Autism spectrum disorder diagnosis method based on fuzzy modeling and multi-label learning

Abstract

The application belongs to the field of intelligent medical computer-aided diagnosis, and relates to a method for diagnosing autism spectrum disorder based on fuzzy modeling and multi-label learning. The method comprises four parts of preprocessing, feature extraction, training stage and use stage. The rs-fMRI image of the autism spectrum disorder sample is processed into digital feature data suitable for method training and use through preprocessing and feature extraction. The training stage comprises sample synthesis, fuzzy uncertain learning, label collaborative learning, linear regression weighting and high-order functional connectivity learning. The combination of sample synthesis and linear regression weighting can effectively alleviate the imbalance between autism spectrum disorder subtypes, label collaborative learning reduces the interference of label noise on method prediction by analyzing the correlation between each autism spectrum disorder subtype, and high-order functional connectivity learning further excavates potential knowledge in the second-order functional connectivity features of autism spectrum disorder based on label prior knowledge.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical computer-aided diagnosis application, and relates to a diagnostic method for autism spectrum disorder based on fuzzy modeling and multi-label learning. Technical Background

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by social impairment, limited interests, and repetitive behaviors. Based on the theory of brain plasticity, early diagnosis, intervention, and treatment of ASD are crucial for altering the developmental trajectory of individuals with ASD. However, the etiology of ASD is complex, and there are no definitive laboratory biochemical indicators for diagnosis. Furthermore, ASD includes subtypes such as autism, Asperger's syndrome, and unclassified pervasive developmental disorder, which lie on a continuous spectrum of symptom severity from mild to severe (autism being at the most severe level), without clear demarcation lines. These realities pose a significant challenge to the clinical diagnosis of ASD. Therefore, computer-aided diagnosis, due to its efficiency and reproducibility, has become a research hotspot in assisting the diagnosis of ASD.

[0003] Current mainstream computer-aided diagnosis of ASD is based on resting-state functional magnetic resonance imaging (rs-fMRI). Unlike behavioral analysis-based autism screening scales and autism diagnostic observation scales, rs-fMRI assesses the functional connectivity of the entire brain by tracking intrinsic fluctuations in blood oxygen level-dependent signals. In other words, rs-fMRI reveals the brain's inherent spontaneous activity patterns, functional connectivity patterns, and brain network topological characteristics, making it a direct and non-invasive tool for measuring ASD. Because rs-fMRI only requires the patient to remain relaxed with their eyes closed without administering numerous instructions, and does not require the involvement of guardians or other observers, it reduces errors in the ASD data acquisition process to some extent. Therefore, it is more popular in the field of computer-aided ASD diagnosis.

[0004] Existing methods have effectively promoted the development of computer-aided diagnosis of ASD. However, almost no method simultaneously addresses the four common challenges faced in computer-aided ASD diagnosis. First, class imbalance between ASD subclasses. This is prevalent during data acquisition. Second, rs-fMRI images contain uncertain information representations. This is mainly due to the instability of acquisition equipment, suboptimal resting conditions, the complexity of pathogenesis, and interference from inherent physiological noise. Third, labeling noise exists in ASD datasets. This is primarily because the boundaries between ASD subclasses are unclear, so mislabeled ASD subclasses are likely to exist in manually labeled ASD datasets. Fourth, limited knowledge mining of ASD datasets. Data preprocessing can only utilize limited methods to reduce the dimensionality of rs-fMRI images into feature data friendly to intelligent models, which to some extent limits the learning strength of intelligent models for ASD knowledge. Generally, these four challenges occur simultaneously in computer-aided ASD diagnosis. Therefore, there is an urgent need to develop an ASD-aided diagnostic method that can address all four challenges. Summary of the Invention

[0005] The purpose of this invention is to provide a diagnostic method for autism spectrum disorder based on fuzzy modeling and multi-label learning, which addresses class imbalance, information uncertainty, and label noise.

[0006] The specific implementation steps are as follows:

[0007] Step 1: Based on autism spectrum disorder image data from six imaging centers, the rs-fMRI images were processed using the standard tool DPARSF provided by ABIDE. The brain was divided into 116 regions of interest according to the automatically anatomically labeled AAL template, and the rs-fMRI images were processed into 116 corresponding average time series.

