Gait analysis disease diagnosis method and system based on feature interaction rebalancing

Abstract

The application belongs to the technical field of disease diagnosis based on gait analysis, and discloses a gait analysis disease diagnosis method and system based on feature interaction rebalancing, which comprises the following steps: feature extraction is performed on gait data to obtain local features and global features, cross enhancement is performed on the two features, splicing operation is performed on the results, classification prediction is performed by using the splicing results, a classification result is obtained, cross entropy loss between the classification result and a real label is calculated, prototype cross entropy loss between weighted local features and weighted global features and a prototype vector is calculated, similarity loss of a similarity matrix, a transpose matrix and the real label is calculated, the losses are combined by weighting, a total loss function of the model is obtained, and the model is trained. The application simultaneously learns intra-joint and inter-joint coupling relations, reduces cross-domain feature distribution difference by aligning cross-domain features, balances fitting speed of a cross-domain extractor, and provides an effective and objective gait-based reference for disease diagnosis.

Description

Technical Field

[0001] This invention relates to the field of gait analysis-based disease diagnosis technology, and in particular to a gait analysis-based disease diagnosis method and system based on feature interaction rebalancing. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Gait analysis is an important auxiliary method for effectively diagnosing diseases with abnormal gait characteristics, which are often misdiagnosed in clinical practice. Given the complexity of multivariate gait time series, deep learning methods have been widely applied to disease diagnosis using gait signal analysis. Although current deep learning methods have significantly improved the accuracy of gait analysis tasks compared to traditional methods, some challenges remain to be overcome. Challenge 1: Complex coupling relationships between and within joints (e.g., Figure 1 As shown, it is difficult to extract. Challenge 2: In the process of multi-feature fusion, due to the different fitting speeds of each feature extraction process, they will interfere with each other (e.g., Figure 2 (As shown). For Challenge 1, existing Graph Neural Networks (GNNs) consider cross-joint coupling, but they face a key shortcoming: they ignore intra-joint coupling features, which are crucial for disease differentiation. For Challenge 2, the state-of-the-art Prototype Modality Rebalancing (PMR) algorithm accelerates the slower fitting process while evaluating the classifier's fit by comparing the distance between features and prototypes in real time. However, PMR's performance is limited by its lack of interaction between cross-domain features, leading to distribution discrepancies. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a gait analysis-based disease diagnosis method and system based on feature interaction rebalancing. This method simultaneously learns intra- and inter-articular coupling relationships, reduces cross-domain feature distribution differences by aligning cross-domain features, and balances the fitting speed of the cross-domain extractor, thus providing an effective and objective gait-based reference for disease diagnosis.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a gait analysis-based disease diagnosis method based on feature interaction rebalancing, comprising the following steps:

[0007] Gait data samples are acquired, divided into multiple batches, and feature extraction is performed on each batch of gait data samples to obtain local and global features.

[0008] A cross-enhancement operation is performed on local and global features to obtain weighted local and weighted global features;

[0009] The weighted local features and weighted global features are concatenated to obtain fused features. The fused features are used for classification prediction to obtain classification results. The cross-entropy loss between the classification results and the true labels is calculated. The prototype vectors of the weighted local features and weighted global features are obtained separately. The prototype cross-entropy loss between the weighted local features and weighted global features and the prototype vectors is calculated separately. The similarity matrix of the weighted local features and weighted global features for each batch is calculated. The similarity loss between the similarity matrix and its transpose matrix and the true labels is calculated separately.

[0010] The above losses are weighted and combined to obtain the total loss function of the model, and the model is then trained.

[0011] As an alternative implementation, the local feature is the coupling within a joint, that is, the relationship between the three degrees of freedom within a joint, and the global feature is the coupling between multiple joints.

[0012] As an alternative implementation, a cross-enhancement operation is performed on local and global features, specifically:

[0013] For local and global features, one feature sequence is used to provide the query sequence, while the other feature sequence provides the key and value sequences. By calculating the similarity between the query sequence and the key sequence, the attention distribution between different sequences is obtained. The attention distribution is then used to perform a weighted summation on the value sequences to obtain weighted local features and weighted global features.

[0014] As an alternative implementation, a local two-dimensional residual module is used to capture intra-joint coupling, and a global dynamic graph learning module is used to capture inter-joint coupling.

