Orthopedic risk intelligent decision system based on multi-modal bioelectric signal fusion

Abstract

The present application relates to the field of medical information processing, and particularly refers to an orthopedic risk intelligent decision system based on multi-modal bioelectric signal fusion, comprising an intelligent recognition module, an individual physiological atlas dynamic modeling module, a decision reasoning module, a dynamic evaluation optimization module and a risk prevention module, the present application constructs an individual physiological atlas based on a graph neural network, fuses historical physiological data of a user with real-time multi-modal bioelectric signals for modeling, realizes dynamic representation of individual physiological states, effectively depicts individual difference characteristics of the user, and thus significantly improves the accuracy of abnormal physiological state recognition; by introducing a pre-clustering mechanism and a muscle-bone collaborative fan-shaped topological subgraph structure, combining a graph neural network model of local neighborhood aggregation and long-distance similar neighborhood aggregation, multi-scale modeling of complex correlation relationships in multi-modal physiological data across regions and modalities is realized, and the stability and reliability of risk prediction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical information processing, specifically to an intelligent decision-making system for orthopedic risks based on the fusion of multimodal bioelectrical signals. Background Technology

[0002] With the development of smart healthcare, orthopedic health status monitoring and risk assessment based on bioelectrical signals has gradually become a research hotspot. However, existing technologies still have the following problems in the fusion analysis of multimodal bioelectrical signals and orthopedic risk decision-making: (1) Insufficient ability to model individual differences Existing methods mostly use a unified model to model users' physiological signals, lacking the ability to jointly model individual historical physiological data and real-time status, making it difficult to construct dynamic physiological representations that can reflect individual differences, resulting in low personalization and insufficient accuracy of orthopedic risk assessment results; (2) Limited ability to depict complex physiological relationships Existing technologies typically process multimodal physiological data based on simple feature splicing or local neighborhood modeling, which makes it difficult to effectively capture the long-distance correlation between the nervous and muscular systems and the multi-scale synergistic mechanism. In complex orthopedic risk prediction scenarios, including orthopedic-related risk scenarios such as osteoarthritis, sports injuries and osteoporosis, there are problems of insufficient modeling ability and poor prediction stability. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent decision-making system for orthopedic risk based on the fusion of multimodal bioelectrical signals. Addressing the problem of insufficient modeling ability for individual differences, this invention constructs an individual physiological atlas based on graph neural networks, fusing historical physiological data with real-time multimodal bioelectrical signals to achieve dynamic representation of individual physiological states, effectively characterizing individual differences and significantly improving the accuracy of abnormal physiological state identification. Addressing the limited ability to characterize complex physiological relationships, this invention introduces a pre-clustering mechanism and a musculoskeletal collaborative fan-shaped topological subgraph structure, combined with a graph neural network model of local neighborhood aggregation and long-distance similar neighborhood aggregation. This enables multi-scale modeling of complex cross-regional and cross-modal relationships in multimodal physiological data, effectively improving the modeling ability of cross-joint kinetic chain collaborative patterns in the musculoskeletal system. Compared to traditional graph neural network methods based solely on local neighborhood propagation, this significantly improves prediction accuracy and result stability in complex orthopedic risk prediction tasks.

[0004] The technical solution adopted in this invention is as follows: The intelligent decision-making system for orthopedic risks based on multimodal bioelectrical signal fusion provided by this invention includes an intelligent identification module, an individual physiological atlas dynamic modeling module, a decision reasoning module, a dynamic evaluation and optimization module, and a risk protection module, specifically including the following: The intelligent recognition module collects multimodal bioelectrical signal data of users, including electromyography (EMG) signals and electroencephalography (EEG) signals, by using a wearable EMG acquisition device. It combines joint motion state signals and musculoskeletal dynamics auxiliary information, performs preprocessing and feature extraction, and obtains the user's musculoskeletal functional state feature vector based on the Transformer multi-task deep neural network model. The individual physiological atlas dynamic modeling module collects users’ historical physiological data, constructs a user-personalized physiological atlas based on graph neural networks, and obtains the user’s current individual physiological atlas representation. The decision reasoning module outputs prediction results of individual orthopedic risk level, probability of musculoskeletal injury, and trend of degenerative lesions based on individual physiological atlas representation, combined with context awareness mechanism and supervised learning-based risk prediction model; The dynamic assessment and optimization module receives updated physiological signal data from the user in real time, dynamically corrects the orthopedic risk assessment results, and uses reinforcement learning algorithms to adaptively optimize the parameters of the risk prediction model. The risk protection module analyzes muscle fatigue evolution trends, joint stability changes, and neuromuscular control abnormalities by constructing a time-series anomaly detection model. When an abnormal risk is detected, it outputs orthopedic risk warning information and corresponding rehabilitation intervention suggestions.