[0008] The second step involves further feature extraction based on the obtained average time series. Using the Pearson correlation coefficient, this invention measures the correlation between any two average time series, obtaining a 116×116 second-order functionally connected feature matrix. In fact, this second-order functionally connected feature matrix is ​​a symmetric matrix, with half of the second-order functional connectivity information being redundant. Therefore, this invention extracts only the elements of the upper triangular matrix from the second-order functionally connected feature matrix as feature vectors. That is, in this invention, the feature dimension of the ASD dataset is 6670.

[0009] Step 3: Perform SMOTE sample synthesis on the imbalanced ASD dataset. The specific process is as follows:

[0010] (1) Input ASD training data The nearest neighbor number k, the sample generation ratio n.

[0011] (2) Initialization

[0012] (3) Find the sample subsets that are associated with each of the L ASD subclass labels. N i N is the number of relevant samples for the i-th ASD subclass label. max ←max N i (1≤i≤L).

[0013] (4) For each ASD tag

[0014] If N i >0 and N i <N max

[0015] 1) In China Find k nearest neighbors

[0016] 2) If the n-fold generation ratio is not satisfied

[0017] a) From Randomly selected samples To synthesize new samples (x) new ,y new ).Right now, δ and It is a random vector in which each element is in the range [0, 1]. This represents the Hadamard product.

[0018] b)

[0019] The training set after sample synthesis is denoted as This invention allows N tr =N+N new , as well as

[0020] Step 4: First, transform the ASD subclass output data into a discrete and clearly defined ASD label output matrix, and then perform fuzzy uncertainty learning. Fuzzy uncertainty learning establishes a fuzzy relation matrix of "ASD instance-ASD cluster" sequentially based on fuzzy logic inference criteria. ASD Clustering-ASD Labeling Fuzzy Relationship Matrix ASD Instances - ASD Tags Fuzzy Relationship Matrix The obtained "ASD instance-ASD tag" relationship matrix is ​​then integrated into a clear tag output matrix to obtain an output matrix containing fuzzy and uncertain information. The specific process is as follows:

[0021] (1) FM_ASD uses fuzzy C-means clustering to cluster ASD instances into C abstract classes and constructs an "ASD instance-ASD cluster" relationship matrix. Each instance corresponds to a fuzzy membership vector belonging to C abstract classes. This vector analyzes the uncertainty of ASD data from the perspective of fuzzy clustering. For ASD instances... Its fuzzy membership value for belonging to the c-th (1≤c≤C) cluster is:

[0022]

[0023]

[0024] in This represents the fuzzy membership degree of the i-th ASD instance belonging to the c-th class. It is the cluster center of the c-th class. m>1 is the fuzzy index. Solving iteratively... and v c The optimal fuzzy membership value of the i-th ASD instance belonging to the c-th class This can be obtained, and further, the fuzzy membership matrix can be obtained. Optimal fuzzy membership vector This reflects the expression uncertainty of the i-th ASD instance.

[0025] (2) Based on the fuzzy membership vector of ASD samples, FM_ASD constructs a fuzzy relation matrix of "fuzzy clustering-ASD label". Specifically, for the j-th row vector F j , (1≤j≤L), which integrates the fuzzy membership vectors of all relevant instances on the j-th ASD tag. That is:

[0026]

[0027]

[0028] Among them 0 1×C This represents the zero vector in row C. In fact, F is a global uncertainty analysis. It is a global fuzzy relation between the fuzzy clusters built upon all ASD instances and the ASD labels.

[0029] (3) Based on the fuzzy logic criterion, FM_ASD transforms the fuzzy relation matrix F of "fuzzy clustering-ASD tag" into the fuzzy relation matrix of "ASD instance-ASD tag". For element B in row i and column j of B ij (1≤i≤L,1≤j≤N tr Its calculation formula is:

[0030]

[0031] Where F ic It is the element in the i-th row and c-th column of F. is the fuzzy membership degree of the j-th ASD instance in the c-th cluster, ∧ represents the smaller operation in the fuzzy matrix synthesis operation, and ∨ represents the larger operation in the fuzzy matrix synthesis operation. In fact, formula (5) is the fuzzy matrix synthesis operation, which can realize the reasoning modeling of fuzzy relations.

