A method, device and equipment for classifying dyspraxia under limited sample conditions

Abstract

The present invention provides a method, apparatus, and device for classifying kinesia under limited sample conditions. The method comprises: reading multimodal behavioral and physiological signal data of a patient, and extracting feature representations for latent class analysis from the data using the Bayesian Information Criterion combined with a maximum expectation algorithm; the feature representations include timestamps and spatial distribution characteristics; inputting the feature representations into a latent class analysis module, determining the optimal number of classes and boundaries using a hierarchical clustering method, and generating candidate subclusters; for the candidate subclusters, re-evaluating the likelihood of each subclusters by introducing information theory enhancement technology to obtain a confidence score for each subclusters; using an improved density peak clustering algorithm to remove redundant subclusters based on the confidence score of each subclusters to obtain final subclusters; and inputting the final subclusters into a trained classification module to estimate the kinesia category of each subclusters, thereby achieving accurate kinesia classification based on the category.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and medical health technology, and specifically relates to a method, device and equipment for classifying phobia under limited sample conditions. Background Art

[0002] Exercise-based cardiac rehabilitation (CR) is a key strategy for the treatment of cardiovascular disease (CVD) and is widely recommended by international medical guidelines. However, over 80% of eligible patients do not participate in CR, particularly among the elderly, where aphobia is prevalent (20%-80%). Therefore, a classification of aphobia is necessary to develop appropriate interventions. However, the classification of aphobia in elderly patients with cardiovascular disease and frailty faces the challenge of insufficient sample size.

[0003] With the rapid development of machine learning and deep learning technologies, classification tasks with limited samples have gradually become a research hotspot, particularly in areas such as medical diagnosis, behavioral analysis, and emotion recognition. However, due to the difficulty in obtaining samples or the high cost of annotation, achieving efficient and accurate classification with limited samples remains a pressing technical challenge. Existing technologies suffer from deficiencies in feature extraction, model generalization, and adaptability to domain-specific data, limiting their performance in real-world scenarios. For example, patent publication CN113378941B proposes a small-sample image classification method using multi-decision fusion. By integrating the decisions of multiple classifiers, this method improves the model's effectiveness and robustness, alleviating the problem of poor adaptability to new categories when training data is limited. However, this method is primarily designed for image classification tasks and relies on a multi-classifier decision fusion strategy. This may have limitations when processing complex non-image data, such as behavioral or physiological signals involved in the classification of aphobia. Furthermore, this method does not fully consider the spatial distribution of sample features, which may lead to reduced classification accuracy when the sample size is extremely limited. Furthermore, it is sensitive to noise and is difficult to adapt to the classification needs of high-dimensional, heterogeneous data. Patent publication number CN112364747B proposes a method for object detection in limited sample conditions. This method extracts target features through a backbone neural network and processes candidate regions in conjunction with a graph structure, achieving object detection tasks in low-sample scenarios. However, this technical solution focuses on target detection rather than classification tasks. Its core lies in bounding box regression and candidate region screening, but it pays insufficient attention to the inter-category differences that require refined distinctions in classification tasks. Furthermore, this method has high data preprocessing requirements and does not fully consider the subtle feature differences between categories in small sample scenarios. This may result in poor performance in tasks requiring high-precision distinctions, such as phobia classification.

[0004] These issues demonstrate that existing finite-sample classification methods still have significant shortcomings in terms of feature extraction sophistication, adaptability to complex domain data, and generalization of classification models. This is particularly true for elderly patients with cardiovascular disease and frailty, where existing technologies struggle to effectively address this issue. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, device and equipment for classifying dyspraxia under limited sample conditions, so as to solve the problems of insufficient sample size, limited feature extraction capability and insufficient model generalization in the existing technology when classifying dyspraxia in elderly patients with cardiovascular disease and frailty.

