A method and system for autism identification based on dynamics and sparse coding

Abstract

This invention relates to the field of autism identification technology, and more specifically, to an autism identification method and system based on dynamic and sparse coding. The invention constructs an identification model, first building a feature extraction template, then training the identification model to obtain a trained model. The trained model is then used to process fMRI brain signal data from G patients to be identified, yielding G identification results. Considering the differences in certain specific brain functions between autistic patients and normal individuals, the identification model constructed in this invention selects brain functional connections with significant differences to form feature selection when extracting features from fMRI brain signal data. It also uses the feature extraction template to extract feature vectors with high F-scores, thereby reducing the amount of subsequent feature processing and achieving better efficiency with less computation. This invention solves the problem of low efficiency in existing models.

Description

Technical Field

[0001] This invention relates to the field of autism recognition technology, and more specifically, to an autism recognition method based on dynamic and sparse coding, and an autism recognition system using this method. Background Technology

[0002] Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder primarily characterized by speech and communication impairments and repetitive, stereotyped behaviors. With the increasing prevalence of ASD cases in recent years, greater attention has been paid to its diagnosis and treatment. Since there are currently no effective medications for autism, early diagnosis is a crucial means of prevention.

[0003] For clinical diagnosis, combining deep learning with fMRI data to assist doctors in making more accurate patient identification is crucial. However, current work still faces the following challenges:

[0004] 1) Some existing models reduce the dimensionality and classify all feature data. However, there are some unimportant feature data among all features. If these features are also reduced in dimensionality and classified, it is equivalent to wasting computing power, which will affect the efficiency of the model and make the model training time-consuming.

[0005] 2) Some models that adopt a self-encoding architecture do not take into account the dynamic characteristics of the data itself and are not linked to dynamic technologies, resulting in room for improvement in the model's recognition rate. Summary of the Invention

[0006] Therefore, it is necessary to provide an autism identification method and system based on dynamics and sparse coding to address the problem of low efficiency in existing models.

[0007] This invention is achieved using the following technical solution:

[0008] In a first aspect, the present invention discloses an autism identification method based on dynamics and sparse coding, comprising the following steps:

[0009] To build a recognition model, first construct a feature extraction template, then train the recognition model to obtain a trained recognition model;

[0010] Obtain fMRI brain signal data from G patients to be identified; G≥1;

[0011] The trained recognition model is used to process the fMRI brain signal data of G patients to be identified, and G recognition results are obtained; where the g-th recognition result corresponds to the fMRI brain signal data of the g-th patient to be identified; g∈[1,G].

[0012] The recognition model includes: a feature extraction unit, a feature enhancement unit, and a classification prediction unit.

[0013] The feature extraction unit processes the fMRI brain signal data of the g-th patient to be identified to obtain Q potential features. The feature extraction unit includes: a sliding window processing subunit, a vector transformation subunit, a feature selection subunit, and a sparse autoencoder subunit. The sliding window processing subunit divides the fMRI brain signal data of the g-th patient to be identified into n data windows according to a sliding window method and converts them into n functional connectivity matrices accordingly. The vector transformation subunit selects the upper triangle of the n functional connectivity matrices and expands the upper triangle of the e-th functional connectivity matrix into S one-dimensional feature vectors, outputting a total of n*S one-dimensional feature vectors; e∈[1,n]. The feature selection subunit extracts P high-value feature vectors with the highest F-scores from the S one-dimensional feature vectors corresponding to the e-th functional connectivity matrix according to the feature extraction template, outputting a total of n*P high-value feature vectors. The sparse autoencoder subunit performs feature dimensionality reduction on the P high-value feature vectors corresponding to the e-th functional connectivity matrix to obtain Q potential features, outputting a total of n*Q potential features.

[0014] The feature enhancement unit is used to enhance the Q latent features corresponding to the e-th functional connection matrix to obtain Q enhanced features, and outputs a total of n*Q enhanced features.

[0015] The classification prediction unit is used to perform feature prediction on n*Q enhanced features based on a one-way LSTM to obtain the recognition result.

[0016] This autism identification method based on dynamics and sparse coding implements the method or process according to embodiments of this disclosure.

[0017] Secondly, the present invention discloses an autism identification system based on dynamics and sparse coding, which uses the autism identification method based on dynamics and sparse coding as disclosed in the first aspect.

[0018] The autism identification system based on dynamic and sparse coding includes: a data acquisition module and an identification model module.