[0032] (4) FM_ASD integrates the fuzzy relation matrix B of "ASD instance-ASD tag" into the ASD tag matrix. Obtain the new tag matrix The "ASD instance-ASD tag" fuzzy relation matrix B implicitly contains the distribution of uncertain information in the ASD data. Therefore, compared to the original clear set tag matrix... It better reflects the distribution characteristics of real ASD data. The specific calculation formula is as follows:

[0033]

[0034] The establishment of fuzzy uncertainty learning for ASD data transforms the originally fuzzy and uncertain ASD data into clear and measurable data. It quantifies the uncertainty in the representation of ASD data based on fuzzy inference modeling. This helps FM_ASD mitigate the impact of uncertain information representation in ASD data to some extent.

[0035] Step 5: In the output matrix containing fuzzy and uncertain information Based on this, perform labeled collaborative learning. Optimize the labeled collaborative learning terms. This establishes the association between individual tags and global tags. The specific process is as follows:

[0036] (1) Let N in the ASD dataset be N tr The label vector formed by the training samples on the l-th label is: In order to reduce To address the interference of labeled noise on subsequent FM_ASD learning of ASD, FM_ASD proposes an explicit mapping vector for the labels. To analyze the impact of all ASD tags on the l-th tag, that is,

[0037]

[0038] Where m l The j-th (1≤j≤L) element m in lj This represents the influence of the j-th ASD tag on the l-th ASD tag. The purpose of minimizing formula (7) is to maintain the consistency of the l-th tag's expression before and after explicit mapping as much as possible when analyzing the effect of all tags on the l-th ASD tag.

[0039] (2) The complete label-based collaborative learning is as follows:

[0040]

[0041] in, The first term in Equation (8) represents the collaborative learning among ASD labels, and the second term represents the regularization term with respect to M. α and γ are two hyperparameters that balance the effects of collaborative learning among ASD labels and model complexity, respectively. Tr(·) represents the trace of the matrix.

[0042] In fact, explicit mapping of ASD label noise involves all ASD training samples. Therefore, the proposed label co-learning is based on a global implementation, which can effectively reduce the impact of label noise in the ASD dataset on FM_ASD's learning of ASD subclasses.

[0043] Step 6: Based on the difference in the number of relevant samples for each label, assign a unique linear regression weight to the linear regression process for each label, with a larger regression weight for smaller relevant samples. The specific process is as follows:

[0044] (1) For the l-th (1≤l≤L) ASD label, its weighted linear regression is defined as:

[0045]

[0046] in It is the regression vector of the l-th label, which can identify the corresponding discriminant features for the l-th ASD label. It is the representation of the l-th label based on label-based collaborative learning, implemented according to formula (7). l It is the linear regression weight corresponding to the l-th label, and

[0047]

[0048] z l =o l / p (11)

[0049]

[0050]

[0051] Where |·| represents the number of elements in the set. Indicates ASD training set The value of the nth sample on the lth label. Formulas (10)-(13) show that the fewer the number of relevant samples for the lth ASD label, the higher h l The larger the weight value, the better.

[0052] (2) The weighted linear regression of all ASD labels can be expressed as:

[0053]

[0054] in It is a diagonal matrix.

[0055] (3) This invention introduces the L1 regularization term into formula (14) to constrain the complexity of the model, that is...

[0056]

[0057] Here, λ is a hyperparameter used to control the impact of model complexity. The introduction of the L1 norm makes the discriminative features learned by FM_ASD for each ASD label more accurate.

[0058] While sample synthesis mechanisms can alleviate the label imbalance problem in the ASD dataset, they cannot achieve complete balance among the ASD labels. Therefore, this invention proposes a weighting mechanism based on linear regression to further reduce ASD decision bias caused by data imbalance.