[0005] Furthermore, the individual physiological atlas dynamic modeling module collects the user's historical physiological data, constructs a user-personalized physiological atlas based on a graph neural network, and obtains the user's current individual physiological atlas representation, specifically including the following steps: Step S1: Graph structure construction, defining the graph structure , The nodes represent various data types, including skeletal structure nodes, neural function region nodes, muscle group nodes, and physiological function indicator nodes. The edges are represented, including muscle-skeletal mechanical coupling relationships, neuromuscular control pathways, inter-joint kinetic chain transmission relationships, and statistical correlation relationships, to construct the initial static physiological graph structure; Step S2: Node feature construction and fusion. The user's musculoskeletal functional state feature vector, the collected multimodal bioelectrical signal data, and the user's historical physiological data are fused to construct the node feature vector. Step S3: Feature dimensionality reduction. Linear discriminant analysis (LDA) is used to reduce the dimensionality of the node feature vectors to obtain the dimensionality-reduced node features. The KMeans++ algorithm is used for pre-clustering to generate K similar node clusters, and the feature centroid of each cluster is calculated. Step S4: Sector topology subgraph construction. A structured subgraph is constructed based on similar node clusters. The node corresponding to the feature centroid of each node cluster is used as the center node to construct a musculoskeletal coordinating sector topology subgraph. Within each node cluster, a similarity metric is constructed based on the feature distance between the node and the center node. A similarity threshold is determined based on the statistical distribution of distances within the node cluster. Nodes whose feature distance to the center node is less than the threshold and satisfy the local density constraint are selected as peripheral nodes. A weighted connection relationship is established between the center node and the peripheral nodes. A distance decay mechanism is introduced to modulate the edge weights to obtain an enhanced graph structure. Step S5: Pre-clustering enhances graph neural network modeling. Construct a graph neural network model that integrates local neighborhood aggregation and long-distance similar neighborhood aggregation. Perform local neighborhood feature aggregation based on the initial static physiological graph structure to obtain the local structural representation of the nodes. Perform long-distance similar node feature aggregation based on the musculoskeletal coordinating fan-shaped topological subgraph to obtain the cross-regional semantic association representation. Fuse the local structural representation and cross-regional semantic association representation of the nodes to complete multi-scale feature learning and obtain the updated node embedding representation. Step S6: Graph output, which integrates the updated node embedding representation and the enhanced graph structure to generate the user's current individual physiological graph representation.

[0006] Furthermore, step S4 specifically includes the following steps: Step S41: Identify the musculoskeletal coordinating sector topology center node. For each similar node cluster, calculate its cluster centroid feature vector as the representative node of the cluster. Calculate the Euclidean distance between all nodes in the cluster and the centroid feature vector, and select the closest node as the musculoskeletal coordinating sector topology center node of the cluster. The formula used is as follows: ; in, Indicates the first The central node of the musculoskeletal co-operational sector topology selected from a cluster of similar nodes. Represents a node The dimensionality-reduced feature vectors Indicates the first The characteristic centroids of a cluster of similar nodes; Step S42: Subgraph node filtering. Within each similar node cluster, sort nodes according to their Euclidean distance to the musculoskeletal co-occurrence sector topology center node before filtering. The set of outer nodes is composed of nodes, represented as follows: ; in, Indicates the first The set of peripheral nodes selected from clusters of similar nodes Indicates candidate outer nodes, Indicates the first The set of all nodes contained in a cluster of similar nodes. Represents a node The feature vector representation after dimensionality reduction in step S3, Represents a node The feature vector representation after dimensionality reduction in step S3, This means taking the nearest neighbor. 1 node This indicates the threshold for the number of peripheral nodes; Step S43: Calculate the distance penalty factor. Define the distance penalty factor and calculate the edge weight for each connected edge using the following formula: ; in, express and The weight of the edges connecting them. Indicates the distance attenuation coefficient; Step S44: Musculoskeletal cooperative sector structure connection: Connect the central node with each node in the outer node set to construct a musculoskeletal cooperative sector topology subgraph. The set of connecting edges is represented as: ; in, Indicates the first The set of musculoskeletal cooperative fan-shaped structure connecting edges corresponding to clusters of similar nodes. Indicates the first The outer nodes selected from clusters of similar nodes Indicates the first The set of subgraph nodes corresponding to each cluster of similar nodes, including and All peripheral nodes; Step S45: Merge into the original graph structure. Merge the musculoskeletal co-operational sector topology subgraph into the initial version of the static graph structure from step S1 to obtain the enhanced graph structure.

