Spine curvature abnormity early risk assessment method and system

Abstract

The invention relates to an early risk assessment method and system for abnormal spine curvature, and the method comprises the steps: carrying out the fusion of space-time sequence features based on obtained multi-modal data, and carrying out the three-dimensional biomechanical modeling, and obtaining a spine-pelvis-lower limb dynamic coupling network; carrying out weighted calculation on the key mechanical transmission path through a graph attention mechanism based on a dynamic coupling network to obtain a path strengthening feature vector; performing multi-modal risk feature decoupling and dynamic weighted aggregation based on the path strengthening feature vector to obtain a structured risk map; and based on the structured risk map, early risk assessment of spine curvature abnormity is carried out through adaptive risk threshold mapping and dynamic risk quantification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided diagnosis, in particular to a scoliosis early risk assessment method and system. BACKGROUND

[0002] Scoliosis (such as adolescent idiopathic scoliosis) is one of the most common deformities in the skeletal development period of adolescents, with a global incidence of about 2%-4%. Early screening and intervention are crucial to prevent the deformity from progressing to surgical indications (Cobb angle > 40°). However, traditional assessment methods have significant limitations: 1) Fragmentation risk of single modality assessment: On the one hand, it relies on Adam forward bending test and other surface observations, but the sensitivity is only 62%-75%, and it is easy to miss mild deformities (Cobb angle < 20°); On the other hand, full-spine X-ray is the gold standard, but the radiation dose (about 0.1-0.3 mSv / time) limits the follow-up frequency, and it cannot capture dynamic functional abnormalities; Finally, surface electromyography (sEMG) or three-dimensional gait analysis can reflect neuromuscular compensation, but lack quantitative correlation with structural deformities. 2) Lag risk of static assessment: Traditional methods mainly rely on "single snapshot" assessment, which cannot capture the dynamic risk of rapid progression of deformities during the growth peak of adolescence (Risser 0-2). 3) Lack of individualization of experience dependence: Risk thresholds are based on population statistics, without considering individual growth potential (such as bone age, Risser sign), muscle compensation capacity (such as trunk rotation angle ATR), etc., leading to over-medicalization (such as overuse of braces) or missed diagnosis. Therefore, in order to overcome the limitations of fragmentation, lag, and individualization of traditional assessment methods, it is necessary to develop a scoliosis early risk assessment method and system that integrates multi-modal data fusion, dynamic quantitative risk, and individualized assessment, to realize "structure-function-compensation" whole-chain quantitative risk and improve the sensitivity and accuracy of screening. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the deficiencies of the prior art, and to provide a scoliosis early risk assessment method and system.

[0004] The technical solution of the present application to solve the above technical problem is as follows: A scoliosis early risk assessment method, the method comprising:

[0005] S1, based on the obtained multi-modal data, performing spatio-temporal sequence feature fusion and three-dimensional biomechanical modeling to obtain a dynamic coupling network of spine-pelvis-lower extremity;

[0006] S2, based on the dynamic coupling network, performing weighted calculation of key mechanical transmission path through graph attention mechanism to obtain a path reinforcement feature vector;

[0007] S3, performing multi-modal risk feature decoupling and dynamic weighted aggregation based on the path reinforcement feature vector to obtain a structured risk graph;

[0008] S4, performing early risk assessment of spinal curvature abnormalities based on the structured risk graph through adaptive risk threshold mapping and dynamic risk quantification.

[0009] Further, in step S1, the multi-modal data includes multi-channel surface electromyography data collected based on a back electrode array, foot region pressure data collected based on a foot pressure sensor, and spinal motion trajectory data collected based on an inertial measurement unit.

[0010] Further, in step S1, the fusion of spatio-temporal sequence features and three-dimensional biomechanical modeling based on the obtained multi-modal data obtain a spine-pelvis-lower limb dynamic coupling network, including:

[0011] S11, performing data preprocessing and spatio-temporal alignment on the obtained multi-modal data to obtain an sEMG signal matrix after spatio-temporal calibration, a foot pressure distribution heat map, and filtered spinal motion trajectory;

[0012] S12, extracting muscle synergistic contraction parameters from the sEMG signal matrix, extracting gait asymmetry parameters based on the foot pressure distribution heat map, and extracting spinal motion harmonic components from the filtered spinal motion trajectory;

[0013] S13, based on the muscle synergistic contraction parameters, gait asymmetry parameters, and spinal motion harmonic components, calculating a cross-modal spatio-temporal feature tensor and an anatomy-motion joint graph through a multi-scale spatio-temporal graph convolution network;

[0014] S14, based on the cross-modal spatio-temporal feature tensor and the anatomy-motion joint graph, performing three-dimensional biomechanical modeling and dynamic coupling network construction to obtain a spine-pelvis-lower limb dynamic coupling network.

[0015] Further, in step S12, the gait asymmetry parameters include a gait cycle asymmetry index, a bilateral lower limb load difference coefficient, and time-frequency features of a pressure center trajectory, wherein:

[0016] The gait cycle asymmetry index is calculated by the following formula:

[0017] ;

[0018] wherein N represents the total number of gait cycles, represents the left-right foot ground contact phase difference, represents the complete gait cycle phase, represents the left foot pressure integral, represents the right foot pressure integral;

[0019] The double lower limb load difference coefficient is calculated by the following formula:

[0020]

[0021] wherein T represents the length of the time window, represents the left foot pressure peak value, represents the right foot pressure peak value.