[0059] Step 7: Construct label co-learning based on prior knowledge of the labels. That is, use the correlation between any two labels to construct the association between their discriminative features, thereby effectively mining hidden information in ASD data.

[0060] The specific process is as follows:

[0061] (1) The implementation of higher-order functional connectivity learning is mainly based on the following two arguments:

[0062] a) Regression vector It can be used as a discriminative feature to identify the i-th (1≤i≤L) ASD tag. id The larger the value of (1≤d≤D), the greater the contribution of the d-th second-order functional connectivity feature to the identification of the i-th ASD tag.

[0063] b) The correlation between any two ASD tags is consistent with the correlation between their discriminative features. That is, in the ASD dataset, the correlation between two ASD tag vectors (such as the i-th (1≤i≤L) ASD tag vector) is consistent. and the j-th (1≤j≤L) ASD tag vector The smaller the similarity between their regression vectors (w and w), the smaller the similarity between their ASD discriminant features, which means that their regression vectors (w and w) are similar. i =[w i1 ,w i2 ,…,w iD ] T and w j =[w j1 ,w j2 ,…,w jD ] T The smaller the similarity between them, the less similar they are to each other.

[0064] (2) For any two label vectors in the ASD dataset and Their higher-order functional connectivity learning can be defined as:

[0065]

[0066] Where r ij =1-s ij s ij Used to measure the i-th ASD tag and the j-th ASD tag The similarity between them. This invention uses cosine similarity to calculate s. ij . and The smaller the similarity, the better. ij The smaller the value of r, the better. ij The larger the value, the better.

[0067] (3) The higher-order function connectivity learning for all ASD tags is as follows:

[0068]

[0069] in β is a hyperparameter used to adjust the impact of higher-order functional connectivity learning.

[0070] Step 8: By integrating the above seven steps, a complete diagnostic objective function for autism spectrum disorder based on fuzzy modeling and multi-label learning is achieved. It can still effectively predict autism spectrum disorder even in scenarios with class imbalance, information uncertainty, and label noise.

[0071] The advantages of this invention include the following:

[0072] 1) Combination of sample synthesis and linear regression weighting. Sample synthesis and linear regression weighting reduce the impact of ASD subclass imbalance from the perspectives of data and algorithm, respectively, thereby effectively reducing decision bias in ASD-assisted diagnostic tasks.

[0073] 2) Construction of fuzzy uncertainty learning. This project utilizes fuzzy membership to quantify the uncertainty of ASD feature representation, and then integrates the obtained fuzzy uncertainty information into the ASD output matrix based on fuzzy inference, thereby realizing the quantitative analysis of uncertainty information in rs-fMRI images.

[0074] 3) Design of label collaborative learning. This project comprehensively analyzes the interaction between ASD labels by modeling the correlation between individual labels and global labels, thereby reducing the impact of ASD label noise (i.e., some ASD samples are mislabeled).

[0075] 4) Construction of higher-order functional connectivity learning. This project utilizes the correlation between any two ASD tags to construct the association between their second-order functional connectivity features, thereby effectively learning the discriminative features of each ASD tag and realizing further mining of the latent knowledge of ASD.

[0076] 5) Construction of a diagnostic method for autism spectrum disorder based on fuzzy modeling and multi-label learning. By integrating sample synthesis, fuzzy uncertainty learning, linear regression weighting, label co-learning, and higher-order functional connectivity learning, a diagnostic method for autism spectrum disorder oriented towards class imbalance, information uncertainty, and label noise was realized, thereby effectively predicting autism spectrum disorder.

[0077] 6) Extensive experimental analysis demonstrates that the present invention is more effective in assisting diagnosis than existing related work on six publicly available benchmark ASD datasets. Attached Figure Description

[0078] Figure 1 This is a general framework diagram of the method of the present invention.

[0079] Figure 2 This is a diagram of the image preprocessing and feature extraction framework of the present invention.

[0080] Figure 3 This is a diagram of the label-based collaborative learning framework of the present invention.

[0081] Figure 4 This is a high-order functional connectivity learning framework diagram of the present invention.

