Sarcopenia risk assessment method based on knowledge graph

Abstract

The application discloses a sarcopenia risk assessment method based on a knowledge graph, relates to the field of medical large models, collects full-dimension multi-source data of a patient and carries out pretreatment; a six-tuple entity relationship model is defined, a dynamic entity life cycle and a heterogeneous relationship type system are introduced, multi-modal feature fusion is realized through modal adaptive weight learning, knowledge conflict resolution and entity evolution embedding; based on an enhanced spatiotemporal graph convolution network, through heterogeneous relationship attention space convolution, spatiotemporal adaptive time convolution and triple contrast learning tasks, enhanced entity embedding is output, and the enhanced entity embedding comprises a historical state, a current state, an evolution rate and a future trend. Through multi-modal spatiotemporal alignment and enhanced knowledge graph construction, the application realizes deep fusion of multi-source heterogeneous data and spatiotemporal dynamic relationship modeling, and significantly improves the comprehensiveness and precision of sarcopenia risk assessment.

Description

Technical Field

[0001] This invention relates to the field of medical large-scale model technology, and in particular to a knowledge graph-based method for sarcopenia risk assessment. Background Technology

[0002] Sarcopenia, a syndrome characterized by decreased muscle mass and declining muscle function, has seen its incidence rise steadily with population aging, becoming a significant public health issue affecting the health and quality of life of the elderly. Currently, risk assessment and intervention for sarcopenia still face numerous technical bottlenecks:

[0003] At the data fusion and modeling level, traditional assessment methods often rely on single-modal data and lack systematic integration of multi-source heterogeneous data. These data have large differences in sampling frequency in the time dimension and heterogeneity in the spatial dimension due to multi-center collection, making it difficult to effectively explore the spatiotemporal correlations between data and failing to comprehensively depict the dynamic development process of sarcopenia.

[0004] At the level of knowledge representation and reasoning, the application of existing knowledge graphs in the medical field is mostly limited to static triples, failing to fully consider the spatiotemporal dynamics of entity relationships. Furthermore, the semantic gap in multimodal data has not been effectively bridged, resulting in insufficient modeling capabilities of knowledge graphs for the complex pathological mechanisms of sarcopenia, ambiguous causal reasoning of risk transmission paths, and difficulty in identifying key intervention nodes.

[0005] Regarding model interpretability and personalized intervention, mainstream risk assessment models, such as deep learning classifiers, are mostly black-box models. While they possess high predictive accuracy, they cannot explain the pathophysiological mechanisms of risk formation and are difficult to generate personalized intervention plans based on individual patient characteristics. In clinical practice, doctors need to subjectively judge the direction of intervention based on experience, leading to inconsistent intervention effects and difficulty in ensuring patient compliance.

[0006] Regarding privacy protection and multi-center collaboration, sarcopenia research involves multi-center data sharing, but patient privacy regulations (such as HIPAA) strictly limit the cross-institutional transfer of raw data. Existing federated learning methods in medical scenarios suffer from low communication efficiency and poor model heterogeneity adaptation, making it difficult to achieve collaborative learning on large-scale multi-center data while protecting privacy. This results in limited model generalization ability and an inability to quickly absorb new medical evidence and clinical data for iterative updates.

[0007] In summary, existing technologies have significant shortcomings in multimodal data fusion, spatiotemporal knowledge modeling, interpretable intervention, and privacy-preserving collaborative learning for sarcopenia risk assessment. There is an urgent need for a systematic approach that can integrate multi-source heterogeneous data, dynamically model spatiotemporal knowledge, provide interpretable intervention plans, and protect privacy, in order to improve the accuracy and clinical applicability of sarcopenia risk assessment. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a knowledge graph-based method for sarcopenia risk assessment. The technical solution adopted is as follows:

[0009] A knowledge graph-based method for sarcopenia risk assessment includes the following steps:

[0010] Step 1: Collect multi-dimensional, multi-source data from patients and perform preprocessing;

[0011] Step 2: Define a six-tuple entity relationship model, introduce a dynamic entity lifecycle and heterogeneous relationship type system, and achieve multimodal feature fusion through modal adaptive weight learning, knowledge conflict resolution and entity evolution embedding;

[0012] Step 3: Based on the enhanced spatiotemporal graph convolutional network, through heterogeneous relational attention spatial convolution, spatiotemporal adaptive temporal convolution and triple contrastive learning task, the enhanced entity embedding is output. The enhanced entity embedding includes historical state, current state, evolution rate and future trend.

[0013] Step 4: Construct a temporal causal graph network. Use a temporal causal enhanced path mining algorithm to mine risky paths through bidirectional biased random walks. After multi-granularity causal verification, output an interventionable risk transmission map.

[0014] Step 5: Output sarcopenia risk level assessment based on the sarcopenia risk level assessment model. The sarcopenia risk level assessment model adopts a hierarchical fusion architecture with dynamic weight adjustment, and combines modal feature dual attention, path attention and adaptive graph readout function. The sarcopenia risk level assessment result is achieved through multi-task learning of inter-task attention mechanism.