[0015] As an alternative implementation, the prototype cross-entropy loss can assess the degree of imbalance between two feature extractors, and the similarity loss enables the model to learn feature representations that reflect similarity relationships in the data.

[0016] As an alternative implementation, the total loss function is:

[0017]

[0018] in, For cross-entropy loss, and These are the prototype cross-entropy losses of the two feature extractors, respectively. For similarity loss, α is a hyperparameter controlling the equilibrium fitting speed, ε is a hyperparameter used for similarity modulation, and β and γ are coefficients.

[0019] Secondly, the present invention provides a gait analysis disease diagnosis system based on feature interaction rebalancing, comprising:

[0020] The feature extraction module is configured to: acquire gait data samples, divide them into multiple batches, and extract features from each batch of gait data samples to obtain local and global features;

[0021] The cross-enhancement module is configured to perform cross-enhancement operations on local and global features to obtain weighted local and weighted global features.

[0022] The loss calculation module is configured to: concatenate weighted local features and weighted global features to obtain fused features; use the fused features to perform classification prediction to obtain classification results; calculate the cross-entropy loss between the classification results and the true labels; obtain the prototype vectors of weighted local features and weighted global features respectively; calculate the prototype cross-entropy loss between weighted local features and weighted global features and prototype vectors respectively; calculate the similarity matrix of weighted local features and weighted global features for each batch; and calculate the similarity loss between the similarity matrix and its transpose matrix and the true labels respectively.

[0023] The model training module is configured to: weight and combine the above losses to obtain the total loss function of the model, and then train the model.

[0024] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0025] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0026] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This disclosure proposes a gait analysis-based disease diagnosis method and system based on feature interaction rebalancing, systematically investigating the challenges of gait-based disease diagnosis: complex coupling relationships and cross-domain fusion interference. A Global-Local Interactive Rebalancing Gait Network (GLIR-GaitNet) is proposed to address these issues. It includes two modules: a Global-Local Joint Coupling Feature Extractor (GL-JCFE) and a Prototype Interactive Rebalancing (PIR). The GL-JCFE module integrates a dynamic graph learning module and a two-dimensional residual module. The dynamic graph learning module extracts global coupling features across joints, while the two-dimensional residual module extracts local coupling features between the three degrees of freedom within a joint. The combination of these two modules aims to capture comprehensive coupling across and within joints, solving the problem of extracting complex couplings between and within joints. Furthermore, the GL-JCFE module introduces MCE to assign appropriate weights to cross-domain features, enhancing the complementarity of cross-domain features while reducing feature redundancy generated by the two parallel networks.

[0029] The introduction of cross-domain similarity supervision in the PIR module helps learn shared feature representations, thereby reducing information conflicts and noise interference, and ultimately improving the model's ability to understand and interpret data. It reduces the differences in feature distributions while balancing the convergence speed of the cross-domain feature extraction network, thus promoting feature integration. It addresses the issue of interference between features in existing gait-based disease diagnosis methods during multi-feature fusion due to varying fitting speeds of different feature extraction processes. Experiments based on self-collected datasets of cervical spondylotic myelopathy (CSM) and Parkinson's syndrome (PS) demonstrate that our method significantly improves disease diagnostic capabilities.

[0030] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0032] Figure 1 This is a schematic diagram illustrating Challenge 1 in the background art of this invention;

[0033] Figure 2 This is a schematic diagram illustrating Challenge 2 in the background art of this invention;

[0034] Figure 3 This is a framework diagram of the GL-JCFE module provided in Embodiment 1 of the present invention;

[0035] Figure 4 This is a framework diagram of the PIR module provided in Embodiment 1 of the present invention;

[0036] Figure 5 This is an overall framework diagram of the gait analysis-based disease diagnosis method based on feature interaction rebalancing provided in Embodiment 1 of the present invention;

[0037] Figure 6 The confusion matrix, scatter plot, and ROC curve for sample classification in this invention are shown. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0040] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0041] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0042] Example 1

[0043] like Figure 5 As shown, this embodiment provides a gait analysis-based disease diagnosis method based on feature interaction rebalancing, including the following steps:

[0044] S1 acquires gait data samples, divides them into multiple batches, and extracts features from each batch of gait data samples to obtain local and global features.