[0082] Figure 5 This is a diagram illustrating the effectiveness analysis of fuzzy uncertainty learning in this invention.

[0083] Figure 6This is a diagram illustrating the effectiveness analysis of the label-based collaborative learning method of this invention.

[0084] Figure 7 This is a graph showing the linear regression weighted effectiveness analysis of the present invention.

[0085] Figure 8 This is a high-order functional connectivity learning effectiveness analysis diagram of the present invention. Detailed Implementation

[0086] The present invention will now be described in detail with reference to the accompanying drawings and embodiments:

[0087] This invention implements a diagnostic method for autism spectrum disorder based on fuzzy modeling and multi-label learning, addressing issues such as class imbalance, uncertain information representation, and label noise. The method comprises four parts: preprocessing, feature extraction, training, and testing. The framework of this invention is as follows: Figure 1 As shown in the diagram. The preprocessing section divides the rs-fMRI images into 116 average time series based on 116 brain regions of interest defined by the AAL template. The feature extraction section first analyzes the correlation between the 116 average time series to obtain a symmetric second-order functional connectivity feature matrix, and then extracts upper triangular matrix elements from the second-order functional connectivity feature matrix as the feature input for the method. Preprocessing and feature extraction are as follows: Figure 2 As shown. The training part includes sample synthesis, fuzzy uncertainty learning, label co-learning, linear regression weighting, and higher-order functional connectivity learning. Sample synthesis generates new samples for the imbalanced ASD dataset, thereby alleviating the problem of ASD subclass imbalance. Fuzzy uncertainty learning uses fuzzy inference to model and quantify the expression uncertainty in ASD feature data. Label co-learning reduces the interference of label noise on the method's learning of ASD knowledge by establishing associations between individual labels and global labels. This framework is as follows. Figure 3 As shown, linear regression weighting assigns a unique regression weight to each ASD label based on the relevant sample size. Linear regression weighting further reduces the impact of imbalanced ASD subclass distributions on the basis of sample synthesis. Higher-order functional connectivity learning enhances the learning of discriminative features by analyzing the correlation between any two labels, thereby effectively uncovering latent ASD knowledge and improving the learning performance of the method. This framework is as follows: Figure 4 As shown.

[0088] Specific steps in the training phase.

[0089] Example 1

[0090] Following the implementation method of the training phase, ASD datasets provided by six imaging centers were used. In assisting the diagnosis of ASD, this invention first preprocesses and extracts features from the rs-fMRI images in the ASD dataset, then trains the model, and finally tests it on a test set to verify the effectiveness of the invention. Specific information about the ASD dataset is shown in Table 1.

[0091] Table 1: Detailed information on ASD datasets provided by the six imaging centers

[0092]

[0093] This invention uses three commonly used evaluation metrics: sensitivity (SEN), accuracy (ACC), and macro-average precision (mAP), and one multi-label evaluation metric: average precision (AP). This invention is compared with six other methods on six publicly available ASD datasets, all employing a five-fold cross-validation strategy. Experimental results are shown in Tables 2-7.

[0094] The results in the table show that the overall performance of this invention is superior to the six comparative methods on six ASD datasets. The main reason is that this invention, in assisting the diagnosis of ASD data, proposes sample synthesis and linear weighted regression to alleviate the problem of ASD subclass imbalance in ASD datasets, uses fuzzy uncertainty learning to measure the uncertainty of ASD feature data representation, reduces the interference of ASD label noise through label co-learning, and further mines hidden knowledge in ASD data through higher-order functional connectivity learning.