[0007] Furthermore, step S5 specifically includes the following steps: Step S51: Initialize the local neighborhood aggregation module. Initialize the graph attention network based on the multi-head graph attention mechanism, learn the attention weights between the dimensionality-reduced node feature vector and its one-hop neighbors, and use learnable attention coefficients to weight and aggregate the features of neighboring nodes to obtain the local neighborhood embedding representation of each node. Step S52: Construction of the long-distance similarity neighborhood aggregation module. Based on the musculoskeletal coordinating fan-shaped topological subgraph constructed in Step S4, for the subgraphs corresponding to the K similar node clusters generated in Step S3, the central node of each subgraph is used as the aggregation core. A graph attention mechanism is introduced in each subgraph to perform weighted aggregation of the features of the peripheral nodes, thereby obtaining the long-distance similarity neighborhood embedding representation of each central node. The formula used is as follows: ; ; in, Indicates the central node The long-distance similar neighborhood embedding representation, Represents a non-linear activation function. Represents a node For nodes Attention weights This represents the learnable linear transformation weight matrix. It is a node The feature representation after dimensionality reduction, It is a node The feature representation after dimensionality reduction, Indicates the attenuation factor. Indicates the structural distance influence factor. It is to traverse the set of outer nodes. index variable, Represents a node The feature vector after dimensionality reduction in step S3, Represents a node With nodes Topological distance; Step S53: Node feature fusion. The local neighborhood embedding representation and the far-distance similar neighborhood embedding representation of each node are subjected to max pooling operation, and feature dimension alignment and fusion processing are performed to generate a unified embedding representation for each node. Step S54: Output update. The unified embedding representation of each node is enhanced with feature representation through residual connection mechanism and nonlinear transformation, and the Dropout mechanism is combined to suppress overfitting to obtain the updated node representation.

[0008] The beneficial effects achieved by the present invention using the above solution are as follows: (1) To address the problem of insufficient individual difference modeling ability, this invention constructs an individual physiological map based on graph neural network, integrates user historical physiological data with real-time multimodal bioelectric signals to achieve dynamic representation of individual physiological state, effectively characterizes the individual difference features of users, and thus significantly improves the accuracy of abnormal physiological state identification. (2) To address the problem of limited ability to characterize complex physiological relationships, this invention introduces a pre-clustering mechanism and a musculoskeletal collaborative fan-shaped topological subgraph structure, combined with a graph neural network model of local neighborhood aggregation and long-distance similar neighborhood aggregation, to achieve multi-scale modeling of complex cross-regional and cross-modal relationships in multimodal physiological data. This effectively improves the ability to model cross-joint kinematic chain collaborative patterns in the musculoskeletal system. Compared with traditional graph neural network methods based only on local neighborhood propagation, this invention significantly improves prediction accuracy and result stability in complex orthopedic risk prediction tasks. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the intelligent decision-making system for orthopedic risks based on multimodal bioelectrical signal fusion proposed in this invention; Figure 2 This is a flowchart illustrating the dynamic modeling module for individual physiological atlases proposed in this invention.

