identifies and defines typical patient groups with different pathophysiological states within the data based on the similarity of patients' multidimensional clinical characteristics, without any pre-defined labels. Then, through supervised association analysis, each data-driven patient type is deeply correlated with its corresponding nutritional intervention outcome data, thereby extracting and refining the assessment focus and effective intervention experience for that type of patient. Finally, this condensed knowledge is systematically encapsulated into an assessment parameter package uniquely bound to each type, constructing a structured and computable large-scale patient classification model that integrates clinical experience with data patterns.

[0064] As in step S201, specifically in this embodiment, basic clinical data of the target patient is obtained through automatic extraction by the electronic medical record system (EMR), manual entry by medical staff (e.g., via mobile APP or web form), and medical device interfaces (e.g., weighing scale, height measuring device). For example, the first screening model is a general nutritional assessment rule engine, which embeds scoring rules from standardized screening tools such as NRS2002 or MUST. By inputting basic clinical data, the rule engine matches and calculates according to preset rules, resulting in a standardized nutritional risk score (e.g., NRS2002 total score) and the corresponding preliminary risk level (e.g., high risk, low risk).

[0065] For example: If a patient has experienced a 5% weight loss in the past month (corresponding to 1 point) and has had difficulty breathing and decreased appetite in the past week, with food intake approximately 60% of normal needs (corresponding to 1 point), then the nutritional status impairment score is 2 points. The system matches the patient's primary diagnosis (e.g., "acute exacerbation of chronic obstructive pulmonary disease") and current condition to the NRS2002 disease severity classification, providing a score. For example, an acute exacerbation of COPD with respiratory failure might be classified as "severe disease," with a score of 3 points. If the patient is ≥70 years old, add 1 point. Adding these three parts (nutritional status 2 points + disease severity 3 points + age 1 point) yields a total NRS2002 score of 6 points (range 0-7). Risk levels are determined based on the total score: an NRS2002 total score ≥5 points indicates high risk, suggesting a high nutritional risk requiring immediate individualized and intensive nutritional intervention and close monitoring.

[0066] As in step S202, specifically in this embodiment, the basic clinical data of the target patient will be... Compared with the initial risk level The data is input into a pre-built large-scale patient classification model, and an intelligent matching algorithm determines the most typical patient type that best matches the patient. It automatically retrieves assessment parameter packages that are bound to the corresponding typical patient types for subsequent personalized assessment.

[0067] First, based on the basic clinical data of the target patients already obtained in S201. Compared with the initial risk level Basic clinical data typically includes structured information such as age, body mass index, and disease diagnosis. The system performs standardization preprocessing on this raw data to eliminate dimensional differences and ensure consistency with the data scale used during model training. For example, continuous numerical features undergo the following standardization process: , in, The first in historical data All possible values ​​of a feature .

[0068] At the same time, the initial risk levels (such as "low risk", "medium risk", "high risk") were also converted into a unified numerical form. Subsequently, the system will generate standardized basic clinical data vectors. Together with the encoded risk level value, they form a complete feature representation vector that comprehensively represents the current patient status. This serves as the input for subsequent fractal matching: .

[0069] The "large-scale patient subtyping model" maintained internally by the system includes A typical patient type obtained through learning from historical data. (For example, "high inflammatory response type", "sarcopenia-dominant type", etc.). Each type $C_k$ is associated with a feature prototype vector. (This can be viewed as the average feature profile of this type of patient) and a set of feature weight configurations. (Which features are most critical for distinguishing this type?)

[0070] The system calculates the current patient feature vector. With each typical patient type prototype vector similarity between The similarity calculation here is not a simple geometric distance, but uses a weighted cosine similarity algorithm: , in, For type The Middle The weights of each feature are derived from the feature weight configuration in the evaluation parameter package for this type. The dimension of the feature vector. This means that when calculating similarity, weights specific to each type will be configured. The design differentiates the importance of different clinical features. For example, for "inflammatory" patients, inflammatory markers (such as C-reactive protein) have a higher weight in similarity calculations; while for "debilitated" patients, muscle function indicators (such as grip strength) are more crucial. This design ensures that the matching process focuses on the clinical features that are most discriminatory to a specific type.