[0022] Further, in step S13, the cross-modal spatiotemporal feature tensor and the anatomic-motor joint atlas are calculated based on the muscle synergistic contraction parameter, the gait asymmetry parameter, and the spinal motion harmonic component through a multi-scale spatiotemporal graph convolution network, comprising:

[0023] S131, different scale spatiotemporal features are extracted from the muscle synergistic contraction parameter, the gait asymmetry parameter, and the spinal motion harmonic component through a multi-scale spatiotemporal graph convolution layer;

[0024] S132, the extracted different modal features are subjected to feature splicing and weighted summation based on a graph attention mechanism to obtain a fused cross-modal spatiotemporal feature tensor;

[0025] S133, the graph structure information of the anatomic-motor joint atlas is determined based on the anatomic structure-motor function correlation and the multi-modal data spatiotemporal dependency;

[0026] S134, the graph structure information is input into the multi-scale spatiotemporal graph convolution network, and the graph structure is optimized based on the graph attention mechanism in the network in combination with anatomical knowledge to obtain the anatomic-motor joint atlas.

[0027] Further, in step S2, the path reinforcement feature vector is obtained by weighted calculation of the key mechanical transmission path based on the dynamic coupling network through a graph attention mechanism, comprising:

[0028] S21, the dynamic coupling network is subjected to spatiotemporal feature encoding based on a graph attention mechanism to obtain an attention- encoded dynamic coupling network;

[0029] S22, the key mechanical transmission path is extracted based on the attention- encoded dynamic coupling network to obtain a key mechanical transmission path set, wherein each path contains a node feature sequence, an edge feature sequence, and a time sequence feature sequence;

[0030] S23, multi-path feature fusion is performed based on the key mechanical transmission path set to obtain a path reinforcement feature vector. ​​

[0031] Further, in step S22, the dynamic coupling network based on the attention encoding extracts the key mechanical transmission path to obtain a key mechanical transmission path set, including:

[0032] S221, the node and edge in the dynamic coupling network are calculated by using the graph attention network to generate the node-level attention coefficient and the edge-level attention coefficient;

[0033] S222, based on the node-level attention coefficient and the edge-level attention coefficient, a depth-first search algorithm is used to identify all potential mechanical transmission paths, and each mechanical transmission path is sorted according to the cumulative attention weight;

[0034] S223, combined with the anatomical prior knowledge, the sorted mechanical transmission path is screened and optimized to remove the path that does not conform to the biomechanical principle, and the key mechanical transmission path is reserved to form the key mechanical transmission path set.

[0035] Further, in step S3, the multi-modal risk feature decoupling and dynamic weighted aggregation based on the path reinforcement feature vector are performed to obtain a structured risk map, including:

[0036] S31, based on the hierarchical variational autoencoder and the dynamic graph attention joint optimization mechanism, the path reinforcement feature vector is decoupled into a structured multi-modal risk feature;

[0037] S32, according to the decoupling quality of the feature, the decoupled multi-modal risk feature is dynamically weighted to obtain a weighted multi-modal risk feature;

[0038] S33, based on the weighted multi-modal risk feature, a graph structure learning mechanism is used to obtain a structured risk map.

[0039] Further, in step S4, based on the structured risk map, an adaptive risk threshold mapping and a dynamic risk quantification are performed for early risk assessment of spinal curvature abnormalities, including:

[0040] S41, the key risk nodes related to the spinal curvature and their associated edge features are extracted from the structured risk map to construct a spinal special risk subgraph;

[0041] S42, based on the individual characteristics and historical risk trajectory of the patient, a personalized risk threshold baseline is generated by a lightweight meta-learning model;

[0042] S43, based on the spinal special risk subgraph and the personalized risk threshold baseline, a dynamic risk value and a risk fluctuation entropy are calculated by a joint quantification model of dynamic time warping and double-flow gated recurrent unit;

[0043] S44, based on the dynamic risk value and risk fluctuation entropy, generating an early risk assessment report including risk classification, trend prediction and personalized intervention suggestions through a multi-modal fusion decision engine.

[0044] In a second aspect, the application discloses an early risk assessment system for abnormal spinal curvature, which comprises a spatiotemporal feature fusion and three-dimensional biomechanical modeling module, a key mechanical transmission path weighting calculation module, a multi-modal risk feature decoupling and dynamic weighting aggregation module, and an early risk assessment module for abnormal spinal curvature, wherein:

[0045] The spatiotemporal feature fusion and three-dimensional biomechanical modeling module is configured to perform spatiotemporal sequence feature fusion and three-dimensional biomechanical modeling based on the acquired multi-modal data, and obtain a dynamic coupling network of the spine-pelvis-lower limb.

[0046] The key mechanical transmission path weighting calculation module is configured to perform weighting calculation of the key mechanical transmission path based on the dynamic coupling network through a graph attention mechanism, and obtain a path reinforcement feature vector.

[0047] The multi-modal risk feature decoupling and dynamic weighting aggregation module is configured to perform multi-modal risk feature decoupling and dynamic weighting aggregation based on the path reinforcement feature vector, and obtain a structured risk graph.

[0048] The early risk assessment module for abnormal spinal curvature is configured to perform early risk assessment of abnormal spinal curvature based on the structured risk graph through adaptive risk threshold mapping and dynamic risk quantification.

[0049] The application has the following beneficial effects: through multi-modal data fusion, key mechanical path analysis, risk feature decoupling and dynamic aggregation, and adaptive risk assessment, the accuracy and individualization level of early screening for abnormal spinal curvature are significantly improved, and the whole-chain risk quantification from structural abnormalities to functional compensation is realized, thereby providing a scientific basis for early intervention. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A flowchart of an early risk assessment method for abnormal spinal curvature provided by the application is shown in the figure.