[0045] S2 performs a cross-enhancement operation on local and global features to obtain weighted local and weighted global features;

[0046] S3 concatenates weighted local features and weighted global features to obtain fused features. It then uses the fused features for classification prediction to obtain classification results. It calculates the cross-entropy loss between the classification results and the true labels. It obtains the prototype vectors of the weighted local features and weighted global features respectively, and calculates the prototype cross-entropy loss between the weighted local features and weighted global features and the prototype vectors respectively. It calculates the similarity matrix of the weighted local features and weighted global features for each batch, and calculates the similarity loss between the similarity matrix and its transpose matrix and the true labels respectively.

[0047] S4 performs a weighted combination of the above losses to obtain the model's total loss function, and then trains the model.

[0048] The local feature is the coupling within a joint, that is, the relationship between the three degrees of freedom within a joint, while the global feature is the coupling between multiple joints.

[0049] The Global-Local Interactive Rebalancing Gait Network (GLIR-GaitNet) disclosed herein is the first gait network to simultaneously learn intra- and inter-joint coupling relationships. It includes the GL-JCFE module (such as...). Figure 3 (as shown) and PIR module (as shown) Figure 4 (As shown). The GL-JCFE module consists of three components: a local 2D residual module for capturing local features within the three degrees of freedom of the same joint; a global dynamic graph learning module for extracting global features between joints; and an MCE module for enhancing the complementarity between these two types of features. The PIR module enhances the interaction between global and local feature information by incorporating SIM loss, thereby addressing the imbalance problem in the multi-feature fusion process. Finally, the fused multi-features are input into a fully connected layer for triple classification to obtain the classification result.

[0050] In the feature extraction section, the proposed GL-JCFE module can capture comprehensive intra- and inter-joint couplings. In this disclosure, local features and global features refer to features learned from intra- and inter-joint relationships, respectively. Specifically, the relationship between three degrees of freedom within a joint is considered a local feature, while the coupling between multiple joints is considered a global feature.

[0051] Local 2D residual modules are used to capture intra-joint couplings as local features. A single gait time series, consisting of three degrees of freedom within the same joint, is viewed as a 2D matrix. Intra-joint coupling features are obtained using a 2D convolutional kernel. The sliding operation of this 2D convolutional kernel on the matrix can be interpreted as multiple small filters to generate new feature maps. These feature maps capture the dependencies between the three degrees of freedom.

[0052] The global dynamic graph learning module captures inter-joint couplings as global features. The dependencies between multiple joints across 18 dimensions can be modeled as a graph, and hidden spatiotemporal dependencies are extracted using existing dynamic graph networks. To establish initial relationships, the graph construction and learning module generates a set of graph adjacency matrices, each corresponding to a different time slot. The elements of the graph adjacency matrices are learned through an iterative process. The core modules of the dynamic graph neural network include dynamic graph transformation and dynamic graph isomorphism networks. Dynamic graph transformation is used to exploit dynamic associations between graphs at different times, while dynamic graph isomorphism networks are used to capture spatiotemporal dependencies.

[0053] The multi-feature cross-enhancement module aims to enhance and integrate information from different feature extractors. In multi-branch feature extraction networks, features from different extractors are highly complementary. What should be enhanced are complementary (irrelevant) information, rather than redundant (correlated) features.

[0054] Inspired by the cross-attention mechanism, this approach utilizes two branches with different parameters to extract features from two extractors. In traditional self-attention mechanisms, there is only one input sequence containing three representations: query, key, and value. Conversely, in the cross-attention mechanism, for two feature sequences... and Provide query sequences, and Provide key and value sequences. The attention distribution between different sequences is obtained by calculating the similarity between the query sequence and the key sequence. Then, a weighted sum is performed on the value sequences using the attention distribution to obtain a representation of the output sequence. The formula for the cross-enhancement module is as follows:

[0055]

[0056] in, and This indicates the input of the MCE module. and Different representations of the input, It is a learnable matrix, norm(·) denotes the linear norm operation, and d is the dimension of the input vector.

[0057] In the multi-feature fusion process, a PIR strategy was established to reduce interference between the two feature extraction processes and enhance the consistency of feature distribution. It accelerates the slower fitting feature extraction process by promoting feature utilization and reduces distribution gaps by performing alignment operations on features from the same sample.