[0006] An embodiment of the present invention provides a method for classifying phobia under limited sample conditions, which includes:

[0007] S101, reading the patient's multimodal behavioral and physiological signal data, and extracting feature representations for latent class analysis from the data using the Bayesian Information Criterion combined with the Maximum Expectation Algorithm;

[0008] S102, inputting the feature representation into a latent class analysis module, determining the optimal number of classes and boundaries through a hierarchical clustering method, and generating candidate subclusters;

[0009] S103, for the candidate subclusters, re-evaluate the likelihood of each subclusters by introducing information theory enhancement technology to obtain a confidence score for each subclusters;

[0010] S104, based on the confidence score of each subcluster, an improved density peak clustering algorithm is used to remove redundant subclustering to obtain the final subclustering;

[0011] S105 , inputting the final sub-clusters into a trained classification module to estimate the kinesiophobia category of each sub-clusters, thereby achieving accurate kinesiophobia classification according to the category.

[0012] Preferably, step S101 specifically includes:

[0013] Use the feature embedding layer guided by prior knowledge to map the input multimodal data into a high-dimensional feature space and obtain the initial feature tensor R D×N×T ; Where D is the original feature dimension, corresponding to the sampling point value of behavioral signals and physiological signals, N is the number of samples, and T is the length of the time series;

[0014] The initial feature tensor is optimized by a maximum expectation algorithm based on the Bayesian information criterion to obtain a latent class feature representation; wherein the optimization process includes iteratively updating latent variables and parameter estimates, and calculating information gain after each iteration until the convergence condition is met; the optimization process also includes sparse constraints on the feature representation to reduce noise interference.

[0015] Preferably, the extraction process of the latent class feature representation is expressed as:

[0016]

[0017] Among them, θ is the feature representation of the latent class, L(θ) represents the objective function, is the prior probability of the kth latent class, is the posterior probability of the kth potential class, K is the optimal number of classes, represents the i-th sample, is the probability density function of the Gaussian distribution, is a sparse regularization term used to control the sparsity of feature representation.

[0018] Preferably, step S102 specifically includes:

[0019] The latent class feature representation is used as input to construct a tree structure using a hierarchical clustering algorithm. The hierarchical clustering algorithm is based on the Ward variance minimization criterion and gradually merges clusters with the highest similarity. The optimal number of classes K is determined by calculating the silhouette coefficient, and candidate subclusters are generated based on class boundaries. The silhouette coefficient is defined as:

[0020]

[0021] Among them, a( ) is a sample The average distance to other samples in its cluster, b( ) is a sample The average distance to the nearest neighbor cluster.

[0022] Preferably, step S103 specifically includes:

[0023] For each candidate subgroup, the mutual information score is calculated by information theory enhancement technology; where the mutual information score Defined as:

[0024]

[0025] Where X is the feature distribution of the candidate subpopulation, Y is the true label distribution, p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions;

[0026] The mutual information score is compared with the preset confidence threshold, and the possibility of each subpopulation is re-evaluated to obtain the confidence score of each subpopulation.

[0027] Preferably, in step S104, an improved density peak clustering algorithm is used to remove redundant subgroups; wherein: first, the local density ρ of each subgroup is calculated i and the relative distance δ i , and select the core subpopulation according to the decision diagram; where the local density is defined as:

[0028]

[0029] in, represents the distance between the i-th subgroup and the j-th subgroup, scales representing subpopulations;

[0030] The relative distance is defined as:

[0031]

[0032] Then, the non-core subclusters are merged or eliminated to obtain the final subclusters.

[0033] Preferably, step S105 specifically includes: aligning the time series features of different subgroups along the time dimension using a dynamic time warping algorithm; flattening the aligned features and performing nonlinear transformation using a multi-layer perceptron network to estimate the probability distribution of the dysphoria category, thereby predicting the dysphoria category of each subgroup; the classification process is defined as:

[0034]

[0035] in, and b k are the weights and biases for the kth class, is the feature representation of the input sample, P(y=k| ) is the probability that the input sample belongs to category k.

[0036] Preferably, for the classification module, three loss functions are used to optimize its learning parameters, including category balance loss, boundary smoothing loss and classification consistency loss.