[0051] Figure 2 A structural diagram of an early risk assessment system for abnormal spinal curvature provided by the application is shown in the figure. DETAILED DESCRIPTION

[0052] The principles and features of the application are described below in combination with the drawings, and the examples are only used to explain the application and not to limit the scope of the application.

[0053] As Figure 1As shown, a method for early risk assessment of spinal curvature abnormalities, the method comprises:

[0054] Step S1, based on the obtained multi-modal data, the fusion of spatio-temporal sequence features and three-dimensional biomechanical modeling are performed to obtain a dynamic coupling network of spine-pelvis-lower limb.

[0055] Step S2, based on the dynamic coupling network, the weighted calculation of key mechanical transmission path is performed through graph attention mechanism to obtain a path reinforcement feature vector.

[0056] Step S3, based on the path reinforcement feature vector, multi-modal risk feature decoupling and dynamic weighted aggregation are performed to obtain a structured risk map.

[0057] Step S4, based on the structured risk map, through adaptive risk threshold mapping and dynamic risk quantification, early risk assessment of spinal curvature abnormalities is performed.

[0058] As can be seen from the above, the method for early risk assessment of spinal curvature abnormalities disclosed in the present application significantly improves the accuracy and individualization level of early screening of spinal curvature abnormalities through multi-modal data fusion, key mechanical path analysis, risk feature decoupling and dynamic aggregation, and adaptive risk assessment, and realizes the whole-chain risk quantification from structural abnormalities to functional compensation, providing a scientific basis for early intervention.

[0059] In one embodiment, in step S1, the multi-modal data includes multi-channel surface electromyography data collected based on a back electrode array, foot region pressure data collected based on a foot pressure sensor, and spinal motion trajectory data collected based on an inertial measurement unit.

[0060] In one embodiment, in step S1, the fusion of spatio-temporal sequence features and three-dimensional biomechanical modeling based on the obtained multi-modal data to obtain a dynamic coupling network of spine-pelvis-lower limb comprises:

[0061] Step S11, data preprocessing and spatio-temporal alignment are performed on the obtained multi-modal data to obtain a spatio-temporally calibrated sEMG signal matrix, a foot pressure distribution thermogram, and a filtered spinal motion trajectory.

[0062] Specifically, the data preprocessing here includes filtering and denoising, rectification and smoothing, signal segmentation, normalization processing, thermogram generation, trajectory smoothing, and coordinate system unification, and the present application is not limited to the specific embodiments. Spatio-temporal alignment, i.e., through time synchronization and space registration, ensures the consistency of multi-modal data in time and space, providing a basis for subsequent multi-modal data fusion and three-dimensional biomechanical modeling.

[0063] Step S12, muscle synergies parameters are extracted from the sEMG signal matrix, gait asymmetry parameters are extracted based on the plantar pressure distribution thermogram, and spinal motion harmonic components are extracted from the filtered spinal motion trajectory.

[0064] Specifically, muscle synergies parameters are extracted from the sEMG signal matrix by a non-negative matrix factorization algorithm. Gait asymmetry parameters are quantified by synthesizing gait cycle asymmetry index, bilateral lower limb load difference coefficient, and center of pressure trajectory time-frequency features. The gait cycle asymmetry index combines phase difference and pressure integral difference to evaluate the timing shift and load difference of the gait cycle. The bilateral lower limb load difference coefficient reflects the static and dynamic asymmetry of the bilateral lower limb load by pressure peak difference and load change rate. The COP time-frequency feature is evaluated by wavelet packet decomposition to assess the balance control ability. The spinal motion harmonic components are extracted by the Hilbert-Huang transform algorithm. The transform specifically adaptively decomposes the spinal motion trajectory into multiple intrinsic mode functions by empirical mode decomposition, then calculates the instantaneous frequency and amplitude by Hilbert transform, finally extracts the main harmonic component and quantifies the periodicity of the spinal motion.

[0065] Step S13, based on the muscle synergies parameters, gait asymmetry parameters, and spinal motion harmonic components, a multi-scale spatio-temporal graph convolution network is used to calculate a cross-modal spatio-temporal feature tensor and an anatomic-motion joint atlas.

[0066] Specifically, the multi-scale spatio-temporal graph convolution network adopts a hierarchical architecture, captures the spatio-temporal features of muscle synergies, gait asymmetry, and spinal motion through multi-scale convolution kernels, and realizes cross-modal feature fusion and graph structure optimization through graph attention mechanism. The input layer of the network is responsible for integrating muscle synergies parameters, gait asymmetry parameters, and spinal motion harmonic components. After extracting local and global features through multi-scale spatio-temporal graph convolution layers, the network dynamically allocates modal weights through self-attention mechanism to perform feature concatenation and weighted summation. Further, the present application also combines anatomical knowledge (such as Fryette mechanism, Gracovetsky spinal engine theory, joint kinematic coupling rules) to constrain and optimize the graph structure, and finally generates a cross-modal feature tensor and a hierarchical anatomic-motion joint atlas containing spatio-temporal dependence and anatomic-functional association.

[0067] Step S14, based on the cross-modal spatio-temporal feature tensor and the anatomic-motion joint atlas, a three-dimensional biomechanical modeling and dynamic coupling network is constructed to obtain a spine-pelvis-lower limb dynamic coupling network.

