A pathological gait phase synchronization compensation analysis method and system based on a kuramoto oscillator graph convolution network

Abstract

This invention discloses a method and system for pathological gait phase synchronization compensation analysis based on a Kuramoto oscillator graph convolutional network, belonging to the fields of visual gait analysis, graph neural networks, and intelligent rehabilitation assessment. The method acquires the temporal sequence of the human gait skeleton and constructs a spatiotemporal skeleton graph. It extracts joint node features using a spatiotemporal graph convolutional network, establishes a Kuramoto phase oscillator model using the skeleton structure adjacency matrix as a structural prior, and completes phase updates based on phase state, natural frequency, and learnable coupling strength. This generates a time-varying functional connectivity matrix and phase-gated signals to enhance the original spatiotemporal features, ultimately outputting the abnormal muscle category, lesion region, or lesion combination. This method can explicitly characterize the non-local compensatory synergistic relationships in pathological gait, improving diagnostic stability, interpretability, and clinical reference value.

Description

Technical Field

[0001] This invention relates to the fields of visual gait analysis, graph neural networks, coupled oscillator modeling and intelligent rehabilitation assessment, and in particular to a pathological gait phase synchronization compensation analysis method based on Kuramoto oscillator graph convolutional network. Background Technology

[0002] Pathological gait analysis aims to identify potential neuromuscular abnormalities from human kinematic signals. While existing graph convolutional networks based on skeleton sequences can characterize joint topological relationships, most employ fixed adjacency matrices or static graph structures, which can only describe local connections between adjacent joints and are insufficient to depict the dynamic collaborative relationships across limbs and body parts caused by compensatory behaviors in pathological gait.

[0003] Clinically, pathological gait is often not a localized abnormality of a single joint, but rather a systemic reorganization process influenced by neural control, musculoskeletal dynamics, and overall balance. Abnormal gait may initially manifest as a synergistic imbalance between the affected and healthy sides, subsequently triggering new functional connections or compensatory linkages between distal joints. Traditional mechanism-agnostic deep learning methods typically treat such dynamic synergistic relationships as statistical correlations, lacking clear physical meaning and interpretability.

[0004] Therefore, there is an urgent need for a gait analysis method that can directly embed the joint phase synchronization mechanism into graph convolutional networks for adaptive reconstruction of time-varying functional connections, thereby improving the reliability and clinical interpretability of pathological gait identification. Summary of the Invention

[0005] The purpose of this invention is to provide a pathological gait phase synchronization compensation analysis method based on Kuramoto oscillator graph convolutional networks, in order to solve the problems that existing skeleton graph convolutional networks are difficult to characterize the non-local cooperative relationships of pathological gait, lack dynamic graph structure, and have insufficient mechanism explanation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: First, the temporal sequence of the human gait skeleton is obtained, a spatiotemporal skeleton graph is constructed, and a high-level spatiotemporal feature is extracted using a spatiotemporal graph convolutional network; then, using the skeleton structure adjacency matrix as the structural prior, a Kuramoto phase oscillator model is established for each joint node, and phase evolution and normalization updates are performed; then, a Gaussian similarity matrix is ​​constructed based on the phase difference, and dynamic edge weights are obtained by combining the sliding window phase locking value, thereby forming a time-varying functional connectivity matrix; simultaneously, phase-gated signals are generated through static and dynamic gated branches to enhance the spatiotemporal features; finally, graph convolution operations are performed on the enhanced features and classification output is performed.

[0007] A method for pathological gait phase synchronization compensation analysis based on Kuramoto oscillator graph convolutional networks includes the following steps: Step S1: Obtain the temporal sequence of the human gait skeleton and construct a spatiotemporal skeleton diagram based on the spatial connection relationship of the human joints; the temporal sequence of the human gait skeleton is a two-dimensional skeleton sequence or a three-dimensional skeleton sequence; the human posture is uniformly represented as a tensor representation containing V joint nodes, T time points, and C-dimensional node attributes.