[0037] An embodiment of the present invention further provides a device for classifying phobia under limited sample conditions, comprising:

[0038] a feature extraction unit for reading the patient's multimodal behavioral and physiological signal data and extracting feature representations for latent class analysis from the data using the Bayesian Information Criterion combined with the maximum expectation algorithm; the feature representations include timestamps and spatial distribution characteristics;

[0039] a candidate subcluster generation unit, configured to input the feature representation into a latent class analysis module, determine the optimal number of classes and boundaries through a hierarchical clustering method, and generate candidate subclusters;

[0040] a confidence score calculation unit, configured to re-evaluate the likelihood of each candidate subcluster by introducing information theory enhancement technology to obtain a confidence score for each subcluster;

[0041] The final subcluster generation unit is used to remove redundant subclusterings according to the confidence score of each subcluster by using an improved density peak clustering algorithm to obtain the final subclustering;

[0042] The kinesiophobia classification unit is used to input the final subclusters into the trained classification module to estimate the kinesiophobia category of each subclusters, thereby achieving accurate kinesiophobia classification according to the category.

[0043] An embodiment of the present invention further provides a device for classifying kinesia under limited sample conditions, which includes a memory and a processor. The memory stores a computer program, and the computer program can be executed by the processor to implement the above-mentioned method for classifying kinesia under limited sample conditions.

[0044] Compared with the prior art, the present invention has at least the following advantages:

[0045] 1) By introducing prior knowledge and information theory enhancement technology, this paper significantly improves the refinement of feature extraction under limited sample conditions, which can effectively address the problem of insufficient sample size for akinesia classification in elderly patients with cardiovascular disease and frailty;

[0046] 2) This paper designs and implements an optimization strategy based on the Bayesian Information Criterion and the Maximum Expectation Algorithm. Through latent class analysis and hierarchical clustering methods, it identifies potential subgroups of patients and determines the optimal number of classes and boundaries, thereby improving classification accuracy and model generalization ability.

[0047] 3) This paper proposes an improved density peak clustering algorithm, which effectively removes redundant subgroups by combining local density and relative distance, thereby improving the robustness and reliability of classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flowchart of a method for classifying phobia under limited sample conditions provided by the first embodiment of the present invention.

[0049] Figure 2 This is a schematic structural diagram of a device for classifying dyspraxia under limited sample conditions provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0050] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0051] This invention provides a method and system for classifying kinesia under limited sample conditions. Its core approach is to address the challenges of insufficient sample size, limited feature extraction capabilities, and insufficient model generalization when classifying kinesia in elderly patients with cardiovascular disease and frailty through feature extraction from multimodal data, latent class analysis, subpopulation generation and optimization, and the design of a final classification module. The following describes specific embodiments of the invention in detail, with reference to the accompanying drawings.

[0052] See also Figure 1 A first embodiment of the present invention provides a method for classifying kinesia under limited sample conditions, which can be performed by a kinesia classification device under limited sample conditions (hereinafter referred to as a classification device), and in particular, by one or more processors within the classification device, to implement the following method:

[0053] S101, reading the patient's multimodal behavior and physiological signal data, and using the Bayesian Information Criterion combined with the Maximum Expectation Algorithm to extract feature representations for latent class analysis from the data; the feature representations include timestamps and spatial distribution characteristics.

[0054] In practical applications, this embodiment mainly processes multimodal behavioral and physiological signal data of elderly patients with cardiovascular disease and frailty to achieve accurate classification of akinesia.

[0055] First, the patient's multimodal data must be read, including but not limited to behavioral and physiological signals such as heart rate, blood pressure, gait parameters, and body position changes. The raw data is then mapped to a high-dimensional feature space, and the latent class feature representation is extracted using the Bayesian Information Criterion (BIC) combined with the Maximum Expectation (EM) algorithm. This process first uses a feature embedding layer guided by prior knowledge to convert the input data into an initial feature tensor R. D×N×T , where D represents the original feature dimension, N is the number of samples, and T is the length of the time series. The initial feature tensor is optimized using the maximum expectation algorithm based on BIC to obtain the latent class feature representation. The optimization objective function is defined as:

[0056]

[0057] Among them, θ is the feature representation of the latent class, L(θ) represents the objective function, is the prior probability of the kth latent class, is the posterior probability of the kth potential class, K is the optimal number of classes, represents the i-th sample, is the probability density function of the Gaussian distribution, is a sparse regularization term used to control the sparsity of feature representation. During the optimization process, latent variables and parameter estimation are completed through iterative updates, and information gain is calculated after each iteration until the convergence condition is met.