[0058] First, the "prototype" concept for these two features is defined as "representative embeddings of class instances." A nonparametric classifier is built using prototypes, and the performance of each fitting process is evaluated by measuring the distance between each sample feature and its respective prototype feature. Processes with lower fitting performance are accelerated by assigning a prototype cross-entropy (PCE) loss coefficient. Furthermore, for each batch of generated sample features, a cross-domain similarity loss (SIM loss) is calculated and added to the backpropagation process. This minimizes the SIM loss, increases the consistency of feature distributions, and reduces information conflicts during the fusion process.

[0059] For inputs f0 and f1 to two feature extraction networks (a global feature extraction network and a local feature extraction network, respectively), the dataset is represented as... It consists of instances and their corresponding labels (x, y), where And y = {0, 1, 2}. The classification task has a total of 3 categories, and the goal is to train the model using data that predicts starting from x.

[0060] For data We can generate representations we will and This represents a subset of data for each category. The prototype is the centroid of the data for each category:

[0061]

[0062] The prototype is used to generate a class-based distribution for the input data x based on the softmax of the distance to the prototype in the embedding space of each modality:

[0063]

[0064] Where d(.,.) is the distance function, i.e., the Euclidean distance in this paper. The imbalance ratio is defined as:

[0065]

[0066] in, and It is a batch of data for training step time t, ρ t This can be used to realistically evaluate the degree of imbalance between two feature extractors. To reduce interference between the two extractors during the melting process, PCE loss is independently introduced to balance the fitting speed:

[0067]

[0068] For the weighted features obtained by the two extractors in each batch and Let F1 and F2 represent them respectively. Local features can be represented by {v0, v1, ..., v...} K} represents the total number of samples in the batch, where K+1 is the total number of samples in the batch. Similarly, the global feature is represented as {t0, t1, ..., t K The time between them. The similarity loss in two different directions is defined as:

[0069]

[0070] Taking the first equation as an example, τ is a hyperparameter. These are all local features. When When the global sample feature is positive, These are negative global sample features that do not match local features. Where f v (·) and f t (·) are two projection heads that map the representation to a space where cross-entropy loss is applied. In summary, we define the SIM loss as:

[0071]

[0072] The loss during backpropagation is a weighted combination of CE loss, PCE loss, and SIM loss:

[0073]

[0074] in, For cross-entropy loss, and These are the prototype cross-entropy losses of the two feature extractors, respectively. For similarity loss, α is a hyperparameter controlling the equilibrium fitting speed, ε is a hyperparameter used for similarity modulation, and β and γ are coefficients.

[0075] Use ρ t To dynamically assess imbalance, the learning speed of each modality can be adjusted by modifying the coefficients β and γ in a simple way:

[0076]

[0077] Here, clip(a,b,c) is a cutoff function that restricts b to between a and c. This approach allows for the convenient use of slower-learning feature extractors while better extractors maintain the original learning strategy, thus mitigating the imbalance phenomenon.

[0078] Our dataset includes healthy subjects (HC), CSM patients, and PS patients. Data from subjects with issues such as insufficient length and poor quality were excluded, resulting in 51 CSM samples, 33 HC samples, and 49 PS samples, as shown in Table 1.

[0079] An inertial motion acquisition system was used to capture motion and output kinematic data from 11 sensors fixed to various parts of the subject's body, focusing primarily on data from six sensors at the lower limb joints. Each data sequence included three directions: sagittal (X), coronal (Y), and horizontal (Z). The human body was modeled using Visual 3D software (version 2021, C-motion Inc, USA), and the kinematic data captured by the sensors was converted into time-series data of joint parameters.

[0080] Table 1: Dataset Category Names and Sample Counts

[0081]

[0082]

[0083] The experiment used two time-series data sets: joint angle and joint angular velocity. Based on three-minute walking data, each time series was 1202 in length and 17 in dimension (six sensors with three degrees of freedom), with two samples per subject. To effectively combine the velocity and acceleration data, the data were laterally concatenated to form a 2404-dimensional dataset as the total input to the network. The proposed framework was implemented using the PyTorch toolbox. Model training was performed on an environment equipped with an NVIDIA RTX 3090 GPU. The batch size was set to 8, and training was conducted over 150 epochs. An SGD optimizer with 0.9 momentum was used, starting with an initial learning rate of 0.001, which was reduced to 0.0001 after 70 epochs.