[0071] The similarity scores between the patient and all types were calculated. Then, the system will determine the similarity threshold based on the preset threshold. A screening process is performed. Only types with a similarity level that meets or exceeds the threshold will be included in the candidate set. .

[0072] If the candidate set contains only one type, that is If so, then directly identify it as the target patient type. As in step S203, this embodiment specifically includes personalized configuration of the second screening model, intelligent identification of data gaps and triggering supplementary data collection.

[0073] If multiple candidate types exist, the system will select the type with the highest similarity as the target, that is: .

[0074] In some advanced settings, the historical confidence level of the type may also be taken into account. Weighted decision-making can be implemented to improve the stability of the results.

[0075] If the similarity scores for all types fail to reach the threshold, it indicates that the current patient characteristics differ significantly from existing known types. In this case, the system will initiate a degradation strategy, such as appropriately relaxing the threshold. Rematch, or mark the patient as "unclassified" and use a generic, risk-conservative default assessment parameter package for subsequent processing, while recording the case for future model updates.

[0076] As in step S203, specifically in this embodiment, based on the evaluation parameter package obtained in S202... The pre-deployed second screening model is personalized, and clinical data is intelligently collected and supplemented to calculate a personalized score. And determine the final risk level. .

[0077] First, the second screening model is customized. The system loads the evaluation parameter package. Then, first configure the vector according to the feature weights. The pre-deployed second screening model is dynamically configured. This weight vector... Defined for target patient types The importance of each clinical feature is given, where $d$ is the total number of features.

[0078] The configuration process is implemented through the following mapping function: , in, This is the second screening model. The configured personalized model includes: (1) updating the weight parameters in the weighted scoring function; (2) adjusting the feature selection threshold; and (3) if the patient classification includes a risk prediction sub-model identifier. If so, then the corresponding dedicated sub-model will be loaded.

[0079] Then, intelligent data supplementation based on the data acquisition template. The system uses the data acquisition template from the patient classification assessment. Identify data gaps for current patients. (Template) The set of data items that must be collected is defined. and the recommended set of data items to be collected The data gap identification algorithm is as follows: , in, For the patient's existing clinical data items, From Based on feature weights The selected high-priority subset.

[0080] For the identified data gaps The system automatically generates data acquisition requests and distributes them to the corresponding clinical workstations. The acquired supplementary clinical data is denoted as a vector. Its dimensions are .

[0081] Furthermore, personalized scores are calculated. The system integrates the patient's basic clinical data. and supplementary clinical data To form a complete feature vector First of all, Perform weighted processing: , in, This represents element-wise multiplication.

[0082] Personalized rating Through the configured model calculate: .

[0083] If the evaluation parameter package contains a dynamic scoring rule template The system will dynamically adjust the base score using rules. Each rule... In the form of: , in, For logical conditional functions, The adjusted value. The final score after applying all matching rules is:

[0084] .

[0085] Ultimately, the risk level of the target patient is determined. The system uses risk assessment thresholds from the assessment parameter package. Map personalized scores to final risk levels For example, logical conditional functions: .

[0086] threshold Multiple threshold values ​​are typically defined, such as ,satisfy Final risk level The rules for determining it are as follows: , in, , , The risk thresholds, individually set for this patient type, are optimized using historical data. For example, for patients with a "high inflammatory response," a threshold might be set... , , For patients with "sarcopenia-dominant" type, a treatment plan may be set... , , The specific values ​​are defined by the evaluation parameter package.

[0087] To enhance clinical transparency and credibility, the system will simultaneously generate a brief report on the matching criteria. This report will clearly state: Final matched patient types and its matching degree A few sets of clinical features that play a key role in matching decisions Deterministic assessment of matching results .