[0015] Optionally, step 6 is also included: a dynamic early warning mechanism based on personalized thresholds, combined with a patient digital twin model to simulate the intervention effect, and generate a personalized intervention plan that meets the constraints and is suitable for compliance.

[0016] The dynamic early warning mechanism sets personalized early warning thresholds based on age, comorbidities, and basic health status, and dynamically adjusts the early warning level in conjunction with the risk evolution rate; the digital twin model maps the patient's physiological state in real time and simulates the implementation effect of different intervention programs.

[0017] Optionally, step 7 is also included, which employs a federated learning framework of federated knowledge distillation and cross-domain knowledge transfer, combined with adaptive differential privacy, encrypted transmission, and privacy leakage detection to achieve privacy protection and model updates.

[0018] Optionally, through dynamic visualization, interactive risk simulation, and multi-dimensional interpretation, a structured report aligned with clinical guidelines can be output, thus building a clinical decision feedback interface.

[0019] Optionally, in step 1, the patient's multi-dimensional and multi-source data includes clinical data, imaging data, time-series functional data, biochemical indicators, nutritional and behavioral data, omics data, environmental data, and subjective feeling data. The patient's multi-dimensional and multi-source data is preprocessed through proactive data completion, cross-scale data fusion, and adaptive spatiotemporal alignment.

[0020] Optionally, the six-tuple entity relationship model described in step 2 consists of a head entity, a relationship type, a tail entity, a timestamp, a spatial context, and a confidence level; the heterogeneous relationship type system includes causal relationships, association relationships, temporal dependencies, and intervention-response relationships.

[0021] Optionally, in step 4, the temporal causal reinforcement path mining algorithm includes the following steps:

[0022] Step 41: Construct a temporal causal graph network to integrate temporal knowledge graphs and causal reasoning rules, and model dynamic causal relationships;

[0023] Step 42: Forwardly explore the risk transmission path and backward trace the risk source node to achieve a two-way biased random walk;

[0024] Step 43: Prioritize risky paths by combining three-dimensional scores of path strength, causal confidence, and intervention feasibility.

[0025] Optionally, the output method for the sarcopenia risk level assessment results in step 5 is as follows:

[0026] Step 51: Input features include enhanced entity embedding, risk transmission path embedding, and patient global graph-level representation; construct input feature matrix X.

[0027] Step 52: Obtain the initial feature representation of each task through the task-specific feature mapping layer. Based on the inter-task correlation matrix C and the average task loss The dynamic weights for each task are calculated using the attention weight formula. ;

[0028] Step 53, using weights By integrating the characteristics of each task, The input to the main task output layer generates a risk level probability distribution through the Softmax function. ;

[0029] Step 54: Determine the risk level based on the clinical calibration threshold and calibrate the confidence level using a temperature scaling algorithm;

[0030] Step 55: The output includes risk level labels, corresponding probability values ​​and confidence levels, and is associated with key supporting features, including the trend of decreasing grip strength, the proportion of muscle fat infiltration and insufficient nutrient intake.

[0031] Optionally, the core formula for predicting sarcopenia risk level is:

[0032] ;

[0033] yes , and The set, , and These are the predicted probability values ​​for low, medium, and high risk levels of sarcopenia for the nth sample output by the sarcopenia risk assessment model. It is the weight matrix of the main task output layer. It is the fusion feature of the nth sample. It is the bias vector of the main task output layer;

[0034] ;

[0035] in This is the sarcopenia risk level of the nth sample. , , and These are risk thresholds calibrated based on clinical data.

[0036] Optionally, a multimodal confidence fusion factor can be introduced into the core formula for predicting sarcopenia risk levels to improve prediction accuracy. The adjusted core formula is as follows:

[0037] ;

[0038] in It is the multimodal confidence fusion factor of the nth sample, which is calculated by weighting multiple dimensions of indicators. This represents an element-wise multiplication operation, using a multimodal confidence fusion factor to fuse features. Dynamic weighting is applied to enhance high-confidence modal features and suppress low-confidence modal features.

[0039] In summary, the present invention has at least one of the following beneficial technical effects:

[0040] This invention provides a knowledge graph-based method for sarcopenia risk assessment. Through multimodal spatiotemporal alignment and enhanced knowledge graph construction, it achieves deep fusion of multi-source heterogeneous data and spatiotemporal dynamic relationship modeling, significantly improving the comprehensiveness and accuracy of sarcopenia risk assessment.

[0041] Based on spatiotemporal graph convolutional networks and temporal causal graph networks, dynamic reasoning and causal verification of risk transmission paths are realized, making risk assessment results highly interpretable and providing clear targets for clinical intervention. Combined with the personalized intervention generation mechanism of digital twins and reinforcement learning, the optimal intervention plan can be output according to the individual characteristics of patients, improving intervention effect and patient compliance.

[0042] By employing federated knowledge distillation and adaptive differential privacy technology, collaborative learning of multi-center data is achieved while protecting patient privacy, effectively improving the model's generalization ability.

[0043] Through multi-dimensional visualization and a closed-loop clinical decision-making system, complex risk assessment results are transformed into clinically understandable structured reports to assist physicians in making efficient decisions. At the same time, the model self-evolves based on clinical feedback to continuously improve the accuracy of assessments. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the knowledge graph-based sarcopenia risk assessment method of the present invention. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings.