[0008] Step S2: Perform spatiotemporal graph convolution feature extraction on the spatiotemporal skeleton graph to obtain the spatiotemporal features of each joint node; Step S3: Using the structural adjacency matrix of the human skeleton as the structural prior, a Kuramoto phase oscillator model is established for each joint node. The phase is updated based on the phase state, natural frequency, and learnable coupling strength of the joint node to obtain the updated phase information. The phase dynamics of each joint node i satisfy:

[0009] in, Let be the natural frequency of node i, and K be the learnable coupling strength. For elements of the structural adjacency matrix, Zero-mean Gaussian phase noise is introduced during the training phase; Phase updates are performed using an explicit Euler method, which involves first obtaining the phase increment based on phase dynamics, and then normalizing the updated phase to be confined to the interval (-π, π]. Step S4: Generate a dynamic gating signal based on the updated phase information, and construct a time-varying functional connection matrix based on the inter-joint phase difference, the adjacency matrix of the human skeleton structure, and the phase lock value of the sliding window. Based on the updated phase information, dynamic gating signals and time-varying functional connection matrices are generated respectively; wherein, the dynamic gating signals are obtained by phase encoding and mapping through a multilayer perceptron, and are used to reweight the channel-node level spatiotemporal features of each joint node; the time-varying functional connection matrix is ​​jointly determined by the inter-joint phase similarity matrix, the human skeleton structure adjacency matrix, and the sliding window phase locking value. The time-varying functional connectivity matrix is ​​obtained as follows: First, a Gaussian similarity matrix is ​​constructed based on the phase difference between joints, serving as the phase similarity matrix between each pair of joints; then, the Gaussian similarity matrix is ​​multiplied element-wise with the adjacency matrix of the human skeleton structure to obtain an initial functional connectivity matrix constrained by structural priors; subsequently, the edge weights of the initial functional connectivity matrix are scaled using the sliding window phase lock value (PLV) to obtain a time-varying functional connectivity matrix constrained by both structural priors and phase synchronization strength.

[0010] Step S5: The spatiotemporal features obtained in step S2 are reweighted element-wise using the dynamic gating signal, and graph convolution is performed based on the time-varying functional connection matrix to obtain enhanced spatiotemporal features; The spatiotemporal features obtained in step S2 are reweighted element-wise using the dynamic gating signal generated in step S4 to obtain gated enhanced features; Symmetric normalized graph convolution is performed on the time-varying functional connection matrix obtained in step S4 to obtain enhanced spatiotemporal features with fused phase synchronization relationships.

[0011] Step S6: Input the enhanced features into the classification head and output the abnormal muscle category, lesion area or lesion combination corresponding to the pathological gait.

[0012] The compensation system used in the pathological gait phase synchronization compensation analysis method based on the Kuramoto oscillator graph convolutional network includes: a data input module, a spatiotemporal feature extraction module, a phase dynamics modeling module, a functional connectivity construction module, a graph convolution enhancement module, and a recognition output module.

[0013] The data input module is connected to the spatiotemporal feature extraction module to obtain the temporal sequence of human gait skeleton, and to perform skeleton scale normalization and filtering noise reduction on the temporal sequence of human gait skeleton to reduce the influence of differences in body proportions of different subjects and sensor noise on joint phase estimation; after preprocessing, a spatiotemporal skeleton map is constructed based on the spatial connection relationship of human joints.

[0014] The spatiotemporal feature extraction module is connected to the phase dynamics modeling module and is used to extract spatiotemporal graph convolution features from the spatiotemporal skeleton graph to obtain the spatiotemporal features of each joint node, and to provide feature input for subsequent joint phase encoding and phase synchronization modeling.

[0015] The phase dynamics modeling module is connected to the functional connection construction module. It is used to establish a Kuramoto phase oscillator model for each joint node using the structural adjacency matrix of the human skeleton as the structural prior. It also performs phase updates based on the phase state, natural frequency, and learnable coupling strength of the joint node to obtain the updated phase information.

[0016] The functional connection construction module is connected to the graph convolution enhancement module and is used to calculate the phase difference between joints based on the updated phase information, construct the phase similarity matrix between joints, and generate a time-varying functional connection matrix by combining the human skeleton structure adjacency matrix and the sliding window phase lock value.

[0017] The functional connection construction module is connected to the graph convolution enhancement module and is used to calculate the inter-joint phase difference based on the updated phase information, construct the inter-joint phase similarity matrix, and generate a time-varying functional connection matrix by combining the human skeleton structure adjacency matrix and the sliding window phase lock value. The graph convolution enhancement module is connected to the recognition output module and is used to perform graph convolution operation based on the time-varying functional connection matrix, and to reweight the spatiotemporal features using the dynamic gating signal generated by phase encoding to obtain enhanced spatiotemporal features that fuse joint phase synchronization relationships. The identification output module is used to input the enhanced spatiotemporal features into the classification head, output the abnormal muscle category, lesion area or lesion combination corresponding to the pathological gait, and output the joint phase synchronization intensity, time-varying functional connectivity relationship or order parameter change results as auxiliary analysis information. Compared with the prior art, the present invention has at least the following beneficial effects: 1. It can characterize the dynamic coordination relationship between joints from the perspective of phase synchronization, and explicitly reflect the compensatory transmission chain in pathological gait.