[0058] S102: Input the feature representation into a latent class analysis module, determine the optimal number of classes and boundaries through a hierarchical clustering method, and generate candidate subclusters.

[0059] In step S102, the classification device uses the extracted latent class feature representation as input and constructs a tree structure using a hierarchical clustering method. The hierarchical clustering algorithm is based on the Ward variance minimization criterion and gradually merges clusters with the highest similarity. To determine the optimal number of classes K, the system evaluates by calculating the silhouette coefficient, which is defined as:

[0060]

[0061] Among them, a( ) is a sample The average distance to other samples in its cluster, b( ) is a sample The average distance to the samples in the nearest neighbor cluster. Based on the calculation results, the system generates candidate subclusters and proceeds to the next step of processing.

[0062] S103 , for the candidate subclusters, re-evaluate the possibility of each subclusters by introducing information theory enhancement technology to obtain a confidence score for each subclusters.

[0063] In this embodiment, the likelihood of each candidate subgroup is re-evaluated by introducing information theory enhancement technology. Specifically, the system calculates the mutual information score for each candidate subgroup, which is defined as:

[0064]

[0065] Where X is the feature distribution of the candidate subpopulation, Y is the true label distribution, p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions.

[0066] The mutual information score is compared with a preset confidence threshold to obtain a confidence score for each subpopulation. This process can effectively identify subpopulations with higher confidence, providing a basis for subsequent redundancy removal.

[0067] S104, based on the confidence score of each subcluster, an improved density peak clustering algorithm is used to remove redundant subclustering to obtain the final subclustering;

[0068] In step S104, an improved density peak clustering algorithm is used to remove redundant subgroups. First, the local density ρ of each subgroup is calculated. i and the relative distance δ i , and select the core subpopulation according to the decision diagram; where the local density is defined as:

[0069]

[0070] in, represents the distance between the i-th subgroup and the j-th subgroup, scales representing subpopulations;

[0071] The relative distance is defined as:

[0072]

[0073] Then, the non-core subclusters are merged or eliminated to obtain the final subclusters.

[0074] In this way, through a comprehensive analysis of local density and relative distance, the system draws a decision diagram and selects core subclusters. For non-core subclusters, the system merges or removes them based on similarity, ultimately generating optimized subclusters.

[0075] S105 , inputting the final sub-clusters into a trained classification module to estimate the kinesiophobia category of each sub-clusters, thereby achieving accurate kinesiophobia classification according to the category.

[0076] In step S105, the final subclusters are fed into the trained classification module to estimate the dysphoria category for each subclusters. The classification module first aligns the time series features of the different subclusters along the temporal dimension using a dynamic time warping algorithm. The aligned features are then flattened and nonlinearly transformed using a multilayer perceptron network. The classification process uses a softmax activation function to estimate the probability distribution of the dysphoria category. The classification process is defined as: ;

[0077] in, and b k are the weights and biases for the kth class, is the feature representation of the input sample, P(y=k| ) is the probability that the input sample belongs to category k.

[0078] To further improve classification performance, the system uses three loss functions to optimize learning parameters, including category balance loss, boundary smoothing loss, and category consistency loss; category balance loss is defined as:

[0079]

[0080] Among them, α k is the class weight; the boundary smoothing loss is defined as:

[0081]

[0082] The classification consistency loss is defined as:

[0083]

[0084] Among them, p i is the generated category label, are pseudo labels generated through data augmentation.

[0085] In actual application scenarios, the embodiments of the present invention have been successfully applied to a group of elderly cardiovascular patients in a hospital. Experimental data show that by introducing prior knowledge and information theory enhancement technology, the system can significantly improve the refinement of feature extraction under limited sample conditions, and effectively deal with the problem of insufficient sample size. At the same time, based on the optimization strategy of the Bayesian Information Criterion and the Maximum Expectation Algorithm, combined with latent class analysis and hierarchical clustering methods, it can accurately identify the potential subgroups of patients and determine the optimal number of categories and boundaries, thereby improving classification accuracy and model generalization ability. In addition, the improved density peak clustering algorithm effectively removes redundant subgroups by combining local density and relative distance, thereby improving the robustness and reliability of the classification results.