[0010] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Example 1, see Figure 1 The intelligent decision-making system for orthopedic risks based on multimodal bioelectrical signal fusion provided by this invention includes an intelligent identification module, an individual physiological atlas dynamic modeling module, a decision reasoning module, a dynamic evaluation and optimization module, and a risk protection module, specifically including the following: The intelligent recognition module collects multimodal bioelectrical signal data of users, including electromyography (EMG) signals and electroencephalography (EEG) signals, by using a wearable EMG acquisition device. It combines joint motion state signals and musculoskeletal dynamics auxiliary information, performs preprocessing and feature extraction, and obtains the user's musculoskeletal functional state feature vector based on the Transformer multi-task deep neural network model. The individual physiological atlas dynamic modeling module collects users’ historical physiological data, constructs a user-personalized physiological atlas based on graph neural networks, and obtains the user’s current individual physiological atlas representation. The decision reasoning module outputs prediction results of individual orthopedic risk level, probability of musculoskeletal injury, and trend of degenerative lesions based on individual physiological atlas representation, combined with context awareness mechanism and supervised learning-based risk prediction model; The dynamic assessment and optimization module receives updated physiological signal data from the user in real time, dynamically corrects the orthopedic risk assessment results, and uses reinforcement learning algorithms to adaptively optimize the parameters of the risk prediction model. The risk protection module analyzes muscle fatigue evolution trends, joint stability changes, and neuromuscular control abnormalities by constructing a time-series anomaly detection model. When an abnormal risk is detected, it outputs orthopedic risk warning information and corresponding rehabilitation intervention suggestions.

[0013] Example 2, see Figure 2 The individual physiological atlas dynamic modeling module collects users' historical physiological data, constructs a user-personalized physiological atlas based on graph neural networks, and obtains the user's current individual physiological atlas representation, specifically including the following steps: Step S1: Graph structure construction, defining the graph structure , The nodes represent various data types, including skeletal structure nodes, neural function region nodes, muscle group nodes, and physiological function indicator nodes. The edges are represented, including muscle-skeletal mechanical coupling relationships, neuromuscular control pathways, inter-joint kinetic chain transmission relationships, and statistical correlation relationships, to construct the initial static physiological graph structure; Step S2: Node feature construction and fusion. The user's musculoskeletal functional state feature vector, the collected multimodal bioelectrical signal data, and the user's historical physiological data are fused to construct the node feature vector. Step S3: Feature dimensionality reduction. Linear discriminant analysis (LDA) is used to reduce the dimensionality of the node feature vectors to obtain the dimensionality-reduced node features. The KMeans++ algorithm is used for pre-clustering to generate K similar node clusters, and the feature centroid of each cluster is calculated. Step S4: Sector topology subgraph construction. A structured subgraph is constructed based on similar node clusters. The node corresponding to the feature centroid of each node cluster is used as the center node to construct a musculoskeletal coordinating sector topology subgraph. Within each node cluster, a similarity metric is constructed based on the feature distance between the node and the center node. A similarity threshold is determined based on the statistical distribution of distances within the node cluster. Nodes whose feature distance to the center node is less than the threshold and satisfy the local density constraint are selected as peripheral nodes. A weighted connection relationship is established between the center node and the peripheral nodes. A distance decay mechanism is introduced to modulate the edge weights to obtain an enhanced graph structure. Step S5: Pre-clustering enhances graph neural network modeling. Construct a graph neural network model that integrates local neighborhood aggregation and long-distance similar neighborhood aggregation. Perform local neighborhood feature aggregation based on the initial static physiological graph structure to obtain the local structural representation of the nodes. Perform long-distance similar node feature aggregation based on the musculoskeletal coordinating fan-shaped topological subgraph to obtain the cross-regional semantic association representation. Fuse the local structural representation and cross-regional semantic association representation of the nodes to complete multi-scale feature learning and obtain the updated node embedding representation. Step S6: Graph output, which integrates the updated node embedding representation and the enhanced graph structure to generate the user's current individual physiological graph representation.