[0018] 2. Simultaneously utilize structural priors and dynamic phase information to reduce the representational limitations caused by fixed topology.

[0019] 3. Thirdly, the obtained phase distribution, dynamic adjacency, and synchronization intensity all have clear interpretive significance and can provide a visual reference for clinical rehabilitation assessment. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the joint phase update process based on the Kuramoto phase oscillator; Figure 3 A schematic diagram comparing the Kuramoto sequence parameters and average coupling strength for normal, mildly damaged, and moderately damaged gaits. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0022] In the embodiment, the overall process is as follows: Figure 1 As shown, the input is a temporal sequence X of the human gait skeleton, whose dimensions can be represented as B×C×T×V, where B is the batch size, C is the feature dimension of each joint, T is the number of time frames, and V is the number of joint nodes. A structural adjacency matrix A is constructed based on the natural topological relationships of the human body, and a spatiotemporal graph convolutional network is used to extract high-level semantic features of each joint.

[0023] For each joint node i, a Kuramoto phase oscillator model is established. Node i has a phase.

[0024] and natural frequency Nodes are coupled through a structural adjacency matrix A and a learnable coupling strength K. Zero-mean Gaussian noise can be introduced during the training phase. To enhance robustness, the noise term is disabled during the inference phase. The phase dynamics can be written as:

[0025] Phase updates are implemented using an explicit Euler integration strategy. First, the phase increment is calculated according to the above formula. Then, the phase is constrained to the (-π, π] interval using a modulo 2π approach to avoid numerical instability caused by phase drift. After these steps, the normalized phase is obtained. This serves as the basis for subsequent functional connection reconstruction and gating enhancement.

[0026] A schematic diagram of the phase update process is shown below. Figure 2 As shown, in the static enhancement branch, The encoded input is a multilayer perceptron, which outputs a node-level static gating signal and normalizes it using spatial Softmax, thereby highlighting key joints while keeping the original skeleton topology unchanged.

[0027] In the dynamic enhancement branch, the phase difference is utilized. Constructing Gaussian similarity:

[0028] Then multiply it element-wise with the structural adjacency matrix A to obtain the initial dynamic graph. Furthermore, the local synchronization degree is measured by the sliding window phase-locked value (PLV), and this synchronization degree is used to perform a secondary scaling of the edge weights to form the final time-varying functional connectivity matrix. .

[0029] Subsequently, based on Symmetrically normalized graph convolution is performed to embed the dynamic graph structure into the skeleton feature propagation process; simultaneously, static or dynamic gating signals are applied element-wise to the input features to obtain an enhanced spatiotemporal representation. Finally, the abnormal muscle category, affected side, or combination of multiple muscle lesions corresponding to the pathological gait is output through the classification head.

[0030] In one optional embodiment, the skeleton sequence can be represented by a three-dimensional skeleton with 23 joint nodes; in another embodiment, a skeleton with other numbers of joints can also be used. As long as the corresponding structural adjacency matrix can be constructed and phase synchronization modeling can be completed, it falls within the protection scope of this invention.

[0031] like Figure 3 As shown in the figure, the Kuramoto analysis results are compared among normal subjects, patients with mild gait impairment, and patients with moderate gait impairment. Among them, Figure 3The top row shows the sequence parameter variation curves of the three types of objects within a gait cycle, and the bottom row shows the corresponding average coupling strength matrix; the horizontal axis represents the time step or joint index, and the vertical axis represents the sequence parameter value or joint index, respectively; the color intensity represents the magnitude of the average coupling strength between joints. Through analysis of... Figure 3 The comparison shows that as the degree of gait impairment increases, the regularity of changes in gait sequence parameters weakens, and the average coupling distribution between joints becomes more discrete and uneven; while the gait phase changes of normal subjects are more orderly and the overall synchronization is higher. Therefore, this invention can distinguish the compensatory coordination patterns of different degrees of pathological gait by using phase synchronization characteristics and coupling intensity distribution, and provide interpretable evidence for visual gait diagnosis.

[0032] In one alternative embodiment, the skeleton sequence can be represented by a three-dimensional skeleton with 23 joint nodes; in another embodiment, it can also be represented by a two-dimensional or three-dimensional skeleton with other numbers of joints. As long as the corresponding skeleton structure adjacency matrix can be constructed, and synchronous modeling, dynamic connection reconstruction, and feature enhancement based on the Kuramoto phase oscillator are completed, it falls within the protection scope of this invention.