[0046] This invention discloses a sarcopenia risk assessment method based on knowledge graphs.

[0047] Reference Figure 1 Example 1, a knowledge graph-based method for sarcopenia risk assessment, includes the following steps:

[0048] Step 1: Collect multi-dimensional, multi-source data from patients and perform preprocessing;

[0049] Step 2: Define a six-tuple entity relationship model, introduce a dynamic entity lifecycle and heterogeneous relationship type system, and achieve multimodal feature fusion through modal adaptive weight learning, knowledge conflict resolution and entity evolution embedding;

[0050] Step 3: Based on the enhanced spatiotemporal graph convolutional network, through heterogeneous relational attention spatial convolution, spatiotemporal adaptive temporal convolution and triple contrastive learning task, the enhanced entity embedding is output. The enhanced entity embedding includes historical state, current state, evolution rate and future trend.

[0051] Step 4: Construct a temporal causal graph network. Use a temporal causal enhanced path mining algorithm to mine risky paths through bidirectional biased random walks. After multi-granularity causal verification, output an interventionable risk transmission map.

[0052] Step 5: Output sarcopenia risk level assessment based on the sarcopenia risk level assessment model. The sarcopenia risk level assessment model adopts a hierarchical fusion architecture with dynamic weight adjustment, and combines modal feature dual attention, path attention and adaptive graph readout function. The sarcopenia risk level assessment result is achieved through multi-task learning of inter-task attention mechanism.

[0053] Example 2 also includes step 6, which uses a dynamic early warning mechanism based on personalized thresholds to simulate the intervention effect using a patient digital twin model, and generates a personalized intervention plan that meets the constraints and is suitable for compliance.

[0054] The dynamic early warning mechanism sets personalized early warning thresholds based on age, comorbidities, and basic health status, and dynamically adjusts the early warning level in conjunction with the risk evolution rate; the digital twin model maps the patient's physiological state in real time and simulates the implementation effect of different intervention programs.

[0055] Example 3 also includes step 7, which adopts a federated learning framework of federated knowledge distillation and cross-domain knowledge transfer, combined with adaptive differential privacy, encrypted transmission and privacy leakage detection to achieve privacy protection and model update.

[0056] By employing the aforementioned technical solutions, multimodal heterogeneous data fusion and spatiotemporal alignment technologies are used to integrate multi-source data from clinical, imaging, time-series functional, biochemical, and nutritional fields, addressing the issue of data heterogeneity (differences in type, sampling frequency, and spatial source). Proactive data completion fills in missing values, cross-scale data fusion links macroscopic and microscopic data, and adaptive spatiotemporal alignment, through a unified time axis and dynamic frequency matching, enables collaborative analysis of different modalities across spatiotemporal dimensions, providing a consistent input foundation for subsequent knowledge graph construction and model learning.

[0057] Based on the theory of dynamic spatiotemporal knowledge representation and multimodal fusion, the traditional triplet is extended to a six-tuple entity relation model. Dynamic entity lifecycles and heterogeneous relation types (causality, association, temporal dependence, intervention response) are introduced, enabling the knowledge graph to dynamically depict the spatiotemporal evolution of sarcopenia (such as the occurrence and development of "grip strength decline events"). Multimodal feature fusion embedding maps heterogeneous features such as text, images, and numerical values ​​to a unified semantic space through modality-adaptive weight learning and knowledge conflict resolution. Simultaneously, entity evolution embedding encodes the dynamic changes in entity states, enhancing the knowledge graph's ability to express the complex pathological mechanisms of sarcopenia.

[0058] Employing a spatiotemporal joint representation learning technique, the spatial dimension aggregates neighbor node information from different types of relationships through a heterogeneous relation attention mechanism to capture the spatial associations of entities in the knowledge graph; the temporal dimension captures the evolutionary patterns of entity states through gated temporal convolution. Triple contrastive learning (entity-level, relation-level, and path-level) enhances embedding robustness, enabling the output enhanced entity embedding to contain four-dimensional information: historical state, current state, evolution rate, and future trend, achieving accurate modeling of the dynamic changes of sarcopenic muscular dystrophy entities.

[0059] Based on the fusion technology of temporal causal reasoning and graph network, a temporal causal graph network is constructed to integrate temporal knowledge graphs and causal rules. Through bidirectional biased random walks, it simultaneously mines the forward paths (from the source of risk to the outcome) and the reverse paths (tracing the source from the outcome). Multi-granularity causal verification verifies the causal strength of the paths at the molecular (omics data), organizational (imaging features), and population (clinical follow-up) levels, clarifies the causal relationship of sarcopenia risk transmission, and outputs an interventionizable risk transmission map.

[0060] This multi-task learning framework, driven by inter-task attention mechanisms, dynamically balances the learning weights of sarcopenia risk level classification (the primary task) with auxiliary tasks such as muscle function decline prediction and nutritional status assessment, leveraging inter-task correlations to improve the accuracy of the primary task. The hierarchical fusion architecture—feature-level, path-level, and graph-level—integrates multi-dimensional features and combines modality-feature dual attention, path attention, and adaptive graph readout functions to achieve accurate risk level assessment, while outputting interpretable results including risk probability, confidence level, and key supporting features.