[0014] In this embodiment, taking the risk assessment scenario of knee osteoarthritis as an example, the user wears an integrated bioelectrical acquisition device, which includes an electromyography (EMG) sensor and an electroencephalogram (EEG) acquisition module. The EMG signal sampling frequency is 1000Hz and the EEG signal sampling frequency is 256Hz. The device continuously collects multimodal bioelectrical signal data of the user in daily activities, and combines it with the user's historical physical examination data, including heart rate variability index, blood oxygen saturation and bone mineral density related indexes, as historical physiological data input system. Constructing physiological map structure The node set V includes nodes of neural functional regions, muscle groups, and physiological function indicators; the edge set E includes connections of neuromuscular control pathways, functional synergistic relationships, and statistical correlation relationships calculated based on historical data. For each node, construct a feature vector, including the following: EEG node characteristics: alpha wave and beta wave power spectral density, and band energy ratio; Electromyographic node characteristics: root mean square (RMS) value, median frequency (MF); Physiological indicator node characteristics: HRV time-domain / frequency-domain indicators, SpO2 level; The above features are then combined with user historical data to form a node feature vector with a unified dimension. The LDA method is used to reduce the original high-dimensional node features to a 32-dimensional semantic space, and the KMeans++ algorithm is used to pre-cluster all nodes to obtain the following categories: C1: Neural regulation-related nodes; C2: Muscle activity-related nodes; C3: Cardiovascular functional node; C4: Metabolic and skeletal-related nodes; C5: Cross-system coupling node; Calculate the eigencentroid vectors of each cluster; Construct a musculoskeletal cooperative sector-shaped topological subgraph for each node cluster: In each cluster, the node closest to the centroid is selected as the central node. The nodes are sorted according to Euclidean distance, and the first T=5 nodes are selected as the outer nodes. The connection relationship between the central node and the outer nodes is constructed. The edge weight is calculated using an exponential decay function. When the distance between nodes is small, the edge weight approaches 1. When the distance increases, the edge weight decreases rapidly. The final result is an enhanced graph structure; Constructing a two-channel graph neural network model includes the following: (1) Local neighborhood aggregation: Based on the initial graph structure, a multi-head graph attention mechanism is used to aggregate one-hop neighbor information to obtain the local structural representation of the node; (2) Aggregation of distant similar neighborhoods: In the musculoskeletal co-operation fan-shaped topological subgraph, attention weighted aggregation of features of peripheral nodes is performed with the central node as the center to obtain cross-regional semantic representation; (3) Feature fusion: The local structural representation and the long-distance semantic representation are fused through max pooling to obtain a unified node embedding; The updated node embedding representation is fused with the enhanced graph structure to generate a representation of the user's current individual physiological map.

[0015] Example 3, this example is based on the above example, step S4 specifically includes the following steps: Step S41: Identify the musculoskeletal coordinating sector topology center node. For each similar node cluster, calculate its cluster centroid feature vector as the representative node of the cluster. Calculate the Euclidean distance between all nodes in the cluster and the centroid feature vector, and select the closest node as the musculoskeletal coordinating sector topology center node of the cluster. The formula used is as follows: ; in, Indicates the first The central node of the musculoskeletal co-operational sector topology selected from a cluster of similar nodes. Represents a node The dimensionality-reduced feature vectors Indicates the first The characteristic centroids of a cluster of similar nodes; Step S42: Subgraph node filtering. Within each similar node cluster, sort nodes according to their Euclidean distance to the musculoskeletal co-occurrence sector topology center node before filtering. The set of outer nodes is composed of nodes, represented as follows: ; in, Indicates the first The set of peripheral nodes selected from clusters of similar nodes Indicates candidate outer nodes, Indicates the first The set of all nodes contained in a cluster of similar nodes. Represents a node The feature vector representation after dimensionality reduction in step S3, Represents a node The feature vector representation after dimensionality reduction in step S3, This means taking the nearest neighbor. 1 node This indicates the threshold for the number of peripheral nodes; Step S43: Calculate the distance penalty factor. Define the distance penalty factor and calculate the edge weight for each connected edge using the following formula: ; in, express and The weight of the edges connecting them. Indicates the distance attenuation coefficient; Step S44: Musculoskeletal cooperative sector structure connection: Connect the central node with each node in the outer node set to construct a musculoskeletal cooperative sector topology subgraph. The set of connecting edges is represented as: ; in, Indicates the first The set of musculoskeletal cooperative fan-shaped structure connecting edges corresponding to clusters of similar nodes. Indicates the first The outer nodes selected from clusters of similar nodes Indicates the first The set of subgraph nodes corresponding to each cluster of similar nodes, including and All peripheral nodes; Step S45: Merge into the original graph structure. Merge the musculoskeletal co-operational sector topology subgraph into the initial version of the static graph structure from step S1 to obtain the enhanced graph structure.