[0033] Example 1: Comparison of this method with a nonlinear compensation analysis method for pathological gait based on Duffing oscillator graph convolutional networks: The Duffing oscillator-based approach primarily targets the local nonlinear dynamic response during joint movement, focusing on characterizing features such as oscillation amplification, state transitions, stability degradation, and position-velocity phase diagram changes in joint trajectories during pathological gait. Its core lies in using the Duffing nonlinear oscillator to describe the nonlinear compensation behavior of individual joints or local joint groups under abnormal muscle drive, making it suitable for analyzing local joint instability, enhanced nonlinear oscillations, and complex compensation modes.

[0034] In contrast, this invention is based on the Kuramoto phase oscillator model and is primarily used to analyze the phase synchronization relationships between the left and right limbs and different joints in pathological gait. Its focus is not on whether a single joint trajectory exhibits nonlinear oscillations, but rather on whether different joint nodes maintain a coordinated phase relationship, especially the degree of synchronization of the left and right limbs during the gait cycle, changes in phase difference, and the resulting time-varying functional connectivity. Therefore, this invention is more suitable for calculating the degree of incoordination in the movement of both limbs of a patient and analyzing abnormalities in cross-joint and cross-limb collaborative compensation.

[0035] From a gait analysis perspective, different individuals have different walking habits; for example, stride length, gait speed, arm swing amplitude, joint range of motion, and movement rhythm may all vary from person to person. Judging directly based on absolute movement amplitude or unilateral joint trajectory is easily influenced by individual behavioral habits. This invention uses the phase synchronization relationship between the left and right limbs as the analysis object, and uses the left and right limb movements of the same subject as a relative reference. By comparing the phase difference, synchronization intensity, and time-varying functional connectivity changes between the two limbs, it can, to a certain extent, reduce the interference caused by individual differences in walking style.

[0036] Furthermore, aging or overall physical decline may lead to a simultaneous weakening of muscle strength in both lower limbs, manifesting as decreased gait speed, reduced joint range of motion, or overall reduced motor ability. However, if the left and right limbs still maintain good phase synchronization and coordinated movement patterns, it is inappropriate to diagnose it as pathological asymmetrical gait solely due to a simultaneous decline in bilateral motor ability. This invention focuses on the coordination imbalance between the left and right limbs, rather than simply a decrease in movement amplitude. Therefore, for cases where bilateral muscle weakness is simultaneous but the left-right movement relationship remains relatively consistent, this invention can distinguish it from unilateral abnormal compensatory patterns through phase synchronization analysis.

Claims

1. A pathological gait phase synchronization compensation analysis method based on a Kuramoto oscillator graph convolution network, characterized by, Includes the following steps: Step S1: Obtain the temporal sequence of the human gait skeleton and construct a spatiotemporal skeleton diagram based on the spatial connection relationship of human joints; Step S2: Perform spatiotemporal graph convolution feature extraction on the spatiotemporal skeleton graph to obtain the spatiotemporal features of each joint node; Step S3: Using the structural adjacency matrix of the human skeleton as the structural prior, a Kuramoto phase oscillator model is established for each joint node. The phase is updated based on the phase state, natural frequency, and learnable coupling strength of the joint node to obtain the updated phase information. The phase dynamics of each joint node i satisfy: wherein, is the natural frequency of node i, K is the learnable coupling strength, is the structural adjacency matrix element, is the zero-mean Gaussian phase noise introduced at the training phase; Phase updates are performed using an explicit Euler method, which involves first obtaining the phase increment based on phase dynamics, and then normalizing the updated phase to be confined to the interval (-π, π]. Step S4: Generate a dynamic gating signal based on the updated phase information, and construct a time-varying functional connection matrix based on the inter-joint phase difference, the adjacency matrix of the human skeleton structure, and the phase lock value of the sliding window. Step S5: The spatiotemporal features obtained in step S2 are reweighted element by element using the dynamic gating signal, and graph convolution operation is performed based on the time-varying functional connection matrix to obtain the enhanced spatiotemporal features. Step S6: Input the enhanced features into the classification head and output the abnormal muscle category, lesion area or lesion combination corresponding to the pathological gait.