[0019] The data acquisition module is used to acquire fMRI brain signal data of G patients to be identified; G≥1. The recognition model module is loaded with a trained recognition model, which is used to process the fMRI brain signal data of the G patients to be identified to obtain G recognition results; where the g-th recognition result corresponds to the fMRI brain signal data of the g-th patient to be identified; g∈[1,G].

[0020] The recognition model module includes a feature extraction module, a feature enhancement module, and a classification prediction module. The feature extraction module processes the fMRI brain signal data of the g-th patient to be identified to obtain n*Q latent features; the feature enhancement module enhances the n*Q latent features to obtain n*Q enhanced features; and the classification prediction module predicts the n*Q enhanced features based on a unidirectional LSTM to obtain the recognition result.

[0021] This autism identification system based on dynamic and sparse coding implements the methods or processes according to embodiments of this disclosure.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. This invention takes into account the differences in certain brain functions between autistic patients and normal people. When extracting features from fMRI brain signal data, it selects brain functional connections with significant differences to form feature selection, and uses feature extraction templates to extract feature vectors with high F-scores, thereby reducing the amount of subsequent feature processing and achieving better efficiency with less computation.

[0024] 2. This invention combines dynamic and self-encoding dimensionality reduction features, which can make data features more representative, thereby improving the prediction accuracy of the model.

[0025] 3. This invention adopts an end-to-end mode during model training, which simplifies the optimization of the entire system model through multi-task learning; and it constructs feature extraction templates during the model training phase, which can significantly reduce the training time of the model. Attached Figure Description

[0026] Figure 1 This is a flowchart of the autism identification method based on dynamics and sparse coding proposed in Embodiment 1 of the present invention;

[0027] Figure 2 This is a data flow diagram of autism identification using the identification model in Embodiment 1 of the present invention;

[0028] Figure 3 This is a structural diagram of the autism identification system based on dynamics and sparse coding in Embodiment 2 of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0031] Unless otherwise defined, 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. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] Example 1

[0033] Please see Figure 1 , Figure 1 This is a flowchart of the autism identification method based on dynamics and sparse coding in Embodiment 1. In general, this autism identification method includes the following steps:

[0034] Step 1: Build a recognition model. First, construct a feature extraction template, and then train the recognition model to obtain a trained recognition model.

[0035] It should be noted that the feature extraction template is constructed based on the F-score before model training, which will not be discussed here.

[0036] The recognition model can be divided into three functional parts: feature extraction, feature enhancement, and classification prediction.

[0037] Step 2: Obtain fMRI brain signal data of G patients to be identified; G≥1.

[0038] The process involves performing brain scans on patients to be identified to obtain brain signal scan images; then, the brain signal scan images are templated using AAL116 to form fMRI brain signal data—which is represented as time-series signals of brain regions.

[0039] Step 3: Use the trained recognition model to process the fMRI brain signal data of G patients to be identified to obtain G recognition results; where the g-th recognition result corresponds to the fMRI brain signal data of the g-th patient to be identified; g∈[1,G].

[0040] The recognition model will be explained in detail below:

[0041] A, the feature extraction unit is used to process the fMRI brain signal data of the g-th patient to be identified to obtain Q potential features.

[0042] See Figure 2 The feature extraction unit includes: a sliding window processing subunit, a vector transformation subunit, a feature selection subunit, and a sparse autoencoder subunit.

[0043] The sliding window processing subunit is used to divide the fMRI brain signal data of the g-th patient to be identified into n data windows in a sliding window manner, and convert them into n functional connectivity matrices accordingly.

[0044] The sliding window mechanism allows for overlapping sliding window processing of fMRI brain signal data, thereby achieving dynamic feature data. In this embodiment 1, n is 16. The size of the sliding window is one-eighth of the time series length, and the step size is one-sixteenth of the time series length. The time series length is the acquisition time T of the fMRI brain signal data of the g-th patient to be identified.

[0045] After obtaining the data windows, a corresponding functional connectivity matrix (i.e., Pearson correlation matrix) is calculated for each data window. The formula for calculating the functional connectivity matrix is:

[0046]

[0047] In the formula, ρ uv Represents the functional connection matrix; u t v represents the t-th element in a brain region time series U of total length T; t Let represent the t-th element in a brain region time series V of total length T; u represents the mean of elements in a brain region time series U of total length T; v represents the mean of elements in a brain region time series V of total length T.

[0048] The vector transformation subunit is used to select the upper triangle of n functional connection matrices and expand the upper triangle of the e-th functional connection matrix into S one-dimensional eigenvectors; e∈[1,n].