[0086] In summary, this embodiment solves the key problems of the existing technology in the classification of phobia under limited sample conditions through a series of innovative technical means, and provides strong technical support for the precision medicine of elderly patients with cardiovascular disease and frailty.

[0087] See also Figure 2 The second embodiment of the present invention provides a device for classifying phobia under limited sample conditions, which includes:

[0088] A feature extraction unit 210 is configured to read the patient's multimodal behavioral and physiological signal data and extract feature representations for latent class analysis from the data using the Bayesian Information Criterion combined with the Expectation Maximum Algorithm; the feature representations include timestamps and spatial distribution characteristics;

[0089] a candidate sub-cluster generation unit 220 for inputting the feature representation into a latent class analysis module, determining the optimal number of classes and boundaries through a hierarchical clustering method, and generating candidate sub-clusters;

[0090] A confidence score calculation unit 230 is configured to re-evaluate the likelihood of each candidate sub-cluster by introducing information theory enhancement technology to obtain a confidence score for each sub-cluster;

[0091] The final subcluster generating unit 240 is configured to remove redundant subclusterings based on the confidence score of each subcluster using an improved density peak clustering algorithm to obtain the final subclusterings;

[0092] The kinesiophobia classification unit 250 is configured to input the final sub-clusters into a trained classification module to estimate the kinesiophobia category of each sub-clusters, thereby achieving accurate kinesiophobia classification according to the category.

[0093] A third embodiment of the present invention further provides a device for classifying kinesia under limited sample conditions, comprising a memory and a processor. The memory stores a computer program, and the computer program can be executed by the processor to implement the method for classifying kinesia under limited sample conditions as described in any of the above embodiments.

[0094] In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of the devices, methods, and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or part of the code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or actions, or can be implemented using a combination of dedicated hardware and computer instructions.

[0095] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0096] If the functions are implemented in the form of software modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which can be a personal computer, electronic device, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or device. Without further constraints, an element defined by the phrase "comprises a..." does not preclude the existence of additional identical elements in the process, method, article or apparatus that includes the element.

[0097] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for classifying phobia under limited sample conditions, characterized in that: include: S101, reading the patient's multimodal behavior and physiological signal data, and extracting feature representations for latent class analysis from the data using the Bayesian Information Criterion combined with the Maximum Expectation Algorithm; wherein the extraction process of the feature representations of the latent classes is expressed as follows: Among them, θ is the feature representation of the latent class, L(θ) represents the objective function, is the prior probability of the kth latent class, is the posterior probability of the kth potential class, K is the optimal number of classes, represents the i-th sample, is the probability density function of the Gaussian distribution, is a sparse regularization term used to control the sparsity of feature representation; S102, inputting the feature representation into a latent class analysis module, determining the optimal number of classes and boundaries through a hierarchical clustering method, and generating candidate subclusters; S103, for the candidate subclusters, re-evaluate the likelihood of each subclusters by introducing information theory enhancement technology to obtain a confidence score for each subclusters; S104, based on the confidence score of each subcluster, an improved density peak clustering algorithm is used to remove redundant subclustering to obtain the final subclustering; S105: Input the final subclusters into the trained classification module to estimate the kinophobia category of each subclusters, thereby achieving accurate kinophobia classification based on the category. Specifically, the dynamic time warping algorithm is used to align the time series features of different subclusters along the time dimension; the aligned features are flattened and nonlinearly transformed using a multi-layer perceptron network to estimate the probability distribution of kinophobia categories, thereby predicting the kinophobia category of each subclusters. The classification process is defined as: in, and b k are the weights and biases for the kth class, is the feature representation of the input sample, P(y=k| ) is the probability that the input sample belongs to category k.