[0084] To comprehensively evaluate the proposed model, two key metrics were used: ACC (Accuracy) and AUC (Area Under the Curve). ACC measures the overall correctness of the model's predictions. Its calculation method is as follows:

[0085]

[0086] TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively. AUC measures the model's ability to distinguish categories. It is the area under the ROC (Receiving Controller Operational Characteristic) curve, plotting the true positive rate (sensitivity) versus the false positive rate (1-specificity) at various threshold settings. A higher AUC indicates better model performance.

[0087] To demonstrate the feasibility of the proposed model, it was compared with several traditional and state-of-the-art multivariate time series classification methods: CNN-LSTM, Deep 1D-Convnet, ALSTM-FCN, TodyNet, and ResNet. The classification results of these algorithms on our own collected dataset are shown in Table 2. For fair comparison, all compared algorithms used the same training and test samples. Experimental results show that our algorithm outperforms all state-of-the-art methods by a significant margin in all metrics, including ACC and AUC. It achieves an overall ACC of 91.45%. Specifically, PS's predicted ACC is significantly higher at 94.34%, while CSM's is relatively lower. Our method improves the average ACC by 2.27% and the average AUC by 3.27%. These improvements are attributed to our comprehensive extraction of coupling relationships while ensuring the minimization of interference and feature consistency during the fusion process. Figure 6 The ROC curves, scatter plots, and confusion matrices for each category are shown, highlighting the superior performance of our algorithm in gait diagnosis.

[0088] Table 2: Quantitative comparison of different methods on self-collected datasets. ACC (%) and AUC (%) are shown, with best results indicated in bold.

[0089]

[0090] To analyze the contribution of each component in the proposed model to the classification performance, experiments were conducted by selectively removing parts of the model. Tables 3 and 4 present the experimental results, where √ indicates the addition of a module.

[0091] Table 3: Ablation Studies of MCE Strategy

[0092]

[0093] The MCE module plays a crucial role in the feature fusion network. The experimental results in Table 3 highlight the effectiveness of the MCE module in enhancing and integrating information from different feature extractors. The results show that the MCE strategy improves ACC and AUC by approximately 1.51% and 3.67%, respectively. Reasonable weight allocation ensures strong complementarity of features from different extractors, reducing the interference of redundant features on classification.

[0094] Table 4: Ablation studies using PIR strategy

[0095]

[0096] Simulation results validate that the integrated PIR strategy improves the model's classification performance, as shown in Table 4. Furthermore, to provide additional evidence of the effectiveness of the PIR strategy, a comparative analysis of the PMR method was conducted. Our results indicate that the PIR strategy improves ACC and AUC by approximately 0.51% and 0.67%, respectively. This improvement stems from the PIR strategy, which overcomes the limitation of inconsistent cross-domain feature distributions observed in the PMR method by introducing cross-domain similarity loss monitoring.

[0097] This publication systematically investigates the challenges of gait-based disease diagnosis and addresses these challenges using GLIR-GaitNet, which comprises a GL-JCFE module for feature extraction and a PIR module for feature fusion. For Challenge 1, involving complex intra- and inter-joint couplings, the GL-JCFE module integrates a local 2D residual module for intra-joint couplings and a dynamic graph learning module for inter-joint global relationships. For Challenge 2, where process interference and inconsistent feature representations hinder multi-feature fusion, the PIR module enhances cross-domain similarity supervision and balances feature extraction speed, improving fusion performance. Experimental results demonstrate the superiority of our method on self-collected datasets for triple classification tasks (CSM, PS, and HC). The model achieves prediction accuracies of 91.45% and 93.48% for ACC and AUC, respectively, representing improvements of 2.77% and 3.27% over existing models.

[0098] Example 2

[0099] This embodiment provides a gait analysis disease diagnosis system based on feature interaction rebalancing, including:

[0100] The feature extraction module is configured to: acquire gait data samples, divide them into multiple batches, and extract features from each batch of gait data samples to obtain local and global features;

[0101] The cross-enhancement module is configured to perform cross-enhancement operations on local and global features to obtain weighted local and weighted global features.