[0049] Since there are n functional connection matrices, the vector transformation subunit outputs a total of n*S one-dimensional feature vectors.

[0050] in, R represents the number of brain regions.

[0051] It should be noted that the functional connection matrix is ​​a symmetric matrix that is symmetrical along its diagonal, so choosing the upper triangle can represent the complete information of the matrix.

[0052] In this embodiment 1, the template used for the functional connectivity matrix is ​​AAL116, i.e., R = 116. Since R is 116, S is 6670. Therefore, one upper triangular matrix is ​​expanded into 6670 one-dimensional eigenvectors; 16 functional connectivity matrices are expanded into 16*6670 one-dimensional eigenvectors.

[0053] The feature selection subunit is used to extract P high-value feature vectors from the S one-dimensional feature vectors corresponding to the e-th functional connectivity matrix based on the feature extraction template.

[0054] Specifically, for the e-th functional connection matrix, its corresponding S one-dimensional feature vectors are input into the feature selection sub-unit, and P high-value feature vectors are extracted from the S one-dimensional feature vectors according to the feature extraction template.

[0055] The feature extraction template is a positional mapping relationship between S one-dimensional feature vectors and P high-value feature vectors. In other words, the feature selection subunit can directly extract P high-value feature vectors from the S one-dimensional feature vectors based on the fixed positional mapping.

[0056] Since there are n functional connection matrices, the feature selection subunit outputs a total of n*P high-value feature vectors.

[0057] The sparse autoencoder subunit is used to perform feature dimensionality reduction on the P high-value eigenvectors corresponding to the e-th functional connectivity matrix to obtain Q latent features.

[0058] Specifically, for the e-th functional connection matrix, its corresponding P high-value feature vectors are processed by a sparse autoencoder submodule and then encoded and decoded by a fully connected layer to obtain Q latent features.

[0059] In particular, KL divergence is used during encoding, which achieves sparsity in the autoencoder while preventing model overfitting.

[0060] Since there are n functional connection matrices, the sparse autoencoder subunit outputs a total of n*Q latent features.

[0061] B, the feature enhancement unit is used to enhance the Q latent features corresponding to the e-th functional connection matrix to obtain Q enhanced features.

[0062] Specifically, for the e-th functional connection matrix, its corresponding Q latent features are input into the feature enhancement unit and processed through a multi-head attention mechanism to make the mean of the Q latent features change less and the variance increase, thus obtaining Q enhanced features.

[0063] Since there are n functional connection matrices, the feature enhancement unit outputs a total of n*Q enhanced features.

[0064] In this embodiment 1, the feature enhancement part adopts the Block part of the Transformer model, which is based on the multi-head attention mechanism. The enhanced features obtained in this way have a more obvious feature distribution.

[0065] C, the classification prediction unit is used to perform feature prediction on n*Q enhanced features based on unidirectional LSTM to obtain the recognition result.

[0066] Specifically, the classification prediction unit first uses a one-way LSTM to predict the n*Q enhanced features to obtain the predicted features, and then uses a Softmax classifier to perform binary classification on the predicted features to obtain probability one and probability two; where probability one is the probability that the g-th patient to be identified is a normal person; probability two is the probability that the g-th patient to be identified is an ASD patient.

[0067] If probability 1 is greater than probability 2, the output is 0, and the g-th identification result is the g-th patient to be identified as a normal person;

[0068] If probability 1 is less than probability 2, the output is 1, then the g-th identification result is the g-th patient to be identified as an ASD patient.

[0069] In summary, by using a trained recognition model to process the fMRI brain signal data of G patients to be identified, G recognition results can be obtained, thus enabling the diagnosis of autism.

[0070] In addition, untrained recognition models need to be trained to obtain trained recognition models that can be applied to the autism recognition methods described above. Before training the model, the feature extraction template must first be determined.

[0071] The specific process of feature template construction is as follows:

[0072] S1.1 Obtain the sample dataset and randomly select four-fifths of the data in the sample dataset as reference data.

[0073] In this Example 1, the sample dataset is the Autism Brain Imaging Data Exchange dataset (abbreviated as ABIDE), which is a consortium of 17 international imaging sites that have collaborated to collect and publicly share neuroimaging data of 505 ASD patients and 530 typical controls.

[0074] By transforming the neuroimaging data, brain region signal files for each subject were obtained. Due to the presence of problematic data in ABIDE, preprocessing of the sample dataset is generally required to remove problematic data (i.e., discarding data with brain region signal values ​​of 0). The final sample dataset includes 948 subjects (457 with autism and 491 normal individuals). The data uses the AAL116 template, resulting in an R-value of 116. The reference dataset includes T subjects, where T = 758.