[0068] Specifically, combined with finite element analysis and multi-body dynamics, the trans-modal feature tensor is used to drive muscle force simulation and joint constraint definition, and the anatomic-motion combined atlas is fused to initialize the bone, ligament and muscle path parameters, and finally a three-dimensional biomechanical model is constructed through simulation simulation. Further need to explain is that the above-mentioned dynamic coupling network construction is mainly based on the nodes (such as vertebrae, pelvis, femur) and dynamic edges (such as muscle activation, joint contact force) predefined by the atlas, and the edge weight is updated by the trans-modal feature tensor to reflect the real-time coupling relationship; finally, Fryette mechanism, Gracovetsky spine engine theory and muscle fascia chain transmission model are integrated to form a dynamic coupling network of spine-pelvis-lower limb containing neuromuscular control and biomechanical constraints.

[0069] In one of the embodiments, in step S12, the gait asymmetry parameters include gait cycle asymmetry index, bilateral lower limb load difference coefficient, and time-frequency features of pressure center trajectory, wherein:

[0070] The gait cycle asymmetry index is calculated by the following formula:

[0071] ;

[0072] Wherein, N represents the total number of gait cycles, represents the left-right foot ground contact phase difference, represents the complete gait cycle phase, represents the left foot pressure integral, represents the right foot pressure integral.

[0073] The bilateral lower limb load difference coefficient is calculated by the following formula:

[0074] ;

[0075] Wherein, T represents the length of the time window, , represents the left foot pressure peak, represents the right foot pressure peak.

[0076] In one of the embodiments, in step S13, based on the muscle synergistic contraction parameters, gait asymmetry parameters, and spinal motion harmonic components, the trans-modal spatiotemporal feature tensor and the anatomic-motion combined atlas are calculated by the multi-scale spatiotemporal graph convolution network, including:

[0077] In step S131, different scales of spatiotemporal features are extracted from muscle synergistic contraction parameters, gait asymmetry parameters and spinal motion harmonic components by multi-scale spatiotemporal graph convolution layer.

[0078] Specifically, the multi-scale spatio-temporal graph convolution layer adopts three sets of parallel convolution kernels (3x3, 5x5, 7x7) to capture local muscle coordination details, periodic gait asymmetry patterns, and macroscopic spinal motion trends, respectively, and finally outputs a multi-scale spatio-temporal feature tensor containing spatio-temporal dependence and anatomical-functional correlation, including muscle coordination contraction features, gait asymmetry features, and spinal motion features.

[0079] In step S132, the extracted different modal features are subjected to feature splicing and weighted summation based on a graph attention mechanism to obtain a fused cross-modal spatio-temporal feature tensor.

[0080] Specifically, the present application dynamically calculates the weights of each modal feature through a self-attention mechanism, and after assigning each weight to each modal feature, the corresponding weighted feature is obtained. Then, splicing each weighted feature can obtain a cross-modal spatio-temporal feature tensor.

[0081] In step S133, the graph structure information of the anatomical-movement joint atlas is determined based on anatomical structure-movement function correlation and multi-modal data spatio-temporal dependence.

[0082] Specifically, the present application will construct an initial graph structure based on an anatomical knowledge graph and a movement function graph, combined with anatomical structure-movement function correlation. Then, the spatio-temporal dependence of the multi-modal data is fused into the initial graph structure through a conditional random field, and the edge weights are dynamically optimized based on the parameter learning and potential function optimization mechanism of the conditional random field, and combined with the penalty term of anatomical constraint, to finally output the graph structure information of the anatomical-movement joint atlas containing anatomical-functional correlation and spatio-temporal dependence.

[0083] In step S134, the graph structure information is input into the multi-scale spatio-temporal graph convolution network, and the graph structure is optimized based on the graph attention mechanism in the network and combined with anatomical knowledge to obtain the anatomical-movement joint atlas.

[0084] Specifically, the present application inputs the graph structure information into the multi-scale spatio-temporal graph convolution network, and captures local anatomical details, periodic motion patterns, and macroscopic biomechanical trends through three sets of parallel convolution kernels (3x3, 5x5, 7x7). At the same time, the edge weights are dynamically adjusted using a graph attention network, and the graph structure is iteratively optimized combined with the penalty term of anatomical constraint. Finally, the multi-scale features are fused through a self-attention mechanism to output the anatomical-movement joint atlas containing fine anatomical structure, dynamic movement function, and spatio-temporal dependence.

[0085] In one embodiment, in step S2, the dynamic coupling network based on the graph attention mechanism is used to calculate the weighted calculation of the key mechanical transmission path to obtain a path enhancement feature vector, including:

[0086] Step S21, based on the graph attention mechanism, the dynamic coupling network is encoded in space-time features to obtain the attention encoded dynamic coupling network.

[0087] Specifically, the node features (such as bone position / speed) of the dynamic coupling network, the edge features (such as muscle force, joint stiffness), and the time sequence information are taken as input data, and the node features are mapped to the time attention space and the spatial attention space through the learnable parameter matrix (W t , W s ) respectively. Then, based on the multi-modal feature fusion and biomechanical constraints, the time dimension attention coefficient α t and the spatial dimension attention coefficient α s are calculated by the following formula:

[0088] α t = softmax(LeakyReLU(a t ^T [W t h i W t h j PE(t)]+ B t (θ joint )));

[0089] α s = softmax(LeakyReLU(a s ^T [W s h i W s h j C anatomy ]+ B s (F muscle )));

[0090] Where PE(t) represents the time sequence phase encoding, T represents the gait cycle time length, h i and h j represent the feature vectors of node i and node j, a t , a s represent the learnable parameter vector, C anatomy represents the anatomical constraint matrix, B t (θ joint )= W bθ θ joint , B s (F muscle )= W bF Fmuscle , θ joint denotes the joint angle vector, F muscle denotes the muscle force vector, W bθ and W bF denote the learnable weight matrix.