[0061] Based on personalized medicine and digital twin technology, a dynamic early warning mechanism customizes warning thresholds according to the patient's age, comorbidities, and baseline health status. Combined with the rate of risk evolution, such as the degree of grip strength decline and the rate of muscle fat infiltration, the warning level is dynamically adjusted to achieve tiered and precise early warning. The digital twin model maps the patient's physiological state in real time, predicts changes in sarcopenia risk after intervention by simulating the effects of different intervention programs, and generates personalized intervention plans that meet individual constraints for each patient.

[0062] Employing federated knowledge distillation and adaptive privacy protection technologies, each medical institution trains its model locally and then uploads only the model knowledge to the central server for aggregation. This achieves multi-center collaborative learning while protecting patient privacy. Adaptive differential privacy adds dynamic noise during knowledge graph queries and publication to prevent individual information from being identified. Federated knowledge distillation improves the efficiency of multi-center model aggregation, supports incremental model learning, continuously absorbs new medical evidence and clinical data, and enables dynamic model updates.

[0063] Example 4: Through dynamic visualization, interactive risk simulation and multi-dimensional interpretation, a structured report aligned with clinical guidelines is output, and a clinical decision feedback interface is constructed.

[0064] By adopting the above technical solutions and using spatiotemporal data visualization technology, the historical evolution, current status and future trends of patients' sarcopenia risk are dynamically presented in the form of time axis and spatial topology map. At the same time, the key paths of risk transmission and intervention opportunity windows are highlighted, enabling clinicians to intuitively grasp the spatiotemporal distribution characteristics of the risk.

[0065] An interpretable interactive interface for the model is constructed to allow doctors to manually adjust intervention parameters, such as protein intake and resistance training frequency. By calling the risk prediction model in real time, the impact of parameter changes on indicators such as sarcopenia risk level and muscle function is simulated, providing a quantitative reference for the formulation of intervention plans.

[0066] This approach transforms the abstract decision-making logic of the model into clinically understandable content from three dimensions: mechanism, evidence, and intervention. The mechanism explanation elucidates the pathophysiological pathways of risk formation; the evidence explanation presents key data supporting the risk assessment, such as the degree of grip strength decline and the proportion of muscle fat infiltration; and the intervention explanation describes the pathway and expected effects of recommended interventions, thus achieving transparency in the risk assessment results.

[0067] It automatically aligns with the latest clinical guidelines, integrating risk assessment results, key evidence, and intervention recommendations into a structured report output; at the same time, it builds a physician feedback interface to receive feedback from clinical experts on the accuracy of risk assessment and the rationality of intervention plans, transforming it into input for model optimization and forming a closed loop for clinical decision-making.

[0068] Example 5: In step 1, the patient's multi-dimensional multi-source data includes clinical data, imaging data, time-series functional data, biochemical indicators, nutritional and behavioral data, omics data, environmental data, and subjective feeling data. The patient's multi-dimensional multi-source data is preprocessed through proactive data completion, cross-scale data fusion, and adaptive spatiotemporal alignment.

[0069] Example 6: In step 2, the six-tuple entity relationship model consists of a head entity, a relationship type, a tail entity, a timestamp, a spatial context, and a confidence level; the heterogeneous relationship type system includes causal relationships, association relationships, temporal dependencies, and intervention-response relationships.

[0070] By employing the aforementioned technical solutions, this approach addresses the pathophysiological mechanisms of sarcopenia, encompassing comprehensive data across clinical diagnosis, imaging representation, functional status, biochemical metabolism, lifestyle, molecular mechanisms, external environment, and subjective experience. Each data type captures sarcopenia-related characteristics from different dimensions: clinical data reflects the background of disease diagnosis and treatment; imaging data visually presents changes in muscle morphology and structure; time-series functional data records the dynamic evolution of muscle function; biochemical indicators reflect metabolic and nutritional status; omics data reveals molecular-level regulatory mechanisms; environmental data captures external influencing factors; and subjective experience data supplements the individual patient experience. Together, these elements construct a comprehensive and multi-dimensional patient data profile, avoiding assessment biases caused by single-dimensional data.

[0071] Based on generative models and clinical domain rules, the inherent distribution patterns and correlation characteristics of multi-source data are learned to reasonably infer and complete missing data. Clinical rules constrain the rationality of the completion results, ensuring that the completed data conforms to medical common sense, avoiding invalid data from interfering with subsequent analysis, and guaranteeing data integrity.

[0072] By employing cross-level feature association technology, we establish a mapping relationship between macro and micro data, explore potential associations between data at different scales, reveal the multi-scale regulatory links of sarcopenia from molecular to clinical phenotype, and improve the utilization rate of data information.

[0073] By using dynamic timeline construction technology to unify the time base of various types of data, and adopting an adaptive frequency matching algorithm to adapt to data with different sampling frequencies, the differences in time dimension are eliminated. At the same time, the spatial heterogeneity of data collected from multiple centers and multiple devices is corrected, so that data from different sources and of different types can achieve consistency in the spatiotemporal dimension, providing standardized input for subsequent multimodal knowledge fusion and model learning.