[0095] Table 2: Experimental results of different methods on the KKI dataset (SEN-ACC-mAP):

[0096]

[0097]

[0098] Table 3: Experimental results of different methods on the Leuven dataset (SEN-ACC-mAP):

[0099]

[0100] Table 4: Experimental results of different methods on the NYU dataset (SEN-ACC-mAP):

[0101]

[0102] Table 5: Experimental results of different methods on the UCLA dataset (SEN-ACC-mAP):

[0103]

[0104]

[0105] Table 6: Experimental results of different methods on the UM dataset (SEN-ACC-mAP):

[0106]

[0107] Table 7: Experimental results of different methods on the USM dataset (SEN-ACC-mAP):

[0108]

[0109] Example 2

[0110] To verify the effectiveness of fuzzy uncertainty learning, this invention conducted comparative experiments on six ASD datasets, namely Group A and Group B. Specifically, this invention obtained the optimal parameters of the invented FM_ASD method on the six ASD datasets based on the AP metric. For each dataset, the hyperparameters of Group A and Group B were set as the corresponding optimal hyperparameters. It should be noted that in Group A, fuzzy uncertainty learning is effective, that is, the training of FM_ASD includes fuzzy uncertainty learning. In Group B, fuzzy uncertainty learning is ineffective, that is, the training of FM_ASD does not consider fuzzy uncertainty learning. Figure 5 The comparison results of Group A and Group B on six ASD datasets are presented. Figure 5 Experimental results show that the construction of fuzzy uncertainty learning can effectively analyze the uncertainty of ASD data representation, thereby improving the performance of FM_ASD in ASD-assisted diagnosis tasks.

[0111] Example 3

[0112] This invention uses the AP (Average Performance) metric as a benchmark and organizes two sets of experiments, Group A and Group B, on six ASD datasets to verify the effectiveness of label-based collaborative learning. The performance results of FM_ASD in Group A are all based on the optimal AP. Label-based collaborative learning is effective in Group A. In Group B, the hyperparameter α corresponding to label-based collaborative learning is set to 0, and the values ​​of other hyperparameters are the same as in Group A. That is, label-based collaborative learning is ineffective in Group B. Figure 6 The results of the comparative experiment are presented. Observation Figure 6 It can be observed that label-based collaborative learning in FM_ASD is crucial, as it can significantly reduce the negative impact of label noise on ASD-assisted diagnosis.

[0113] Example 4

[0114] To demonstrate the effectiveness of the weighting strategy, this invention conducted comparative analyses on six datasets. Specifically, in Group A, this invention assigned different weights to the regression of ASD markers based on the differences in the number of relevant samples for different ASD markers. In Group B, the regression weights for all ASD markers were set to 1. Furthermore, this invention set the optimal parameters of the invented FM_ASD method on each of the six datasets using the AP metric as a benchmark. The comparative results are as follows: Figure 7 As shown. Figure 7 The experimental results show that the weighted regression strategy involved in this invention is effective in improving the performance of ASD-assisted diagnosis.

[0115] Example 5

[0116] Higher-order functional connectivity learning (HFM) utilizes the correlation between any two labels in the ASD dataset to construct the association between their second-order functional connectivity features, thereby further analyzing the differences between the discriminative features of each ASD label and improving FM_ASD's learning ability for each ASD label's discriminative features. To verify the effectiveness of HFM, this invention organized two sets of experiments for comparative analysis. Specifically, this invention determined the optimal parameters of FM_ASD on six datasets based on the AP metric. It should be noted that the HFM mechanism is effective in Group A (i.e., β is a suitable non-zero hyperparameter) but ineffective in Group B (i.e., β = 0). Experimental results are as follows: Figure 8 As shown. Figure 8 The comparative experimental results show that the high-order functional connectivity learning method created in this invention can effectively improve the auxiliary diagnostic performance of FM_ASD for ASD.