[0091] Based on the above formula, it needs to be explained that the timing phase encoding PE(t) in the formula maps the timing information in the gait cycle to a high-dimensional space, enhancing the model's perception ability of periodic motion patterns. The introduction of the anatomical constraint matrix C anatomy ensures the physiological reasonableness of the calculation results and avoids the emergence of non-physiological mechanical transmission paths. In addition, the design of the dynamic bias term B t (θ joint ), B s (F muscle ) enables the model to adapt to the mechanical transmission characteristics under different joint angles and muscle force conditions, improving the model's generalization ability.

[0092] Finally, the time dimension attention coefficient α t and the spatial dimension attention coefficient α s are fused by element-wise multiplication, that is, α t α s , to obtain the fused space-time attention coefficient α. Through the fused space-time attention coefficient α, feature weighting aggregation and residual update processing are performed to obtain the attention encoded dynamic coupling network. For each node i, the application will take the fused attention coefficient α as the weight, and based on the features of the neighbor node j, weighted summation is performed (for reference, the formula is as follows: , is a learnable feature transformation matrix), to obtain the aggregated features . The aggregated features are connected in residual connection with the original node features (residual connection design can alleviate gradient disappearance and preserve original dynamic features), and the LeakyReLU activation function is used for non-linear expression ability enhancement (for reference, the formula is as follows: ), to realize the update of the node i features.

[0093] Further, the application will also update the original edge weight based on the fused attention coefficient α, which can be specifically referred to in the following formula: (wherein, denotes the anatomical prior weight), wherein the anatomical prior weight The introduction of the attention mechanism further makes the model more inclined to select a mechanical transmission path that conforms to anatomical rules during optimization. Finally, based on the updated node features and edge weights a new network is constructed, i.e. a dynamic coupling network that considers both the periodicity of the gait in the time dimension and the mechanical transmission pattern in the spatial dimension, which provides a high-quality feature representation for subsequent key mechanical transmission path search.

[0094] Step S22, extracting a key mechanical transmission path based on the attention- encoded dynamic coupling network to obtain a key mechanical transmission path set, wherein each path includes a node feature sequence, an edge feature sequence, and a time sequence feature sequence.

[0095] Step S23, performing multi-path feature fusion based on the key mechanical transmission path set to obtain a path reinforcement feature vector.

[0096] Specifically, the present application fuses the feature of each path in the key mechanical transmission path set through attention weighted summation and anatomical constraint guided graph aggregation operation. First, the attention mechanism is used to calculate the weight of each path, which reflects the importance of the path in the overall mechanical transmission. Second, the anatomical prior knowledge is combined to guide the constraint of the path feature (i.e. the path feature is weighted and modified by the anatomical constraint matrix, and the feature value is clipped by introducing the physiological range limit), to ensure that the fused feature conforms to the biomechanical principle. Finally, through the graph aggregation operation, the weighted path feature is integrated into a unified path reinforcement feature vector, which not only retains the key information of each path, but also reflects the overall mechanical transmission pattern.

[0097] In one embodiment, in step S22, the key mechanical transmission path is extracted based on the attention-encoded dynamic coupling network to obtain a key mechanical transmission path set, which includes:

[0098] Step S221, using a graph attention network to calculate the attention weight of the nodes and edges in the dynamic coupling network to generate node-level attention coefficients and edge-level attention coefficients.

[0099] Specifically, the graph attention network calculates the attention weight of the neighbor nodes (i.e. node-level attention coefficients) for each node and the attention weight of the transmitted force (i.e. edge-level attention coefficients) for each edge through the self-attention mechanism.

[0100] In step S222, based on the node-level attention coefficient and the edge-level attention coefficient, a depth-first search algorithm is used to identify all potential mechanical transmission paths, and each mechanical transmission path is sorted according to the cumulative attention weight.

[0101] It should be noted that the cumulative attention weight refers to the product of all edge-level attention coefficients on the path, and its quantification reflects the comprehensive importance of the path in the mechanical transmission network.

[0102] In step S223 , the sorted mechanical transmission pathways are screened and optimized based on prior anatomical knowledge to remove pathways that do not conform to biomechanical principles, retain key mechanical transmission pathways, and form a set of key mechanical transmission pathways.

[0103] Specifically, this application combines anatomical prior knowledge to perform a two-stage screening of the sorted paths. First, non-physiological paths (such as knee hyperextension paths) are filtered out using an anatomical constraint matrix. Second, the validity of each remaining path is evaluated using a path validity evaluation function. The optimal path is screened based on a validity threshold comparison and used as the key mechanical transmission path to ensure that the final set conforms to biomechanical principles and covers the main mechanical transmission modes.

[0104] In one embodiment, the path validity evaluation function is as shown in the following formula:

[0105] ;

[0106] in, represents each edge on the path, Represents an edge The actual mechanical transmission amount on Represents an edge The maximum mechanical transfer capacity on Represents an edge Energy consumption on Represents an edge The maximum energy consumption allowed, 、 Represents the weight coefficient.

[0107] It should be noted that the evaluation function comprehensively considers the mechanical transmission efficiency and energy consumption through weighted summation, so that the evaluation results can quantitatively reflect the comprehensive performance of the path in mechanical transmission, thereby guiding the algorithm to give priority to efficient transmission paths.

[0108] In one embodiment, in step S3, the multimodal risk feature decoupling and dynamic weighted aggregation based on the path reinforcement feature vector to obtain a structured risk map includes:

[0109] Step S31, based on the hierarchical variational autoencoder and dynamic graph attention joint optimization mechanism, the path reinforcement feature vector is decoupled into structured multi-modal risk features.