[0074] Building upon the traditional triplet, three key dimensions—timestamp, spatial context, and confidence level—are added to achieve a dynamic and precise characterization of sarcopenia-related entity relationships. The timestamp records the specific time when the entity relationship occurs, capturing its temporal attributes and adapting to the dynamic development characteristics of sarcopenia. Spatial context annotates spatial information such as the anatomical location and data collection scenario corresponding to the entity relationship, strengthening its spatial relevance. The confidence level quantifies the reliability of the entity relationship, assigning values ​​based on factors such as data quality and the strength of evidence support, providing a credibility reference for subsequent knowledge reasoning and preventing low-confidence relationships from affecting assessment accuracy. The synergistic effect of the six-tuple structure enables the knowledge graph to accurately express the spatiotemporal dynamics and reliability of sarcopenia-related entity relationships.

[0075] To address the complex association patterns in sarcopenia risk assessment, four core relationship types are defined to enable refined modeling of associations of different natures. Causal relationships depict causal links that directly lead to the occurrence or progression of sarcopenia, such as the causal association between insufficient nutrient intake and muscle mass loss. Association relationships capture entity connections without clear causality but with statistical correlation, such as the association between aging and muscle function decline. Temporal dependencies reflect the sequential dependence of entity states over time, such as the temporal association between a continuous decline in grip strength and an increased risk of subsequent falls. Intervention-response relationships describe the correspondence between intervention measures and their effects, such as the response association between resistance training and muscle strength improvement. These four relationship types comprehensively cover various association patterns related to sarcopenia, providing structured knowledge support for risk transmission path discovery and causal reasoning.

[0076] Example 7, step 4, the temporal causal enhancement path mining algorithm includes the following steps:

[0077] Step 41: Construct a temporal causal graph network to integrate temporal knowledge graphs and causal reasoning rules, and model dynamic causal relationships;

[0078] Step 42: Forwardly explore the risk transmission path and backward trace the risk source node to achieve a two-way biased random walk;

[0079] Step 43: Prioritize risky paths by combining three-dimensional scores of path strength, causal confidence, and intervention feasibility.

[0080] By adopting the above technical solution, the temporal knowledge graph provides data-driven temporal associations of entities, while the causal reasoning rules originate from medical consensus and pathophysiological mechanisms. The two are integrated to construct a temporal causal graph network. By embedding domain causal rules into the entity relationships of the temporal knowledge graph, the spuriousness of purely data-driven associations is constrained. At the same time, temporal attributes are used to characterize the temporal order of causal relationships, achieving accurate modeling of dynamic causal relationships and providing a reliable graph structure foundation for subsequent path mining.

[0081] Bidirectional biased random walks overcome the limitations of traditional unidirectional path mining. They forward-track the transmission chain along the timeline from the risk source to the sarcopenia phenotype, clearly identifying the risk diffusion path; and backward-track from the sarcopenia phenotype to pinpoint the core risk source node, providing complete coverage of the entire chain. Biasedness, by assigning higher jump probabilities to high-confidence, strongly correlated relationships, avoids meaningless random traversal, improving the efficiency and targeting of path mining.

[0082] The three-dimensional scoring system quantifies the value of risk pathways from three core dimensions: relevance, reliability, and usability. Path strength reflects the cumulative degree of correlation between entities along the path, demonstrating the tightness of risk transmission; causal confidence quantifies the true causal strength of the path based on causal verification results, excluding statistically spurious associations; intervention feasibility assesses the interveneability of path nodes by combining factors such as clinical technology maturity and patient tolerability. The three factors are weighted to calculate the overall path score, and high-value paths are prioritized after ranking by score, providing clear and feasible core targets for clinical intervention and avoiding interference from ineffective paths in decision-making.

[0083] Example 8, the output method of the sarcopenia risk level assessment results in step 5 is as follows:

[0084] Step 51: Input features include enhanced entity embedding, risk transmission path embedding, and patient global graph-level representation; construct input feature matrix X.

[0085] Step 52: Obtain the initial feature representation of each task through the task-specific feature mapping layer. Based on the inter-task correlation matrix C and the average task loss The dynamic weights for each task are calculated using the attention weight formula. ;

[0086] Step 53, using weights By integrating the characteristics of each task, The input to the main task output layer generates a risk level probability distribution through the Softmax function. ;

[0087] Step 54: Determine the risk level based on the clinical calibration threshold and calibrate the confidence level using a temperature scaling algorithm;

[0088] Step 55: The output includes risk level labels, corresponding probability values ​​and confidence levels, and is associated with key supporting features, including the trend of decreasing grip strength, the proportion of muscle fat infiltration and insufficient nutrient intake.

[0089] Example 9: The core formula for predicting sarcopenia risk level is:

[0090] ;

[0091] yes , and The set, , and These are the predicted probability values ​​for low, medium, and high risk levels of sarcopenia for the nth sample output by the sarcopenia risk assessment model. It is the weight matrix of the main task output layer. It is the fusion feature of the nth sample. It is the bias vector of the main task output layer;

[0092] ;

[0093] in This is the sarcopenia risk level of the nth sample. , , and These are risk thresholds calibrated based on clinical data.