Claims

1. A diagnostic method for autism spectrum disorder based on fuzzy modeling and multi-label learning, the steps of which are as follows: Step 1: Process rs-fMRI images using DPARSF; based on the automatic anatomical marker template, process the rs-fMRI images into 116 average time series corresponding to 116 brain regions of interest; Step 2: Perform feature extraction based on the average time series; use the Pearson correlation function to measure the correlation between any two average time series, obtaining a feature of size... The second-order functionally connected feature matrix is ​​obtained, and then the elements of the upper triangular matrix are extracted from the second-order functionally connected feature matrix to generate a 6670-dimensional feature vector. Step 3: Perform SMOTE sample synthesis on the imbalanced ASD dataset; the training set after sample synthesis is denoted as... ;make , ,as well as ; Step 4: First, transform the ASD subclass output data into a discrete and clearly defined ASD label output matrix, and then perform fuzzy uncertainty learning; fuzzy uncertainty learning establishes a "ASD instance-ASD cluster" fuzzy relation matrix sequentially based on fuzzy logic inference criteria. "ASD clustering-ASD tagging" fuzzy relation matrix "ASD Instance - ASD Tag" Fuzzy Relationship Matrix The obtained "ASD instance-ASD tag" relationship matrix is ​​then integrated into a clear tag output matrix to obtain an output matrix containing fuzzy and uncertain information. ; Step 5: In the output matrix containing fuzzy and uncertain information Based on this, perform labeled collaborative learning; Optimize the labeling of collaborative learning items To establish a connection between individual tags and global tags; Step 6: Based on the difference in the number of relevant samples for each label, assign a unique linear regression weight to the linear regression process for each label, with a larger regression weight for smaller relevant samples; Step 7: Construct label-based collaborative learning based on prior knowledge of the labels; that is, use the correlation between any two labels to construct the association between their discriminative features, thereby effectively mining hidden information in ASD data; Step 8: By integrating the above seven steps, a complete diagnostic objective function for autism spectrum disorder based on fuzzy modeling and multi-label learning is realized; it can still effectively predict autism spectrum disorder in scenarios with class imbalance, information uncertainty, and label noise. The fifth step, as described above, consists of the following steps: 5.1 Let the ASD dataset contain The training sample at the th ... The tag vector formed by the tags is In order to reduce To address the interference of labeled noise on subsequent FM_ASD learning of ASD, FM_ASD proposes an explicit mapping vector for the labels. To analyze all ASD tag pairs The effect of each marker, ,Right now, (7) in The first in element Representing the The ASD tag pair of the first The impact of ASD tags ; The purpose of minimizing formula (7) is to analyze all the pairs of labels. When the ASD tag is active, try to keep the first one as active as possible. Consistency in the representation of each tag before and after explicit mapping; 5.2 The complete label-based collaborative learning is as follows: (8) in, The first term of formula (8) represents collaborative learning among ASD labels, and the second term represents the learning about... The regularization term; and These are two hyperparameters that balance the impact of collaborative learning among ASD tags and model complexity, respectively. Represents the trace of a matrix.

2. The autism spectrum disorder diagnostic method based on fuzzy modeling and multi-label learning as described in claim 1, characterized in that: The third step, as described above, consists of the following steps: 3.1 Input ASD training data Nearest neighbor number k Sample generation ratio n ; 3.2 Initialization ; 3.3 Search for and respectively L A subset of samples associated with ASD subclass labels , , It is the first i The number of relevant samples for each ASD subclass label , ; 3.4 For each ASD tag if and 1) In China Search k The nearest residence , ; 2) If not satisfied n Multiplier a) From Randomly selected samples To synthesize new samples , ;Right now, and ; / / It is a random vector in which each element is in the range [0, 1]; ) represents the Hadamard product; b) 。 3. The autism spectrum disorder diagnostic method based on fuzzy modeling and multi-label learning as described in claim 2, characterized in that: The fourth step, as described above, consists of the following steps: 4.1FM_ASD utilizes fuzziness C Mean clustering clusters ASD instances into C An abstract class is used to construct an "ASD instance-ASD cluster" relationship matrix. ; Each instance corresponds to a member C A fuzzy membership vector for each abstract class, which analyzes the uncertainty of ASD data from the perspective of fuzzy clustering; for ASD instances , It belongs to the first The fuzzy membership values ​​of each cluster are: (1) (2) in, , Indicates the first i The ASD instance belongs to the first c Fuzzy membership degree of a class; It is the first c Cluster centers of a class; It is a fuzzy index; solved iteratively. and , No. i The ASD instance belongs to the first c Optimal fuzzy membership value of a class This can be obtained, and further, the fuzzy membership matrix can be obtained. Optimal fuzzy membership vector Reflects the first i The expression uncertainty of an ASD instance; 4.2 FM_ASD constructs a fuzzy relation matrix of "fuzzy clustering - ASD labeling" based on the fuzzy membership vector of ASD samples. Specifically, for The j row vector , It was in the j A fuzzy membership vector that integrates all relevant instances on each ASD tag; that is: (3) (4) in Indicates 1 row C The zero vector of the column; in fact, It is a global uncertainty analysis; it is based on the overall fuzzy relationship between fuzzy clustering of all ASD instances and ASD tags. 4.3FM_ASD, based on fuzzy logic criteria, transforms the fuzzy relation matrix of "fuzzy clustering-ASD labeling" into... Transformed into a fuzzy relation matrix of "ASD instance - ASD tag" ;for The first in i Line 1 j Column elements , Its calculation formula is: (5) in yes The i Line 1 c Column elements, It is the first j The ASD instance in the ... c Fuzzy membership degree on a cluster This represents the minimum operation in fuzzy matrix composition. This represents the maximum operation in the fuzzy matrix synthesis operation; in fact, formula (5) is the fuzzy matrix synthesis operation, which can realize the reasoning modeling of fuzzy relations. 4.4FM_ASD will generate a fuzzy relationship matrix of "ASD instance - ASD tag". Integrating into the ASD tag matrix Obtain the new tag matrix ; "ASD Instances - ASD Tags" Fuzzy Relationship Matrix The distribution of uncertain information in ASD data is implicitly contained; therefore, compared to the original clear set label matrix... , More closely reflects the actual distribution characteristics of ASD data; The specific calculation formula is as follows: (6) The establishment of fuzzy uncertainty learning for ASD data makes the originally fuzzy and uncertain ASD data clear and measurable. It quantifies the uncertainty of ASD data expression based on fuzzy inference modeling. This helps FM_ASD to mitigate the impact of uncertain information expression in ASD data to some extent.