[0016] In this embodiment, for the cluster of similar nodes in the "femoral anteflexion muscle group" category, let the cluster contain the following dimensionality-reduced feature vectors of nodes: Node A: Quadriceps femoris ; Node B: Sartorius muscle ; Node C: Rectus femoris ; Node D: Tensor fasciae latae ; Node E: Electrode ; After performing KMeans++ clustering on these nodes, the centroid feature vector of the cluster is obtained, and then the Euclidean distance from each node to the centroid is calculated: ; ; ; ; ; Choose the quadriceps femoris muscle as the actual node closest to the center of mass. That is, the musculoskeletal co-operational fan-shaped topology center node of the cluster; Using the quadriceps femoris muscle as the central node, calculate the Euclidean distances to the other nodes: ; ; ; ; Sort by distance and select the first T=3 nodes: ; Let distance penalty factor Calculate the corresponding edge weights according to the formula: ; ; ; Constructing a sector-shaped subgraph ,in: ; ; Will Insert the initial static graph structure obtained in step S1, and update to obtain the enhanced graph structure.

[0017] Example 4: This example is based on the above examples. Step S5 specifically includes the following steps: Step S51: Initialize the local neighborhood aggregation module. Initialize the graph attention network based on the multi-head graph attention mechanism, learn the attention weights between the dimensionality-reduced node feature vector and its one-hop neighbors, and use learnable attention coefficients to weight and aggregate the features of neighboring nodes to obtain the local neighborhood embedding representation of each node. Step S52: Construction of the long-distance similarity neighborhood aggregation module. Based on the musculoskeletal coordinating fan-shaped topological subgraph constructed in Step S4, for the subgraphs corresponding to the K similar node clusters generated in Step S3, the central node of each subgraph is used as the aggregation core. A graph attention mechanism is introduced in each subgraph to perform weighted aggregation of the features of the peripheral nodes, thereby obtaining the long-distance similarity neighborhood embedding representation of each central node. The formula used is as follows: ; ; in, Indicates the central node The long-distance similar neighborhood embedding representation, Represents a non-linear activation function. Represents a node For nodes Attention weights This represents the learnable linear transformation weight matrix. It is a node The feature representation after dimensionality reduction, It is a node The feature representation after dimensionality reduction, Indicates the attenuation factor. Indicates the structural distance influence factor. It is to traverse the set of outer nodes. index variable, Represents a node The feature vector after dimensionality reduction in step S3, Represents a node With nodes Topological distance; Step S53: Node feature fusion. The local neighborhood embedding representation and the far-distance similar neighborhood embedding representation of each node are subjected to max pooling operation, and feature dimension alignment and fusion processing are performed to generate a unified embedding representation for each node. Step S54: Output update. The unified embedding representation of each node is enhanced with feature representation through residual connection mechanism and nonlinear transformation, and the Dropout mechanism is combined to suppress overfitting to obtain the updated node representation.

[0018] In this embodiment, the code used for the main graph neural network model is as follows: class PersonalizedStimGNN(nn.Module): def __init__(self, in_dim, hidden_dim, out_dim, heads=4, dropout=0.3): super().__init__() self.local_gat = GATConv(in_dim, hidden_dim, heads=heads, dropout=dropout) self.remote_gat = RemoteGATLayer(in_dim, hidden_dim) self.fusion = nn.Linear(hidden_dim 2, out_dim) self.residual = nn.Linear(in_dim, out_dim) self.dropout = nn.Dropout(dropout) def forward(self, x, edge_index, batch, cluster_nodes, center_idx,dist_matrix): # S51: Local Neighborhood Aggregation (Local GAT) h_local = self.local_gat(x, edge_index) h_local = F.elu(h_local) # S52: Aggregation of Long-Distance Similar Neighborhoods h_remote = self.remote_gat(x, cluster_nodes, center_idx, dist_matrix) # S53: Feature Fusion (Max Pooling) h_combined = torch.max(h_local, h_remote) # or torch.cat followed by the wiring layer # S54: Output Update + Residual Join + Dropout out = self.fusion(h_combined) + self.residual(x) out = self.dropout(F.elu(out)) return out.