[0075] S1.2, the reference data is processed through the sliding window processing subunit, and each subject receives n functional connectivity matrices (which can be called functional connectivity reference matrices), for a total of T*n functional connectivity reference matrices;

[0076] The T*n functional connection reference matrices are then processed by the vector transformation subunit, and each functional connection reference matrix yields S corresponding one-dimensional feature vectors (which can be called one-dimensional feature reference vectors), for a total of T*n*S one-dimensional feature reference vectors.

[0077] Then, the T*n*S one-dimensional feature reference vectors are averaged to obtain T*S one-dimensional feature average vectors. Specifically, averaging is performed using a sliding window to eliminate dimension n, thus obtaining T*S one-dimensional feature average vectors.

[0078] S1.3 Calculate S F-scores based on T*S one-dimensional feature average vectors, and retain the P F-scores whose values ​​are in the top P positions.

[0079] The F-score is used to initially distinguish which features can effectively differentiate between ASD patients and normal individuals. In other words, the F-score is used as an indicator to measure the difference between two groups of real numbers (features of ASD patients and features of normal individuals).

[0080] The formula for calculating the F-score is:

[0081]

[0082] In the formula, F(i) represents the i-th F-score, i∈[1,S]; This represents the average value of the i-th characteristic of the population; This represents the average value of the i-th characteristic of ASD patients. This represents the average value of the i-th characteristic of a normal person; This represents the i-th characteristic of the k-th ASD patient. n represents the i-th characteristic of the k-th normal person; a n represents the number of ASD patients among the T subjects. cT represents the number of normal individuals among the T subjects; T = n a +n c .

[0083] The S F-scores are obtained from the T*S one-dimensional feature average vectors calculated as described above. It should be noted that the positions of the S F-scores correspond one-to-one with the positions of the S one-dimensional feature average vectors of any subject, and can be considered identical.

[0084] Based on the positions of P F-scores in S F-scores, a positional mapping relationship is established, thus obtaining the feature extraction template.

[0085] In other words, P F-scores can be used to find the corresponding P one-dimensional feature average vectors in the S one-dimensional feature average vectors based on their positions in the S F-scores, thus establishing a positional mapping relationship. Therefore, referring to the above, the feature selection subunit can also directly extract P high-value feature vectors from the S one-dimensional feature vectors.

[0086] In this embodiment 1, since S is 6670, 6670 F-scores are obtained. These 6670 F-scores are then sorted from largest to smallest, and the top P F-scores are retained. The value of P is adjusted based on the actual situation because: too small a value will result in missing feature references, while too large a value will lead to excessive data volume and computational complexity. In this embodiment 1, considering the subsequent autoencoder dimensionality reduction, the optimal value of P is determined to be 573.

[0087] After the feature extraction template is determined, model training can proceed. Directly using the feature extraction template during model training can significantly reduce data processing volume and shorten model training time. Specifically, the model training method includes the following steps:

[0088] S2.1, Obtain the sample dataset.

[0089] The sample dataset here is the same as in S1.1. The final retained sample dataset includes 948 subjects (457 with autism and 491 normal individuals). The data uses the AAL116 template, so R is 116.

[0090] S2.2, divide the sample dataset into training and test sets.

[0091] In S2.2, the sample dataset is divided into five folds, and five-fold cross-validation is used to divide it into training and test sets.

[0092] S2.3, train the recognition model using the training set and introduce a loss function. Perform backpropagation to adjust the model parameters.

[0093] in,

[0094] In the formula, The loss representing the sparse autoencoder subunit:

[0095] In the formula, x is a one-dimensional feature vector input into the sparse autoencoder subunit. It is a potential feature of the output of sparse autoencoder subunits.

[0096] Represents binary cross-entropy loss:

[0097] In the formula, N represents the number of subjects participating in the prediction.

[0098] y i It is the actual label of the object (0 or 1); It is a probability prediction of the object, that is, a predicted value from 0 to 1.

[0099] The KL divergence loss is represented as a sparse penalty term:

[0100] In the formula, ρ represents the sparsity constraint constant; h n represents the number of neurons in the innermost layer of the sparse autoencoder; m represents the number of subjects added to the model each time. a represents the average activation level of neuron j; j (x(i)) represents the activation of hidden unit j in a specific hidden layer; x(i) represents the number of neurons in the hidden layer.