[0110] Specifically, the total correlation penalty and the path-aware sparse constraint are combined to construct a cross-modal contrastive loss function. Based on this function, the modal decoupling and structured risk representation of the path reinforcement feature vector are performed through the joint optimization framework of hierarchical variational autoencoder and dynamic graph attention adversarial training. The total correlation penalty minimizes the total correlation of latent variables to enforce modal feature statistical independence. The path-aware sparse constraint dynamically adjusts the L1 regularization strength based on the path length, and applies stronger sparsity penalty to long path features to effectively suppress redundant information propagation. The cross-modal contrastive loss function constructs an adaptive weight adjustment mechanism by combining the total correlation penalty and the path-aware sparse constraint, that is, when calculating the inter-modal similarity, higher attention is given to feature samples with high decoupling quality (i.e., low total correlation penalty) and long path length. The calculation formula is as follows:

[0111] ;

[0112] wherein, represents the total correlation of the modal features of the i-th sample, represents the maximum total correlation within the batch, represents a decoupling quality sensitive coefficient. It can be understood that when is small (i.e., high decoupling quality), is close to 1, and the weight enhancement effect is amplified by ; otherwise, when is close to , the adjustment of the weight will tend to 1 (i.e., no enhancement processing is performed). represents the path length of the i-th sample, represents a path length sensitive coefficient, and I(*) represents an indicator function that outputs 1 when the condition in the parentheses is met, and otherwise outputs 0. It can be understood that when , the application applies an additional weight to long path features through the ( ) term to suppress redundant information propagation and improve feature quality.

[0113] Step S32, according to the decoupling quality of the features, the multi-modal risk features after decoupling are dynamically weighted to obtain weighted multi-modal risk features.

[0114] Specifically, the present application dynamically weights the decoupled multi-modal risk features according to the decoupling quality of the features. In the process, the importance weight of each feature is learned through an attention mechanism, and the decoupling quality value of the feature (such as the reciprocal of the total correlation penalty TC value, where a lower TC value indicates higher decoupling quality) is mapped to the corresponding weight adjustment coefficient according to the decoupling quality through a pre-set decoupling quality mapping table. The coefficient is multiplied by the original weight in the attention mechanism to dynamically adjust the weight, highlighting features with high decoupling quality and suppressing features with low decoupling quality.

[0115] Step S33, based on the weighted multi-modal risk features, a structured risk graph is obtained through a graph structure learning mechanism.

[0116] Specifically, the present application uses the weighted multi-modal risk features as graph nodes and constructs an initial graph structure through a feature similarity measurement edge connection strategy. Subsequently, a graph optimization algorithm (such as a graph attention network or a graph convolution network) is used to combine a local structure preservation loss (which preserves the local mode of risk propagation path) and a global alignment loss (which ensures cross-modal risk semantic consistency) to jointly optimize the edge weight and node representation of the graph structure. Finally, through iterative optimization, a structured risk graph that can accurately depict the complex associations between risk features is obtained.

[0117] In one embodiment, in step S4, based on the structured risk graph, early risk assessment of spinal curvature abnormalities is performed through adaptive risk threshold mapping and dynamic risk quantification, including:

[0118] Step S41, from the structured risk graph, extract key risk nodes related to spinal curvature and their associated edge features, and construct a spinal-specific risk subgraph.

[0119] Specifically, based on the extracted key risk nodes and associated edge features, the present application constructs a risk subgraph focused on spinal curvature abnormalities. It should be noted that the subgraph is a directed and weighted graph, where the nodes in the graph represent key risk factors, the edges in the graph represent the associations between these factors, and the weights of the edges reflect the strength or direction of the associations.

[0120] Step S42, based on the individual characteristics and historical risk trajectory of the patient, generate a personalized risk threshold baseline through a lightweight meta-learning model.

[0121] Specifically, in order to generate a personalized risk threshold baseline, the present application designs a lightweight meta-learning model. The model takes into full consideration the individual characteristics of the patient, such as age, gender, physiological indicators, etc., as well as the historical risk trajectory of the patient. Through a meta-learning algorithm, the model can quickly learn and adapt to new tasks from a small number of samples, thereby generating a personalized risk threshold baseline for each patient.

[0122] Step S43, based on the spine-specific risk subgraph and the personalized risk threshold baseline, a dynamic risk value and a risk fluctuation entropy are calculated by a joint quantization model of dynamic time warping and double-flow gated recurrent unit.

[0123] Specifically, the dynamic time warping algorithm will be used to align the multi-time phase spine-specific risk subgraph in time to eliminate the noise caused by irregular examination frequency, and then the time-aligned spatiotemporal graph sequence will be input into the spatiotemporal graph convolution network to extract the spatiotemporal joint features. Further, the joint quantization model of double-flow gated recurrent unit will be used to capture the changing trajectory of the risk value over time based on the spatiotemporal joint features, and a dynamic risk value will be generated in combination with the preset threshold baseline. Finally, the fluctuation mode of the risk value sequence is encoded by LSTM, and the risk fluctuation entropy is calculated by the following steps:

[0124] 1) Extract the risk value sequence {h1, h2,..., h T} output by the LSTM hidden layer;

[0125] 2) Based on the risk value sequence {h1, h2,..., h T}, the mean and the standard deviation are calculated by maximum likelihood estimation;

[0126] 3) Based on the mean and the standard deviation , the risk fluctuation entropy is calculated by the following formula:

[0127] ;

[0128] Wherein, denotes the standardized deviation of the risk value from the mean , and denotes the time decay weight, and m denotes the pattern matching length. Based on the formula, the mean is taken as the "baseline" of the risk value, and quantifies the degree of deviation of each risk value from the baseline. By explicitly modeling the baseline and fluctuation range of the risk value through the mean and the standard deviation, and by combining the time weighting and the dynamic sensitivity adjustment, the calculation of the risk fluctuation entropy has individualization, time sensitivity and clinical interpretability.