[0019] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0020] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0021] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent decision-making system for orthopedic risk based on multimodal bioelectrical signal fusion, characterized in that: It includes an intelligent recognition module, an individual physiological atlas dynamic modeling module, a decision reasoning module, a dynamic evaluation and optimization module, and a risk protection module, specifically including the following: The intelligent recognition module collects multimodal bioelectrical signal data of users, including electromyography (EMG) signals and electroencephalography (EEG) signals, by using a wearable EMG acquisition device. It combines joint motion state signals and musculoskeletal dynamics auxiliary information, performs preprocessing and feature extraction, and obtains the user's musculoskeletal functional state feature vector based on the Transformer multi-task deep neural network model. The individual physiological atlas dynamic modeling module collects users’ historical physiological data, constructs a user-personalized physiological atlas based on graph neural networks, and obtains the user’s current individual physiological atlas representation. The decision reasoning module outputs prediction results of individual orthopedic risk level, probability of musculoskeletal injury, and trend of degenerative lesions based on individual physiological atlas representation, combined with context awareness mechanism and supervised learning-based risk prediction model; The dynamic assessment and optimization module receives updated physiological signal data from the user in real time, dynamically corrects the health risk assessment results, and uses reinforcement learning algorithms to adaptively optimize the parameters of the risk prediction model. The risk protection module analyzes muscle fatigue evolution trends, joint stability changes, and neuromuscular control abnormalities by constructing a time-series anomaly detection model. When an abnormal risk is detected, it outputs orthopedic risk warning information and corresponding rehabilitation intervention suggestions.

2. The intelligent decision-making system for orthopedic risk based on multimodal bioelectrical signal fusion according to claim 1, characterized in that: The individual physiological atlas dynamic modeling module collects users' historical physiological data, constructs a user-personalized physiological atlas based on graph neural networks, and obtains the user's current individual physiological atlas representation. Specifically, it includes the following steps: Step S1: Graph structure construction, defining the graph structure , The nodes represent various data types, including skeletal structure nodes, neural function region nodes, muscle group nodes, and physiological function indicator nodes. The edges are represented, including muscle-skeletal mechanical coupling relationships, neuromuscular control pathways, inter-joint kinetic chain transmission relationships, and statistical correlation relationships, to construct the initial static physiological graph structure; Step S2: Node feature construction and fusion. The user's musculoskeletal functional state feature vector, the collected multimodal bioelectrical signal data, and the user's historical physiological data are fused to construct the node feature vector. Step S3: Feature dimensionality reduction. Linear discriminant analysis (LDA) is used to reduce the dimensionality of the node feature vectors to obtain the dimensionality-reduced node features. The KMeans++ algorithm is used for pre-clustering to generate K similar node clusters, and the feature centroid of each cluster is calculated. Step S4: Sector topology subgraph construction. A structured subgraph is constructed based on similar node clusters. The node corresponding to the feature centroid of each node cluster is used as the center node to construct a musculoskeletal coordinating sector topology subgraph. Within each node cluster, a similarity metric is constructed based on the feature distance between the node and the center node. A similarity threshold is determined based on the statistical distribution of distances within the node cluster. Nodes whose feature distance to the center node is less than the threshold and satisfy the local density constraint are selected as peripheral nodes. A weighted connection relationship is established between the center node and the peripheral nodes. A distance decay mechanism is introduced to modulate the edge weights to obtain an enhanced graph structure. Step S5: Pre-clustering enhances graph neural network modeling. Construct a graph neural network model that integrates local neighborhood aggregation and long-distance similar neighborhood aggregation. Perform local neighborhood feature aggregation based on the initial static physiological graph structure to obtain the local structural representation of the nodes. Perform long-distance similar node feature aggregation based on the musculoskeletal coordinating fan-shaped topological subgraph to obtain the cross-regional semantic association representation. Fuse the local structural representation and cross-regional semantic association representation of the nodes to complete multi-scale feature learning and obtain the updated node embedding representation. Step S6: Graph output, which integrates the updated node embedding representation and the enhanced graph structure to generate the user's current individual physiological graph representation.