[0102] The loss calculation module is configured to: concatenate weighted local features and weighted global features to obtain fused features; use the fused features to perform classification prediction to obtain classification results; calculate the cross-entropy loss between the classification results and the true labels; obtain the prototype vectors of weighted local features and weighted global features respectively; calculate the prototype cross-entropy loss between weighted local features and weighted global features and prototype vectors respectively; calculate the similarity matrix of weighted local features and weighted global features for each batch; and calculate the similarity loss between the similarity matrix and its transpose matrix and the true labels respectively.

[0103] The model training module is configured to: weight and combine the above losses to obtain the total loss function of the model, and then train the model.

[0104] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0105] In further embodiments, the following is also provided:

[0106] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0107] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0108] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0109] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0110] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0111] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0112] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0113] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0114] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and so on. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0115] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0116] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A gait analysis-based disease diagnosis method based on feature interaction rebalancing, characterized in that, Includes the following steps: Gait data samples are acquired, divided into multiple batches, and feature extraction is performed on each batch of gait data samples to obtain local and global features. Local features are the couplings within a joint, i.e. the relationship between the three degrees of freedom within a joint, while global features are the couplings between multiple joints. Local two-dimensional residual modules are used to capture the couplings within a joint, and global dynamic graph learning modules are used to capture the couplings between joints. A cross-enhancement operation is performed on local and global features to obtain weighted local and weighted global features; The weighted local features and weighted global features are concatenated to obtain fused features. The fused features are used for classification prediction to obtain classification results. The cross-entropy loss between the classification results and the true labels is calculated. The prototype vectors of the weighted local features and weighted global features are obtained separately. The prototype cross-entropy loss between the weighted local features and weighted global features and the prototype vectors is calculated separately. The similarity matrix of the weighted local features and weighted global features for each batch is calculated. The similarity loss between the similarity matrix and its transpose matrix and the true labels is calculated separately. The above losses are weighted and combined to obtain the total loss function of the model, and the model is then trained.

2. The gait analysis-based disease diagnosis method based on feature interaction rebalancing as described in claim 1, characterized in that, The operation involves cross-enhancement of local and global features, specifically: For local and global features, one feature sequence is used to provide the query sequence, while the other feature sequence provides the key and value sequences. By calculating the similarity between the query sequence and the key sequence, the attention distribution between different sequences is obtained. The attention distribution is then used to perform a weighted summation on the value sequences to obtain weighted local features and weighted global features.

3. The gait analysis-based disease diagnosis method based on feature interaction rebalancing as described in claim 1, characterized in that, The prototype cross-entropy loss assesses the degree of imbalance between two feature extractors, while the similarity loss enables the model to learn feature representations that reflect similarity relationships in the data.

4. The gait analysis-based disease diagnosis method based on feature interaction rebalancing as described in claim 1, characterized in that, The total loss function is: in, For cross-entropy loss, and These are the prototype cross-entropy losses of the two feature extractors, respectively. For similarity loss, It is a hyperparameter that controls the speed of equilibrium fitting. These are hyperparameters used for similarity modulation. and is a coefficient.

5. A gait analysis disease diagnosis system based on feature interaction rebalancing, characterized in that, include: The feature extraction module is configured to: acquire gait data samples, divide them into multiple batches, and extract features from each batch of gait data samples to obtain local and global features; Local features are the couplings within a joint, i.e. the relationship between the three degrees of freedom within a joint, while global features are the couplings between multiple joints. Local two-dimensional residual modules are used to capture the couplings within a joint, and global dynamic graph learning modules are used to capture the couplings between joints. The cross-enhancement module is configured to perform cross-enhancement operations on local and global features to obtain weighted local and weighted global features. The loss calculation module is configured to: concatenate weighted local features and weighted global features to obtain fused features; use the fused features to perform classification prediction to obtain classification results; calculate the cross-entropy loss between the classification results and the true labels; obtain the prototype vectors of weighted local features and weighted global features respectively; calculate the prototype cross-entropy loss between weighted local features and weighted global features and prototype vectors respectively; calculate the similarity matrix of weighted local features and weighted global features for each batch; and calculate the similarity loss between the similarity matrix and its transpose matrix and the true labels respectively. The model training module is configured to: weight and combine the above losses to obtain the total loss function of the model, and then train the model.

6. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-4.

8. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-4.

Images