[0129] Step S44, based on the dynamic risk value and the risk fluctuation entropy, an early risk assessment report containing risk classification, trend prediction and personalized intervention suggestions is generated by a multi-modal fusion decision engine.

[0130] Specifically, the present application will use a multi-modal fusion decision engine to take the dynamic risk value as the core risk quantification indicator, combine the risk uncertainty represented by the risk fluctuation entropy, and train the historical risk data and the corresponding risk level labels through a supervised learning framework to form a corresponding risk grading model. It should be noted that the risk grading model can automatically divide the risk into different levels, such as low, medium, and high, to help clinicians quickly understand the risk status of patients. At the same time, in order to predict the future trend of the risk value, the present application also uses a time series analysis algorithm to capture the change pattern of the risk value over time and generate a predicted value for a future period of time to provide prospective risk information for clinicians. Further, according to the risk grading and trend prediction results, the present application will also generate personalized intervention recommendations in combination with the individual characteristics (such as age, gender, physiological indicators) of the patient and the clinical guidelines. These recommendations include specific intervention measures (such as physical therapy, brace therapy, surgical treatment) and implementation timing, aiming to help clinicians develop targeted treatment plans. Finally, the present application will integrate risk grading, trend prediction, and personalized intervention recommendations to generate a structured early risk assessment report, with charts and graphs in the report illustrating risk assessment recommendations, so that clinicians can quickly understand the risk status and make decisions.

[0131] Please refer to Figure 2 The present application discloses a system for early risk assessment of abnormal spinal curvature, which comprises a spatiotemporal feature fusion and three-dimensional biomechanical modeling module, a key mechanical transmission path weighting calculation module, a multi-modal risk feature decoupling and dynamic weighting aggregation module, and an early risk assessment module for abnormal spinal curvature.

[0132] The spatiotemporal feature fusion and three-dimensional biomechanical modeling module is used to fuse spatiotemporal sequence features and perform three-dimensional biomechanical modeling based on the obtained multi-modal data, to obtain a dynamic coupling network of the spine-pelvis-lower limb.

[0133] The key mechanical transmission path weighting calculation module is used to perform weighting calculation of the key mechanical transmission path based on the dynamic coupling network through a graph attention mechanism, to obtain a path reinforcement feature vector.

[0134] The multi-modal risk feature decoupling and dynamic weighting aggregation module is used to perform multi-modal risk feature decoupling and dynamic weighting aggregation based on the path reinforcement feature vector, to obtain a structured risk atlas.

[0135] The early risk assessment module for abnormal spinal curvature is used to perform early risk assessment of abnormal spinal curvature based on the structured risk atlas through adaptive risk threshold mapping and dynamic risk quantification.

[0136] In one of the embodiments, the modules are further configured to implement a method for early risk assessment of abnormal spinal curvature as described in any of the preceding method embodiments, which are not limited herein.

[0137] As can be seen from the above, the early risk assessment system for abnormal spinal curvature disclosed by the present application significantly improves the accuracy and individualization level of early screening of abnormal spinal curvature through multi-modal data fusion, key mechanical path analysis, risk feature decoupling and dynamic aggregation, and adaptive risk assessment, and realizes the whole-chain risk quantification from structural abnormalities to functional compensation, providing a scientific basis for early intervention.

[0138] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for early risk assessment of spinal curvature abnormality, characterized in that: The method includes: S1. Based on the acquired multimodal data, the spatiotemporal sequence features are integrated and three-dimensional biomechanical modeling is performed to obtain a dynamic coupling network of the spine, pelvis and lower limbs. S2. Based on the dynamic coupling network, a graph attention mechanism is used to perform weighted calculation on key mechanical transmission paths to obtain a path reinforcement feature vector. S3. Decoupling and dynamically weighted aggregation of multimodal risk features are performed based on the path-enhanced feature vector to obtain a structured risk map; S4. Based on the structured risk map, early risk assessment of spinal curvature abnormalities is performed through adaptive risk threshold mapping and dynamic risk quantification.

2. The method according to claim 1, characterized in that In step S1, the multimodal data includes multi-channel surface electromyography data collected based on the back electrode array, pressure data of various areas of the sole of the foot collected based on the sole pressure sensor, and spinal motion trajectory data collected based on the inertial measurement unit.

3. The method according to claim 2, characterized in that In step S1, the spatiotemporal sequence feature fusion and three-dimensional biomechanical modeling based on the acquired multimodal data are performed to obtain a spine-pelvis-lower limb dynamic coupling network, including: S11, performing data preprocessing and spatiotemporal alignment on the acquired multimodal data to obtain a spatiotemporally calibrated sEMG signal matrix, a plantar pressure distribution heat map, and a filtered spinal motion trajectory; S12, extracting muscle co-contraction parameters from the sEMG signal matrix, extracting gait asymmetry parameters based on the plantar pressure distribution thermogram, and extracting spinal motion harmonic components from the filtered spinal motion trajectory; S13. Based on the muscle co-contraction parameters, gait asymmetry parameters, and spinal motion harmonic components, a cross-modal spatiotemporal feature tensor and an anatomical-motion joint atlas are calculated using a multi-scale spatiotemporal graph convolutional network; S14. Based on the cross-modal spatiotemporal feature tensor and the anatomical-motion joint atlas, three-dimensional biomechanical modeling and dynamic coupling network construction are performed to obtain a spine-pelvis-lower limb dynamic coupling network.