3. The intelligent decision-making system for orthopedic risk based on multimodal bioelectrical signal fusion according to claim 2, characterized in that: Step S4 specifically includes the following steps: Step S41: Identify the musculoskeletal coordinating sector topology center node. For each similar node cluster, calculate its cluster centroid feature vector as the representative node of the cluster. Calculate the Euclidean distance between all nodes in the cluster and the centroid feature vector, and select the closest node as the musculoskeletal coordinating sector topology center node of the cluster. The formula used is as follows: ； in, Indicates the first The central node of the musculoskeletal co-operational sector topology selected from a cluster of similar nodes. Represents a node The dimensionality-reduced feature vectors Indicates the first The characteristic centroids of a cluster of similar nodes; Step S42: Subgraph node filtering. Within each similar node cluster, sort nodes according to their Euclidean distance to the musculoskeletal co-occurrence sector topology center node before filtering. The set of outer nodes is composed of nodes, represented as follows: ； in, Indicates the first The set of peripheral nodes selected from clusters of similar nodes Indicates candidate outer nodes, Indicates the first The set of all nodes contained in a cluster of similar nodes. Represents a node The feature vector representation after dimensionality reduction in step S3, Represents a node The feature vector representation after dimensionality reduction in step S3, This means taking the nearest neighbor. 1 node This indicates the threshold for the number of peripheral nodes; Step S43: Calculate the distance penalty factor. Define the distance penalty factor and calculate the edge weight for each connected edge using the following formula: ； in, express and The weight of the edges connecting them. Indicates the distance attenuation coefficient; Step S44: Musculoskeletal cooperative sector structure connection: Connect the central node with each node in the outer node set to construct a musculoskeletal cooperative sector topology subgraph. The set of connecting edges is represented as: ； in, Indicates the first The set of musculoskeletal cooperative fan-shaped structure connecting edges corresponding to clusters of similar nodes. Indicates the first The outer nodes selected from clusters of similar nodes Indicates the first The set of subgraph nodes corresponding to each cluster of similar nodes, including and All peripheral nodes; Step S45: Merge into the original graph structure. Merge the musculoskeletal co-operational sector topology subgraph into the initial version of the static graph structure from step S1 to obtain the enhanced graph structure.

4. The intelligent decision-making system for orthopedic risk based on multimodal bioelectrical signal fusion according to claim 3, characterized in that: Step S5 specifically includes the following steps: Step S51: Initialize the local neighborhood aggregation module. Initialize the graph attention network based on the multi-head graph attention mechanism, learn the attention weights between the dimensionality-reduced node feature vector and its one-hop neighbors, and use learnable attention coefficients to weight and aggregate the features of neighboring nodes to obtain the local neighborhood embedding representation of each node. Step S52: Construction of the long-distance similarity neighborhood aggregation module. Based on the musculoskeletal coordinating fan-shaped topological subgraph constructed in Step S4, for the N graphs corresponding to the K similar node clusters generated in Step S3, the central node of each subgraph is used as the aggregation core. A graph attention mechanism is introduced in each subgraph to perform weighted aggregation of the features of the peripheral nodes, thereby obtaining the long-distance similarity neighborhood embedding representation of each central node. The formula used is as follows: ； ； in, Indicates the central node The long-distance similar neighborhood embedding representation, Represents a non-linear activation function. Represents a node For nodes Attention weights This represents the learnable linear transformation weight matrix. It is a node The feature representation after dimensionality reduction, It is a node The feature representation after dimensionality reduction, Indicates the attenuation factor. Indicates the structural distance influence factor. It is to traverse the set of outer nodes. index variable, Represents a node The feature vector after dimensionality reduction in step S3, Represents a node With nodes Topological distance; Step S53: Node feature fusion. The local neighborhood embedding representation and the far-distance similar neighborhood embedding representation of each node are subjected to max pooling operation, and feature dimension alignment and fusion processing are performed to generate a unified embedding representation for each node. Step S54: Output update. The unified embedding representation of each node is enhanced with feature representation through residual connection mechanism and nonlinear transformation, and the Dropout mechanism is combined to suppress overfitting to obtain the updated node representation.

Images