[0088] Based on the aforementioned steps S201 to S203, the system completes a comprehensive hierarchical nutritional risk assessment process. This process begins with S201, a rapid initial screening based on standardized rules (such as NRS2002), aimed at efficiently identifying at-risk individuals and determining the necessity of subsequent assessments. Then, in S202, the system uses a constructed large-scale patient typing model to intelligently categorize patients who tested positive in the initial screening, matching them to the most representative clinical subgroup and obtaining a customized assessment parameter package for that subgroup. Subsequently, in S203, the system dynamically configures the assessment model based on this patient typing, intelligently drives the collection of supplementary data, and integrates personalized weights and thresholds to calculate the final risk score and level. Thus, the system successfully achieves a progressive analysis from general screening to precise typing, and then to individualized assessment, laying a solid foundation for generating a comprehensive report that integrates standardized results with in-depth personalized insights.

[0089] As in step S300, specifically in this embodiment, the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level are integrated to generate a comprehensive nutritional risk assessment report.

[0090] The system first performs multi-source data aggregation and alignment. Using a unified patient identifier and timestamp, it synchronously collects and integrates all data from the three core processing modules for this assessment cycle, including: standardized screening results from step S201, such as the name of the screening tool used, various sub-scores and their basis, and the standardized total score (…). ) and preliminary risk level ( ); patient classification results from step S202, such as the target patient type ( Matching similarity () ) and key matching features; personalized assessment results from step S203, such as personalized scores ( ), final risk level ( ), application feature weight configuration ( ), supplementary data collection list, and triggered dynamic rules ( ) and personalized thresholds ( Simultaneously, recommended intervention plans for the corresponding target types are extracted from the evaluation parameter package. The system performs automatic consistency checks to ensure data integrity and source consistency.

[0091] Subsequently, a structured report is generated. The report generation module, based on a preset template, transforms the aggregated data into a logically clear and hierarchically structured narrative report. The report begins with a summary of key findings, summarizing the core conclusions and highest-priority recommendations in one or two sentences. It then elaborates on the standardized screening results, clearly listing the scoring criteria and calculation basis for each item. The core of the report is an in-depth personalized assessment analysis, systematically explaining the assessment logic: explaining the basis for patient classification (such as matching features and similarity); explaining the feature weighting configuration used (such as the clinical reasons for increasing the weight of inflammatory markers); listing the data items intelligently supplemented for accurate assessment and their clinical significance; demonstrating the triggered dynamic scoring rules and their specific adjustments to the scores; and comparing the differences between the personalized thresholds used for the final risk level determination and the general standards, thus providing a complete explanatory chain across multiple levels, including classification reasoning, weighting configuration, data supplementation, rule application, and threshold personalization.

[0092] Finally, the system performs report formatting and multimodal output. The generated structured content is automatically converted into multiple concurrent formats to adapt to different clinical scenarios: generating structured data that can be directly integrated into electronic medical records (EMR) (such as FHIR format); generating detailed graphic reports for clinicians to review (such as PDF); and simultaneously extracting concise action lists for bedside nurses to execute. Before report release, the system performs automated quality control and flagging, checking logical consistency and marking anomalies to ensure the accuracy and reliability of the report.

[0093] Through the above process, step S300 integrates the standardized, categorized, and personalized results generated in the preceding steps (S201, S202, S203) into a comprehensive assessment report that is authoritative, accurate, interpretable, and operable, providing a complete chain of evidence for clinical decision-making, from risk identification to intervention recommendations.

[0094] As in step S400, specifically in this embodiment, it is the model iterative update process. This step aims to build a closed-loop learning mechanism that drives continuous model optimization based on real-world application feedback while strictly protecting data privacy, enabling the entire system to continuously evolve with clinical practice, thereby achieving the goal of "getting smarter with use".

[0095] The entire iterative process begins with the completion of a single assessment and application cycle. For example, upon patient discharge or the end of the current nutritional intervention cycle, the system automatically initiates the feedback process. It encapsulates the key data generated in steps S202 and S203 for this case—including supplementary clinical data for subtyping and assessment, the comprehensive nutritional risk assessment report generated in step S300 (containing system judgment logic and recommended plans), and the actual implemented nutritional intervention plan and final clinical outcome data (such as complications, length of hospital stay, and changes in nutritional indicators)—into a desensitized and encrypted "feedback training sample package," which is securely stored locally.