[0094] Example 10: A multimodal confidence fusion factor is introduced into the core formula for predicting sarcopenia risk level to improve prediction accuracy. The adjusted core formula is as follows:

[0095] ;

[0096] in It is the multimodal confidence fusion factor of the nth sample, which is calculated by weighting multiple dimensions of indicators. This represents an element-wise multiplication operation, using a multimodal confidence fusion factor to fuse features. Dynamic weighting is applied to enhance high-confidence modal features and suppress low-confidence modal features.

[0097] By adopting the above technical solutions, enhanced entity embedding characterizes the dynamic features of local entities in patients, risk transmission path embedding reflects the key link information of risk transmission, and the patient's global graph-level representation captures the overall health status. The three components construct an input feature matrix from three dimensions: local, link, and global, to achieve comprehensive coverage of multi-scale features and provide a rich information foundation for risk assessment.

[0098] The task-specific feature mapping layer transforms shared input features into task-specific features, ensuring that each task receives targeted input. The inter-task correlation matrix quantifies the degree of association between tasks, and the task average loss reflects the learning difficulty and data quality of the task. The two are combined to calculate dynamic weights through the attention weight formula, so that the model automatically tilts resources toward tasks with high information value and stable learning, thereby improving the evaluation accuracy of the main task.

[0099] Dynamic weights are used to weight and fuse the features of each task, highlighting the role of high-contribution task features. The fused features are then input into the main task output layer. The Softmax function transforms the linear results of the output layer into a probability distribution of risk levels that conforms to probability axioms, intuitively presenting the model's confidence level for each risk level.

[0100] The clinical calibration threshold is obtained through statistical optimization based on a large amount of clinical data to ensure that the risk level determination meets clinical practice standards. The temperature scaling algorithm calibrates the confidence level of the model prediction by adjusting the steepness of the output probability distribution, avoiding overconfidence or underconfidence, and improving the reliability of the results.

[0101] The output integrates risk level labels, probability values, and confidence levels to meet the clinical need for quantitative assessment results. It also links key supporting features such as the trend of declining grip strength and the proportion of muscle fat infiltration, combining abstract probability outputs with specific medical indicators. This allows doctors to clearly trace the basis of risk assessment and achieve interpretability of assessment results.

[0102] , and These correspond to low, medium, and high risk confidence levels, respectively, and are clinically calibrated thresholds. , , and Based on statistical analysis of clinical case data, the judgment rules are determined to ensure that they conform to actual diagnosis and treatment scenarios. Risk levels are determined by combining multiple conditions to avoid misjudgments that may be caused by the principle of maximizing a single probability. For example, when the probability of low risk does not reach the high threshold but the probability of medium risk meets the conditions, it is judged as medium risk, thus achieving rigorous clinical-oriented risk classification.

[0103] Multimodal confidence fusion factor The value is calculated by weighting multiple dimensions of indicators such as data integrity, modality acquisition quality, and feature stability. Its magnitude directly reflects the overall reliability of the multimodal data of the nth sample. This is achieved through element-wise multiplication. Fusion features The features of each dimension are dynamically weighted; the feature dimensions corresponding to high confidence modalities are strengthened, while the feature dimensions corresponding to low confidence modalities are suppressed. This approach can effectively reduce the interference of low-quality data (such as noisy image data and time series data with many missing data) on the prediction results, making the model rely more on high-quality features for inference, thereby improving the accuracy and robustness of sarcopenia risk level prediction.

[0104] The following describes the implementation principle of the present invention using specific embodiments:

[0105] A tertiary hospital's geriatrics department selected a 68-year-old male patient for a sarcopenia risk assessment. The procedure is as follows:

[0106] Step 1: Collect multi-source data including clinical, imaging, temporal functional, biochemical, nutritional and behavioral, omics, environmental, and subjective feelings. Actively complete missing grip strength data, fuse gut microbiota and muscle function indicators across scales, and adaptively align the time axis to unify the time axis and correct for spatial heterogeneity.

[0107] Step 2: Construct a six-tuple such as "insufficient nutrient intake - causal relationship - muscle mass reduction - March 2024 - lower limb muscle - 0.92" to label the entity lifecycle. Achieve multimodal feature fusion through modality adaptive weight learning and knowledge conflict resolution to encode entity state changes.

[0108] Step 3: Based on the enhanced spatiotemporal graph convolutional network, through heterogeneous relational attention spatial convolution and spatiotemporal adaptive temporal convolution, combined with triple contrastive learning, the output is an enhanced entity embedding containing history, current state, evolution rate, and future trend.

[0109] Step 4: Construct a temporal causal graph network, mine risk transmission paths through bidirectional biased random walks, sort them by three-dimensional scoring, and output high-value interventionable paths such as "insufficient nutrient intake → reduced muscle mass → decreased grip strength".

[0110] Step 5: The model inputs enhanced entity embedding, path embedding and global graph-level representation, calculates the main task weight of 0.6 and the auxiliary task weight of 0.4, and generates a probability distribution through Softmax after feature fusion. Combined with clinical thresholds, the risk level is determined and associated with key supporting features.