4. The method according to claim 3, characterized in that In step S12, the gait asymmetry parameters include the gait cycle asymmetry index, the load difference coefficient of both lower limbs, and the time-frequency characteristics of the pressure center trajectory, wherein: The gait cycle asymmetry index is calculated by the following formula: ； Where N represents the total number of gait cycles, Indicates the phase difference between the left and right feet touching the ground. represents the phase of the complete gait cycle, represents the left foot pressure integral, represents the right foot pressure integral; The lower limb load difference coefficient is calculated by the following formula: ； Where T represents the time window length, 、, Indicates the peak pressure of the left foot, Indicates peak right foot pressure.

5. The method according to claim 3, characterized in that In step S13, based on the muscle co-contraction parameters, gait asymmetry parameters, and spinal motion harmonic components, a cross-modal spatiotemporal feature tensor and an anatomical-motion joint atlas are calculated by a multi-scale spatiotemporal graph convolutional network, including: S131. Through the multi-scale spatiotemporal graph convolution layer, spatiotemporal features of different scales are extracted from muscle co-contraction parameters, gait asymmetry parameters and spinal motion harmonic components; S132. Perform feature concatenation and weighted summation based on the graph attention mechanism on the extracted features of different modalities to obtain a fused cross-modal spatiotemporal feature tensor; S133, determining graph structure information of the anatomical-motor joint atlas based on the anatomical structure-motor function association and the spatiotemporal dependency of the multimodal data; S134. Input the graph structure information into a multi-scale spatiotemporal graph convolutional network, optimize the graph structure based on the graph attention mechanism in the network and in combination with anatomical knowledge to obtain an anatomical-motion joint atlas.

6. The method according to claim 1, characterized in that In step S2, the weighted calculation of the key mechanical transmission paths based on the dynamic coupling network is performed through the graph attention mechanism to obtain the path reinforcement feature vector, including: S21. Encoding the spatiotemporal features of the dynamic coupling network based on a graph attention mechanism to obtain an attention-encoded dynamic coupling network; S22. Extracting key mechanical transmission paths based on the attention-encoded dynamic coupling network to obtain a set of key mechanical transmission paths, wherein each path includes a node feature sequence, an edge feature sequence, and a time series feature sequence; S23. Perform multi-path feature fusion based on the key mechanical transfer path set to obtain a path enhancement feature vector.

7. The method according to claim 6, characterized in that In step S22, extracting key mechanical transmission paths based on the attention-encoded dynamic coupling network to obtain a set of key mechanical transmission paths includes: S221. Calculate attention weights for nodes and edges in the dynamic coupling network using a graph attention network to generate node-level attention coefficients and edge-level attention coefficients. S222. Based on the node-level attention coefficient and the edge-level attention coefficient, a depth-first search algorithm is used to identify all potential mechanical transmission paths, and each mechanical transmission path is sorted according to the cumulative attention weight; S223. Combined with anatomical prior knowledge, the sorted mechanical transmission pathways are screened and optimized to remove pathways that do not conform to the principles of biomechanics, retain key mechanical transmission pathways, and form a set of key mechanical transmission pathways.

8. The method according to claim 1, characterized in that In step S3, multimodal risk feature decoupling and dynamic weighted aggregation are performed based on the path reinforcement feature vector to obtain a structured risk map, including: S31. Based on the joint optimization mechanism of hierarchical variational autoencoder and dynamic graph attention, the path reinforcement feature vector is decoupled into a structured multimodal risk feature. S32. Dynamically weighting the decoupled multimodal risk features according to the decoupling quality of the features to obtain weighted multimodal risk features; S33. Based on the weighted multimodal risk features, a structured risk map is obtained through a graph structure learning mechanism.

9. The method according to claim 1, characterized in that In step S4, the early risk assessment of spinal curvature abnormality is performed based on the structured risk map through adaptive risk threshold mapping and dynamic risk quantification, including: S41. Extract key risk nodes related to spinal curvature and their associated edge features from the structured risk graph to construct a spinal-specific risk subgraph; S42. Generate personalized risk threshold baselines based on individual patient characteristics and historical risk trajectories through a lightweight meta-learning model; S43. Based on the spinal-specific risk subgraph and the personalized risk threshold baseline, a dynamic risk value and a risk fluctuation entropy are calculated using a joint quantitative model of dynamic time warping and dual-stream gated recurrent units. S44. Based on the dynamic risk value and risk fluctuation entropy, an early risk assessment report including risk grading, trend prediction and personalized intervention suggestions is generated through a multimodal fusion decision engine.

10. An early risk assessment system for spinal curvature abnormality, characterized in that: The system includes a spatiotemporal feature fusion and three-dimensional biomechanical modeling module, a key mechanical transmission path weighted calculation module, a multimodal risk feature decoupling and dynamic weighted aggregation module, and an early risk assessment module for spinal curvature abnormalities, wherein: The spatiotemporal feature fusion and three-dimensional biomechanical modeling module is used to fuse spatiotemporal sequence features and perform three-dimensional biomechanical modeling based on the acquired multimodal data to obtain a dynamic coupling network of the spine, pelvis and lower limbs; The key mechanical transmission path weighted calculation module is used to perform weighted calculation of the key mechanical transmission path based on the dynamic coupling network through the graph attention mechanism to obtain a path reinforcement feature vector; The multimodal risk feature decoupling and dynamic weighted aggregation module is used to perform multimodal risk feature decoupling and dynamic weighted aggregation based on the path reinforcement feature vector to obtain a structured risk map; The spinal curvature abnormality early risk assessment module is used to perform early risk assessment of spinal curvature abnormality based on the structured risk map through adaptive risk threshold mapping and dynamic risk quantification.

Images