[0096] After entering a regular (e.g., weekly) model update cycle, local nodes in each medical institution use the newly added feedback sample packages to incrementally train the locally maintained large-scale patient subtyping model. The training focuses on two aspects: first, at the unsupervised clustering level, using the feature data of new patients to optimize the definition of existing patient type clusters, and even discover new potential subtypes; second, at the supervised association analysis level, the core is to verify and optimize the assessment parameter package associated with each patient type based on real outcome data. For example, if the system finds that for "high inflammatory response" patients, the efficacy of treatment plan B in actual clinical practice is consistently better than the originally recommended treatment plan A for the patient subtype, it will automatically adjust the priority of the intervention plan in that patient type subtype. Simultaneously, the system will also fine-tune the feature weight configuration and risk assessment threshold for that type, making the assessment strategy more accurately aligned with actual clinical outcomes.

[0097] To achieve cross-institutional knowledge sharing and collaborative evolution while ensuring data privacy, this embodiment employs a federated learning framework. Each local node only encrypts and uploads the updated model parameters (not any original data) to the central coordination node. The central node securely aggregates the encrypted updates uploaded by each node, generates a global model update, and then distributes it back to all participating nodes. Each node applies this global update to complete the synchronous upgrade of its local model. Through this mechanism, the system can integrate clinical experience from multiple centers, continuously improve the accuracy of subtyping and the effectiveness of intervention recommendations, while ensuring that the patient's original data does not leave the local machine, fundamentally meeting the regulatory requirements for data security and privacy protection.

[0098] Through step S400, this invention constructs a complete closed loop from historical data learning (S100) to clinical assessment application (S201-S300), and then drives continuous model iteration through practice feedback (S400), realizing the dynamic co-evolution of the assessment system and the clinical practice effect.

[0099] like Figure 6 As shown in the embodiments of the present invention, a hybrid model-driven nutritional risk prediction system for precision nutritional intervention is also provided to implement the aforementioned method. This system specifically includes a large-scale model construction module, a standardized initial screening module, a patient classification and parameter matching module, a personalized precision assessment module, and a comprehensive report generation module.

[0100] Specifically, the large-scale model construction module is used to construct a large-scale patient classification model based on the patient group's historical clinical characteristic data, nutritional risk scores, and nutritional intervention outcome data, through unsupervised clustering algorithms and supervised association analysis. This large-scale model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package includes at least feature weight configurations, data collection templates, and risk assessment thresholds for the corresponding typical patient type. The standardized initial screening module is used to acquire the target patient's basic clinical data and, using a pre-deployed first screening model based on preset standardized rules, obtain a standardized nutritional risk score and its corresponding preliminary risk level. The patient classification and parameter matching module is used to input the target patient's basic clinical data and preliminary risk level into the large-scale patient classification model to determine the target patient type; and to retrieve the assessment parameter package associated with the target patient type from the large-scale model. The personalized assessment module is used to personalize the configuration of the pre-deployed second screening model based on the feature weights configured in the assessment parameter package. The second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package. Based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment thresholds in the assessment parameter package, the second screening model calculates and outputs a personalized nutritional risk score and the corresponding final risk level. The comprehensive report generation module integrates the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

[0101] Scenario Example 1

[0102] Scenario: In the surgical oncology department of a hospital, three newly admitted postoperative gastric cancer patients (Wang, Li, and Zhao) all underwent standard nutritional risk screening (NRS2002) and received the same score of 5, thus all were classified as "high-risk." Based solely on this standardized result, clinicians can only obtain a general indication that "nutritional risk exists and intervention is necessary," making it difficult to make differentiated decisions.

[0103] After applying the system of this invention, in-depth personalized evaluation is achieved through the process of steps S201 to S300: Step S201: Based on the basic clinical data of the three patients, the system calculates their respective NRS2002 scores. Due to factors such as age, weight loss, and surgical trauma, all three patients received a score of 5, with an initial risk level of "high risk." This result automatically triggers the system's in-depth assessment process.

[0104] Step S202: The system further extracts and analyzes richer clinical characteristics of each patient and matches them with the pre-built "patient subtyping model".