[0111] Based on the patient's age, comorbid hypertension, and baseline health status, personalized early warning thresholds were set. A medium-risk warning was triggered by a monthly decrease in grip strength of 0.3 kg. A digital twin model was constructed to map the physiological state. After simulating three intervention programs, the "daily increase of 20g protein + resistance training twice a week" program was selected. This program predicted that the risk would decrease to low risk after 3 months. Personalized intervention and follow-up plans were then generated based on the patient's individual constraints.

[0112] Multi-center model optimization was conducted in collaboration with three top-tier hospitals:

[0113] After each hospital trains its model locally, it only uploads the distilled parameters to the central server, integrating heterogeneous data through cross-domain knowledge transfer. Parameter transmission uses homomorphic encryption, and adaptive differential privacy noise is added to knowledge graph queries, embedding a privacy leakage detection module. New data and medical evidence are regularly absorbed, and the model and knowledge graph are updated through incremental learning.

[0114] The dynamic visualization interface displays changes in muscle mass and grip strength over time, while a spatial topology map shows the risk transmission path and highlights the intervention window. Doctors can adjust intervention parameters through an interactive interface and view real-time risk predictions. The system interprets results from three dimensions: mechanism, evidence, and intervention, automatically aligns with clinical guidelines to generate structured reports, and receives feedback from doctors to update treatment plans and models, forming a closed-loop decision-making process.

[0115] Eight categories of comprehensive data were collected. In the preprocessing stage, missing dietary data were supplemented by generating models and combining them with clinical rules. The gut microbiota and muscle mass were correlated across levels. A unified time reference was established and the differences in sampling frequency were adapted. Measurement biases from multiple devices were corrected to achieve spatiotemporal alignment of data.

[0116] We constructed a six-tuple system, including "resistance training - intervention response relationship - muscle strength improvement - May 2024 - upper limb training - 0.95", which covers four types of relationships: causality, association, temporal dependence, and intervention response, accurately depicting the dynamic association and evolution process of entities.

[0117] Step 41: Integrate temporal knowledge graphs and medical causal rules to model dynamic causal relationships; Step 42: Use bidirectional biased random walks to discover positive transmission paths and reverse risk sources; Step 43: Output the optimal path "insufficient nutrient intake → vitamin D deficiency → muscle mass reduction → decreased grip strength" through three-dimensional scoring of path strength, causal confidence, and intervention feasibility.

[0118] Step 51: Input features include enhanced entity embedding (containing muscle function history, current, evolution, and trend information), risk transmission path embedding (feature representation of core paths), and patient global graph-level representation (overall health status features), constructing a 3×128 dimensional input feature matrix.

[0119] Step 52: Through the task-specific feature mapping layer, the input features are transformed into initial feature representations for the four tasks. Based on the correlation matrix between tasks and the average loss of each task, the weight of the main task (sarcopenia risk classification) is calculated to be 0.62, and the weights of the auxiliary tasks are 0.21 for nutritional status assessment, 0.13 for muscle function decline prediction, and 0.04 for fall risk prediction.

[0120] Step 53: Use the dynamic weights described above to fuse the features of each task to obtain the fused feature Ffusion, input it into the output layer of the main task, and generate the risk level probability distribution through the Softmax function. ,in (Low risk) is 0.32. (Medium risk) is 0.58. (High risk) is 0.10.

[0121] Step 54: Based on clinically calibrated thresholds θ1=0.65, θ2=0.35, θ3=0.45, θ4=0.55, the patient... =0.32<θ2=0.35, and =0.58≥θ3=0.45, the risk level is determined to be medium risk for Graden; the confidence level is calibrated using a temperature scaling algorithm, after calibration... For 0.30, 0.61 It is 0.09.

[0122] Step 55: The output includes a medium-risk level label. =0.32、 =0.58、 The probability value of 0.10 and the calibrated confidence level are used to correlate key supporting features such as the decreasing grip strength trend, high muscle fat infiltration ratio, and insufficient nutrient intake.

[0123] Application of the core formula for risk prediction: A multimodal confidence fusion factor τn is introduced into the core formula. This factor is calculated by weighting indicators such as patient data completeness, image data quality, and feature stability, and its value is 0.88. Through element-wise multiplication, τn dynamically weights the fusion feature Ffusion,n, enhancing the contribution of high-quality DXA image features and clinical data features, and suppressing the interference of subjective perception data features with significant missing data. The adjusted result is obtained through the formula calculation. =0.31、 =0.60、 =0.09, the risk level remains medium, but the robustness of the forecast results has been significantly improved.