2. The pathological gait phase synchronization compensation analysis method based on Kuramoto oscillator graph convolution network according to claim 1, characterized in that, In step 1, the human gait skeleton time sequence is a two-dimensional skeleton sequence or a three-dimensional skeleton sequence; the human posture is uniformly represented as a tensor representation containing V joint nodes, T time points, and C-dimensional node attributes.

3. The Kuramoto oscillator graph convolution network-based pathological gait phase synchronization compensation analysis method according to claim 1, characterized in that, In step S4, a dynamic gating signal and a time-varying functional connection matrix are generated based on the updated phase information. The dynamic gating signal is obtained by mapping phase encoding through a multilayer perceptron and is used to reweight the channel-node level spatiotemporal features of each joint node. The time-varying functional connection matrix is ​​determined by the inter-joint phase similarity matrix, the human skeleton structure adjacency matrix, and the sliding window phase lock value. The time-varying functional connectivity matrix is ​​obtained as follows: First, a Gaussian similarity matrix is ​​constructed based on the phase difference between joints, serving as the phase similarity matrix between each pair of joints; then, the Gaussian similarity matrix is ​​multiplied element-wise with the adjacency matrix of the human skeleton structure to obtain an initial functional connectivity matrix constrained by structural priors; subsequently, the edge weights of the initial functional connectivity matrix are scaled using the sliding window phase lock value (PLV) to obtain a time-varying functional connectivity matrix constrained by both structural priors and phase synchronization strength.

4. The method for pathological gait phase synchronization compensation analysis based on Kuramoto oscillator graph convolutional network according to claim 1, characterized in that, In step S5, the spatiotemporal features obtained in step S2 are reweighted element by element using the dynamic gating signal generated in step S4 to obtain gating enhanced features. Then, a symmetric normalized graph convolution operation is performed on the time-varying functional connection matrix obtained in step S4 to obtain the enhanced spatiotemporal features with fused phase synchronization relationships.

5. A compensation system used in the pathological gait phase synchronization compensation analysis method based on Kuramoto oscillator graph convolutional networks as described in any one of claims 1-4, characterized in that, include: The system includes a data input module, a spatiotemporal feature extraction module, a phase dynamics modeling module, a functional connectivity construction module, a graph convolution enhancement module, and a recognition output module.

6. The compensation system according to claim 5, characterized in that, The data input module is connected to the spatiotemporal feature extraction module and is used to acquire the human gait skeleton temporal sequence and perform skeleton scale normalization and filtering noise reduction on the human gait skeleton temporal sequence to reduce the influence of differences in body proportions of different subjects and sensor noise on joint phase estimation. After preprocessing, a spatiotemporal skeleton diagram is constructed based on the spatial connection relationship of human joints.

7. The compensation system according to claim 5, characterized in that, The spatiotemporal feature extraction module is connected to the phase dynamics modeling module and is used to extract spatiotemporal graph convolution features from the spatiotemporal skeleton graph to obtain the spatiotemporal features of each joint node, and to provide feature input for subsequent joint phase encoding and phase synchronization modeling.

8. The compensation system according to claim 5, characterized in that, The phase dynamics modeling module is connected to the functional connection construction module. It is used to establish a Kuramoto phase oscillator model for each joint node using the structural adjacency matrix of the human skeleton as the structural prior. It also performs phase updates based on the phase state, natural frequency, and learnable coupling strength of the joint node to obtain the updated phase information.

9. The compensation system according to claim 5, characterized in that, The functional connection construction module is connected to the graph convolution enhancement module and is used to calculate the phase difference between joints based on the updated phase information, construct the phase similarity matrix between joints, and generate a time-varying functional connection matrix by combining the human skeleton structure adjacency matrix and the sliding window phase lock value.

10. The compensation system according to claim 5, characterized in that, The functional connection construction module is connected to the graph convolution enhancement module and is used to calculate the inter-joint phase difference based on the updated phase information, construct the inter-joint phase similarity matrix, and generate a time-varying functional connection matrix by combining the human skeleton structure adjacency matrix and the sliding window phase lock value. The graph convolution enhancement module is connected to the recognition output module and is used to perform graph convolution operation based on the time-varying functional connection matrix, and to reweight the spatiotemporal features using the dynamic gating signal generated by phase encoding to obtain enhanced spatiotemporal features that fuse joint phase synchronization relationships. The identification output module is used to input the enhanced spatiotemporal features into the classification head, output the abnormal muscle category, lesion area or lesion combination corresponding to the pathological gait, and output the joint phase synchronization intensity, time-varying functional connectivity relationship or sequence parameter change results as auxiliary analysis information.

Images