[0105] For patient Wang, the extended characteristics showed significantly abnormal inflammatory markers (CRP 180 mg / L, PCT 8.0 ng / mL), while muscle mass was relatively normal. System calculations showed a high degree of agreement (92%) with the "high inflammatory response / sepsis" archetype in the model.

[0106] For patient Li, inflammatory markers were only slightly elevated (CRP 22 mg / L), but imaging and functional assessments indicated severe muscle loss and a significant decrease in grip strength. Systematic assessment determined that his characteristic combination most closely matched the "sarcopenia / fainting-dominant type," with a match rate of 88%.

[0107] For patient Zhao, all extended characteristics showed mild abnormalities, lacking any particularly prominent contradictory indicator. The system calculated that his match with "mild stress / metabolic type" was the highest.

[0108] Based on the matching results above, the system identified different target patient types for the three patients and loaded the corresponding assessment parameter packages.

[0109] Step S203: The system performs personalized configuration and calculation based on the specific assessment parameter package loaded for each patient.

[0110] For patient Wang (inflammatory type), the system applied a configuration that significantly increased the weight of "inflammatory indicators." It automatically suggested and imported targeted test results such as blood cultures. The extremely high inflammation value, after weighting, resulted in a personalized score of 6.8. Referring to the strict threshold for this type (≥5.0 points is considered extremely high risk), the final risk level was determined to be "extremely high risk." The system's output clearly indicated that infection and a hypermetabolic state were the core risks.

[0111] For patient Li (sarcopenia type), the system was configured according to the patient's classification, focusing on "muscle loss" and "functional status," and automatically generated more detailed body composition and grip strength tests. Severe muscle loss was calculated with high weight, resulting in a personalized score of 5.5, and the final risk level remained "high risk," but the core assessment pointed to chronic wasting and functional decline.

[0112] For patient Zhao (mild stress type), the system adopted a more balanced weighting configuration in the patient classification for this type and applied a relatively high risk threshold (e.g., 4.5 points). The calculated personalized score was 3.8 points. Because this did not reach the high-risk threshold for this type, the final risk level was adjusted to "low risk." The system concluded that the main risk stemmed from short-term insufficient intake.

[0113] Step S300: The system generated clearly structured comprehensive nutritional risk assessment reports for the three patients. Each report not only listed the standardized screening results, but also detailed the complete personalized decision-making logic chain from patient classification, weight adjustment, application of supplementary data collection, to threshold comparison, making the basis for changes in risk level readily apparent.

[0114] Understandably, according to embodiments of the present invention, when faced with three patients with identical standardized screening results, the system reveals the essential differences in their underlying pathophysiological states through a large-scale patient subtyping model—corresponding to acute infection, chronic wasting, and short-term stress, respectively. This not only achieves precise differentiation of risk levels (extremely high risk, high risk, low risk), but more importantly, the system uses subtyping intelligence to drive differentiated in-depth examinations (such as targeted inflammation or muscle function assessments), avoiding indiscriminate comprehensive examinations. Ultimately, the interpretable report output by the system, along with the differentiated risk levels, directly correlates with and recommends distinctly different intervention focuses (such as anti-inflammatory nutrition, rehabilitation combined with nutritional support, and basic nutritional supplementation), thereby achieving a leap from homogeneous risk alerts to personalized clinical decision support.

[0115] Scenario Example 2

[0116] Scenario: Mr. Zhou, 68 years old, was admitted to the hospital due to "fatigue and poor appetite for 1 month, which worsened and was accompanied by shortness of breath for 3 days". His past medical history was extensive: 10-year history of coronary heart disease (post-coronary stent implantation), 15-year history of type 2 diabetes (generally controlled with oral hypoglycemic agents), and stage 3 chronic kidney disease (CKD).