4. The autism spectrum disorder diagnostic method based on fuzzy modeling and multi-label learning as described in claim 3, characterized in that: The sixth step, as described above, consists of the following specific steps: 6.1 For the first ASD tags, Its weighted linear regression is defined as: (9) in It is the first The regression vector of the nth label, which can be used for the nth label Each ASD tag identifies the corresponding discriminant feature; It is based on formula (7) and is the first label-based collaborative learning method. The expression of a tag; It is the first The linear regression weights corresponding to each label, and (10) (11) (12) (13) in Indicates the number of elements in the set; Indicates ASD training set The Middle The sample at the th The value on the mark; formulas (10)-(13) show that the first mark The fewer the number of relevant samples for each ASD label, the better. The larger the weight value; 6.2 The weighted linear regression of all ASD labels can be expressed as: (14) in ; It is a diagonal matrix; 6.3 The L1 regularization term is introduced into formula (14) to constrain the complexity of the model, i.e. (15) in It is a hyperparameter used to control the impact of model complexity; the introduction of the L1 norm makes the discriminative features learned by FM_ASD for each ASD label more accurate.

5. The autism spectrum disorder diagnostic method based on fuzzy modeling and multi-label learning as described in claim 4, characterized in that: The seventh step, as described above, consists of the following specific steps: 7.1 The implementation of higher-order functional connectivity learning is mainly based on the following two arguments: a) Regression vector Can be used to identify the first Discriminative features of ASD tags, ; The larger the value, the better. This indicates the first The second-order functional connectivity feature is useful for identifying the first... The greater the contribution of each ASD tag; b) The correlation between any two ASD tags is consistent with the correlation between their discriminative features; that is, in the ASD dataset, the correlation between the two ASD tag vectors is consistent. ASD tag vectors and the ASD tag vectors The smaller the similarity, The smaller the similarity of their ASD discriminant features, the smaller their regression vectors will be. and The smaller the similarity, the better; 7.2 For any two label vectors in the ASD dataset and Their higher-order functional connectivity learning can be defined as: (16) in, , , Used to measure the ASD tags and the ASD tags Similarity between them; calculated using cosine similarity. ; and The smaller the similarity, The smaller the value, The larger the value; 7.3 The higher-order function connectivity learning for all ASD tags is as follows: (17) in ; It is a hyperparameter used to adjust the impact of higher-order functional connectivity learning.

Images