2. The method for classifying dyspraxia under limited sample conditions according to claim 1, characterized in that: Step S101 specifically includes: Use the feature embedding layer guided by prior knowledge to map the input multimodal data into a high-dimensional feature space and obtain the initial feature tensor R D×N×T ; Where D is the original feature dimension, corresponding to the sampling point value of behavioral signals and physiological signals, N is the number of samples, and T is the length of the time series; The initial feature tensor is optimized by a maximum expectation algorithm based on the Bayesian information criterion to obtain a latent class feature representation; wherein the optimization process includes iteratively updating latent variables and parameter estimates, and calculating information gain after each iteration until the convergence condition is met; the optimization process also includes sparse constraints on the feature representation to reduce noise interference.

3. The method for classifying dyskinesia under limited sample conditions according to claim 1, characterized in that: Step S102 specifically includes: taking the latent class feature representation as input, constructing a tree structure through a hierarchical clustering algorithm; wherein the hierarchical clustering algorithm is based on the Ward variance minimization criterion, gradually merging clusters with the highest similarity; determining the optimal number of categories K by calculating the silhouette coefficient, and generating candidate subclusters based on category boundaries; wherein the silhouette coefficient is defined as: Among them, a( ) is a sample The average distance to other samples in its cluster, b( ) is a sample The average distance to the nearest neighbor cluster.

4. The method for classifying dyskinesia under limited sample conditions according to claim 1, characterized in that: Step S103 specifically includes: for each candidate subgroup, calculating its mutual information score by using information theory enhancement technology; wherein the mutual information score Defined as: Where X is the feature distribution of the candidate subpopulation, Y is the true label distribution, p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions; The mutual information score is compared with the preset confidence threshold, and the possibility of each subpopulation is re-evaluated to obtain the confidence score of each subpopulation.

5. The method for classifying dyskinesia under limited sample conditions according to claim 1, characterized in that: In step S104, an improved density peak clustering algorithm is used to remove redundant subgroups; wherein: first, the local density ρ of each subgroup is calculated i and the relative distance δ i , and select the core subpopulation according to the decision diagram; where the local density is defined as: in, represents the distance between the i-th subgroup and the j-th subgroup, scales representing subpopulations; The relative distance is defined as: Then, the non-core subclusters are merged or eliminated to obtain the final subclusters.

6. The method for classifying dyspraxia under limited sample conditions according to claim 1, characterized in that: For the classification module, three loss functions are used to optimize its learning parameters, including category balance loss, boundary smoothing loss and classification consistency loss.

7. A device for classifying phobia under limited sample conditions, characterized in that: include: A feature extraction unit is used to read the patient's multimodal behavior and physiological signal data, and use the Bayesian Information Criterion combined with the Maximum Expectation Algorithm to extract feature representations for latent class analysis from the data; the feature representations include timestamps and spatial distribution characteristics; wherein the extraction process of the feature representations of the latent classes is expressed as follows: Among them, θ is the feature representation of the latent class, L(θ) represents the objective function, is the prior probability of the kth latent class, is the posterior probability of the kth potential class, K is the optimal number of classes, represents the i-th sample, is the probability density function of the Gaussian distribution, is a sparse regularization term used to control the sparsity of feature representation; a candidate subcluster generation unit, configured to input the feature representation into a latent class analysis module, determine the optimal number of classes and boundaries through a hierarchical clustering method, and generate candidate subclusters; a confidence score calculation unit, configured to re-evaluate the likelihood of each candidate subcluster by introducing information theory enhancement technology to obtain a confidence score for each subcluster; The final subcluster generation unit is used to remove redundant subclusterings according to the confidence score of each subcluster by using an improved density peak clustering algorithm to obtain the final subclustering; A kinephobia classification unit is configured to input the final subclusters into a trained classification module to estimate the kinephobia category of each subclusters, thereby achieving accurate kinephobia classification based on the category; align the time series features of different subclusters along the time dimension using a dynamic time warping algorithm; flatten the aligned features and perform a nonlinear transformation using a multilayer perceptron network to estimate the probability distribution of kinephobia categories, thereby predicting the kinephobia category of each subclusters; the classification process is defined as: in, and b k are the weights and biases for the kth class, is the feature representation of the input sample, P(y=k| ) is the probability that the input sample belongs to category k.

8. A device for classifying phobia under limited sample conditions, characterized in that: The system comprises a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the method for classifying kinesia under limited sample conditions according to any one of claims 1 to 6.

Images