[0117] Step S201: Based on the basic clinical data collected upon admission, an automatic NRS2002 screening was performed: age 68 years (+1 point), BMI 21.0 (0 points, but close to the lower limit), weight loss of approximately 6% in the past month (+1 point), reported poor appetite with food intake about half of normal (+1 point), disease severity (considering multiple chronic diseases, but no obvious signs of acute exacerbation upon admission, the preliminary assessment is "mild disease", +1 point). Standardized total score = 4 points, preliminary risk level is "medium risk". This conclusion may lead the clinical team to focus on simple nutritional supplementation.

[0118] Step S202: Integrate various baseline data collected upon patient admission that far exceed the standard screening scope, including: age 68 years, coexisting coronary artery disease, diabetes, chronic kidney disease, etc., resting heart rate 98 bpm, blood pressure 100 / 65 mmHg, hemoglobin 98 g / L, albumin 32 g / L, serum creatinine 150 μmol / L, estimated glomerular filtration rate (eGFR) 45 mL / min / 1.73 m², glycated hemoglobin (HbA1c) 8.5%, electrocardiogram and echocardiography indicating old anterior wall myocardial infarction, and left ventricular ejection fraction (LVEF) 45%. These multidimensional features are input into the patient classification model for matching and calculation. Based on patterns learned from historical data, the model identified that the aforementioned combination of features (advanced age, heart failure, renal failure, anemia, poor glycemic control, and recent weight loss) was highly similar to the feature prototype of the patient type "chronic multiple organ dysfunction with pre-cardiac cachexia" in the model, with a matching degree of 85%. Based on this, this type was determined as the target patient type, and an associated assessment parameter package was loaded. This patient subtype defines the core assessment dimensions and risk calculation rules for this type of patient.

[0119] Step S203: Based on the data acquisition template defined in the loaded assessment parameter package, automatically identify the current data gap and generate structured supplementary assessment suggestions, which include: high-sensitivity inflammatory markers (such as IL-6), prealbumin, body composition analysis (focusing on assessing skeletal muscle mass), grip strength test, and cardiopulmonary function or endurance test (such as the 6-minute walk test).

[0120] Following supplemental assessment, new data was obtained: IL-6 was 25 pg / mL, body composition analysis confirmed muscle loss, grip strength was below the normal reference range, and the 6-minute walking distance was 280 meters. Based on the feature weighting configuration set for this type in the assessment parameter package (with "inflammatory markers," "muscle loss," and "functional status" assigned higher weights), all features (including baseline and supplemental clinical data) were weighted and calculated. The results showed that the intermediate-risk score from the original standardized screening (NRS2002) contributed relatively little weight in this weighted calculation. The supplementally obtained features such as "elevated inflammatory markers," "muscle loss," and "decreased functional reserve," combined with the cardiac and renal function indicators from the baseline data, resulted in a weighted personalized score of 8.2.

[0121] The personalized score (8.2 points) is compared with the preset risk assessment threshold in the assessment parameter package for this type (for example, for this type, set...). The scores were compared. Because the score (8.2 points) significantly exceeded the high-risk threshold (7.0 points), the final risk level was determined to be "extremely high risk." The output conclusion included a warning message, indicating that the patient's condition conformed to the characteristics of chronic multiple organ dysfunction progressing to cachexia, suggesting that routine nutritional support strategies may not be applicable, and recommending the initiation of multidisciplinary collaborative management.

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented in software, the above embodiments can be stored, in whole or in part, in a readable storage medium, such as a memory. The memory can be a read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; the embodiments of the present invention do not specifically limit this.

[0124] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different systems to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

Claims

1. A hybrid model-driven nutritional risk prediction method for precision nutritional intervention, characterized in that, The method includes the following steps: S100. Based on the patient population's historical clinical characteristics data, nutritional risk scores, and nutritional intervention outcome data, a large-scale patient classification model is constructed using unsupervised clustering algorithms and supervised association analysis. The large-scale patient classification model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package at least includes feature weight configurations, data collection templates, and risk assessment thresholds for the corresponding typical patient type. S201. Obtain basic clinical data of the target patients and use a pre-deployed first screening model based on preset standardized rules to obtain a standardized nutritional risk score and the corresponding preliminary risk level. S202. Input the basic clinical data and preliminary risk level of the target patient into the patient classification model to determine the target patient type; obtain the assessment parameter package associated with the target patient type from the patient classification model; S203. Based on the feature weight configuration in the assessment parameter package, a pre-deployed second screening model is configured in a personalized manner; the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package; the second screening model calculates and outputs a personalized nutritional risk score and the corresponding final risk level based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment threshold in the assessment parameter package. S300. Integrate the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

2. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, In step S100, the supervised association analysis specifically includes: analyzing the nutritional intervention outcomes of historical patients associated with each typical patient type, configuring one or more recommended nutritional intervention programs for each typical patient type, and incorporating the recommended nutritional intervention programs as components of the assessment parameter package.

3. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, In step S202, determining the target patient type to which the target patient belongs specifically includes: calculating the similarity between the basic clinical data and preliminary risk level and the feature representation of each typical patient type, and selecting one or more types with the highest similarity as the target patient type.

4. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, In step S203, the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package, specifically including: Compare the existing clinical data of the target patient with the data items defined in the data collection template to identify data item gaps; Based on the data item gaps, a data acquisition request is generated and sent.

5. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, The preset standardized rules used by the first screening model are the scoring rules in NRS2002 or the MUST Nutrition Risk Screening Tool.

6. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, In step S203, calculating a personalized nutritional risk score based on the weighted clinical characteristic data and the supplementary clinical data specifically includes: A trained risk prediction sub-model is invoked for processing; the structure or hyperparameters of the risk prediction sub-model are defined by the evaluation parameter package.

7. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, The method further includes step S400: feeding the comprehensive nutritional risk assessment report generated in step S300, as well as the actual nutritional intervention plan and final outcome data of the target patient, into the historical data of the patient group as new training samples for periodically updating the large patient subtyping model.

8. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 7, characterized in that, The periodic updating of the patient subtyping model is performed using a federated learning framework, specifically including: At each local node, the locally maintained large-scale patient subtyping model is incrementally trained using the newly added training samples. Upload the updated parameters of the trained model to the central node; At the central node, model update parameters from multiple local nodes are securely aggregated to generate a global model update. The global model update is distributed to each local node.

9. The hybrid model-driven nutritional risk prediction method for precision nutritional intervention as described in claim 1, characterized in that, The assessment parameter package also includes a dynamic scoring rule template, which is used in step S203 to dynamically generate or adjust temporary scoring rules for the target patient based on the collected supplementary clinical data; the second screening model, based on the dynamic scoring rule template, performs secondary processing and scoring fusion on the weighted clinical feature data and the supplementary clinical data to calculate the personalized nutritional risk score.

10. A hybrid model-driven nutrition risk prediction system for precision nutrition intervention, characterized in that, The system includes: The large model building module is used to construct a large patient classification model based on the patient population's historical clinical characteristic data, nutritional risk scores, and nutritional intervention outcome data, through unsupervised clustering algorithms and supervised association analysis. The large patient classification model contains multiple typical patient types, and each typical patient type is associated with an assessment parameter package. The assessment parameter package includes at least the feature weight configuration, data collection template, and risk assessment threshold for the corresponding typical patient type. The standardized initial screening module is used to acquire basic clinical data of target patients and use a pre-deployed first screening model based on preset standardized rules to obtain a standardized nutritional risk score and corresponding preliminary risk level. The patient subtyping and parameter matching module is used to input the basic clinical data and preliminary risk level of the target patient into the patient subtyping model to determine the target patient type to which the target patient belongs; and to obtain the assessment parameter package associated with the target patient type from the patient subtyping model. The personalized assessment module is used to configure a pre-deployed second screening model based on the feature weight configuration in the assessment parameter package; the second screening model collects supplementary clinical data of the target patient according to the data collection template in the assessment parameter package; the second screening model calculates and outputs a personalized nutritional risk score and the corresponding final risk level based on the weighted clinical feature data and the supplementary clinical data, and according to the risk assessment threshold in the assessment parameter package. The comprehensive report generation module is used to integrate the preliminary risk level, the final risk level, and the assessment basis corresponding to the final risk level to generate a comprehensive nutritional risk assessment report.

Images