[0124] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for sarcopenia risk assessment, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional, multi-source data from patients and perform preprocessing; Step 2: Define a six-tuple entity relationship model, introduce a dynamic entity lifecycle and heterogeneous relationship type system, and achieve multimodal feature fusion through modal adaptive weight learning, knowledge conflict resolution and entity evolution embedding; Step 3: Based on the enhanced spatiotemporal graph convolutional network, through heterogeneous relational attention spatial convolution, spatiotemporal adaptive temporal convolution and triple contrastive learning tasks, the enhanced entity embedding is output. The enhanced entity embedding includes historical state, current state, evolution rate and future trend; the triple contrastive learning tasks are entity-level contrastive learning, relation-level contrastive learning and path-level contrastive learning. Step 4: Construct a temporal causal graph network. Use a temporal causal enhanced path mining algorithm to mine risky paths through bidirectional biased random walks. After multi-granularity causal verification, output an interventionable risk transmission map. Step 5: Output sarcopenia risk level assessment based on the sarcopenia risk level assessment model. The sarcopenia risk level assessment model adopts a hierarchical fusion architecture with dynamic weight adjustment, and combines modal feature dual attention, path attention and adaptive graph readout function. The sarcopenia risk level assessment result is achieved through multi-task learning of inter-task attention mechanism. The six-tuple entity relationship model described in step 2 consists of a head entity, a relationship type, a tail entity, a timestamp, a spatial context, and a confidence level; the heterogeneous relationship type system includes causal relationships, association relationships, temporal dependencies, and intervention-response relationships.

2. The knowledge graph-based sarcopenia risk assessment method according to claim 1, characterized in that, It also includes step 6, a dynamic early warning mechanism based on personalized thresholds, which combines the patient's digital twin model to simulate the intervention effect and generate a personalized intervention plan that meets the constraints and is suitable for compliance. The dynamic early warning mechanism sets personalized early warning thresholds based on age, comorbidities, and basic health status, and dynamically adjusts the early warning level in conjunction with the risk evolution rate. Digital twin models can map a patient's physiological state in real time and simulate the effects of different intervention programs.

3. The knowledge graph-based sarcopenia risk assessment method according to claim 2, characterized in that, It also includes step 7, which adopts a federated learning framework of federated knowledge distillation and cross-domain knowledge transfer, and combines adaptive differential privacy, encrypted transmission and privacy leakage detection to achieve privacy protection and model update.

4. The knowledge graph-based sarcopenia risk assessment method according to claim 3, characterized in that, Through dynamic visualization, interactive risk simulation, and multi-dimensional interpretation, it outputs structured reports aligned with clinical guidelines and builds a clinical decision feedback interface.

5. The knowledge graph-based sarcopenia risk assessment method according to claim 4, characterized in that, In step 1, the patient's multi-dimensional and multi-source data includes clinical data, imaging data, time-series functional data, biochemical indicators, nutritional and behavioral data, omics data, environmental data, and subjective feeling data. The patient's multi-dimensional and multi-source data is preprocessed through proactive data completion, cross-scale data fusion, and adaptive spatiotemporal alignment.

6. The knowledge graph-based sarcopenia risk assessment method according to claim 5, characterized in that, In step 4, the temporal causal reinforcement path mining algorithm includes the following steps: Step 41: Construct a temporal causal graph network to integrate temporal knowledge graphs and causal reasoning rules, and model dynamic causal relationships; Step 42: Forwardly explore the risk transmission path and backward trace the risk source node to achieve a two-way biased random walk; Step 43: Prioritize risky paths by combining three-dimensional scores of path strength, causal confidence, and intervention feasibility.

7. The knowledge graph-based sarcopenia risk assessment method according to claim 6, characterized in that, The output method for the sarcopenia risk level assessment results in step 5 is as follows: Step 51: Input features include enhanced entity embedding, risk transmission path embedding, and patient global graph-level representation; construct input feature matrix X. Step 52: Obtain the initial feature representation of each task through the task-specific feature mapping layer. Based on the inter-task correlation matrix C and the average task loss The dynamic weights for each task are calculated using the attention weight formula. ; Step 53, using weights By integrating the characteristics of each task, The input to the main task output layer generates a risk level probability distribution through the Softmax function. ; Step 54: Determine the risk level based on the clinical calibration threshold and calibrate the confidence level using a temperature scaling algorithm; Step 55: The output includes risk level labels, corresponding probability values ​​and confidence levels, and is associated with key supporting features, including the trend of decreasing grip strength, the proportion of muscle fat infiltration and insufficient nutrient intake.

8. The knowledge graph-based sarcopenia risk assessment method according to claim 7, characterized in that, The core formula for predicting the risk level of sarcopenia is: ; yes , and The set, , and These are the predicted probability values ​​for low, medium, and high risk levels of sarcopenia for the nth sample output by the sarcopenia risk assessment model. It is the weight matrix of the main task output layer. It is the fusion feature of the nth sample. It is the bias vector of the main task output layer; ； in This is the sarcopenia risk level of the nth sample. , , and These are risk thresholds calibrated based on clinical data.

9. The knowledge graph-based sarcopenia risk assessment method according to claim 8, characterized in that, To improve prediction accuracy, a multimodal confidence fusion factor is introduced into the core formula for predicting sarcopenia risk levels. The adjusted core formula is as follows: ; in It is the multimodal confidence fusion factor of the nth sample, which is calculated by weighting multiple dimensions of indicators. This represents an element-wise multiplication operation, using a multimodal confidence fusion factor to fuse features. Dynamic weighting is applied to enhance high-confidence modal features and suppress low-confidence modal features.

Images