[0101] β is the weighting coefficient; β is the sparse penalty term coefficient.

[0102] Loss function If the value is below the preset threshold or the recognition model converges, a well-trained recognition model is obtained.

[0103] In addition, a test set is used to evaluate the performance of the trained recognition model in order to keep a close eye on the training effect of the model.

[0104] This recognition model selects features based on brain functional connections with significant differences and introduces the F-score as a metric, thereby reducing the amount of subsequent feature processing. This allows for better efficiency with less computation and significantly reduces the training time of the model.

[0105] Furthermore, in practical applications, this recognition model directly uses feature extraction templates for feature extraction, which requires less computation and can also reduce the time spent on model prediction.

[0106] Example 2

[0107] This embodiment 2 discloses an autism recognition system based on dynamics and sparse coding, which uses the autism recognition method based on dynamics and sparse coding from embodiment 1.

[0108] See Figure 3 The autism identification system based on dynamics and sparse coding includes: a data acquisition module and an identification model module.

[0109] The data acquisition module is used to acquire fMRI brain signal data of G patients to be identified; G≥1.

[0110] The recognition model module is loaded with a trained recognition model, which is used to process the fMRI brain signal data of G patients to be identified and obtain G recognition results; where the g-th recognition result corresponds to the fMRI brain signal data of the g-th patient to be identified; g∈[1,G].

[0111] The recognition model module includes a feature extraction module, a feature enhancement module, and a classification prediction module. The feature extraction module processes the fMRI brain signal data of the g-th patient to be identified to obtain n*Q latent features; the feature enhancement module enhances the Q latent features to obtain n*Q enhanced features; and the classification prediction module predicts the n*Q enhanced features based on a unidirectional LSTM to obtain the recognition result.

[0112] The feature extraction module can be further subdivided into: a sliding window processing submodule, a vector transformation submodule, a feature selection submodule, and a sparse autoencoder submodule. The sliding window processing submodule divides the fMRI brain signal data of the g-th patient to be identified into n data windows using a sliding window method, and converts them into n functional connectivity matrices accordingly. The vector transformation submodule selects the upper triangle of the n functional connectivity matrices and expands the upper triangle of the e-th functional connectivity matrix into S one-dimensional feature vectors, outputting a total of n*S one-dimensional feature vectors; e∈[1,n]. The feature selection submodule extracts P high-value feature vectors with the highest F-scores from the S one-dimensional feature vectors corresponding to the e-th functional connectivity matrix based on the feature extraction template, outputting a total of n*P high-value feature vectors. The sparse autoencoder submodule performs feature dimensionality reduction on the P high-value feature vectors corresponding to the e-th functional connectivity matrix to obtain Q latent features, outputting a total of n*Q latent features.

[0113] Since the autism identification system uses the autism identification method of Example 1, it also has the advantages of Example 1, which will not be repeated here.

[0114] Example 3

[0115] This embodiment 3 compares and verifies the method of embodiment 1:

[0116] 1. An existing feature-free selection method (i.e., the method in the background technique that performs dimensionality reduction and classification on all feature data) was introduced, and compared with the method in Example 1. The training time and recognition accuracy of the two models were examined, and the results are shown in Table 1 below:

[0117] Table 1 Comparison of Time and Recognition Accuracy

[0118]

[0119] It can be seen that the accuracy of the two methods is the same, but the method in Example 1 significantly reduces the model training time compared to the existing methods.

[0120] 2. Introduce an existing autoencoder combined with an F-value model and compare it with the method in Example 1.

[0121] In this study, the trained recognition model from Example 1 was designated as Model 1, and the trained autoencoder combined with the F-score model was designated as Model 2. Model 1 and Model 2 were then used to identify individuals on the same validation dataset. The validation dataset consisted of one-fifth of the ABIDE dataset, containing real labels that could be used to determine whether an individual had autism. The accuracy comparison results between Model 1 and Model 2 are shown in Table 2 below.

[0122] Table 2 Comparison of Recognition Accuracy

[0123]

[0124] It can be seen that the recognition accuracy of Model 1 is higher than that of Model 2, and it has also been improved.

[0125] In summary, experimental verification has confirmed that the method in Example 1 can achieve better efficiency with less computation, significantly reduce model training time, and improve recognition accuracy.

[0126] Example 4

[0127] This embodiment 4 also discloses a readable storage medium storing computer program instructions. When the computer program instructions are read and run by a processor, the autism identification method based on dynamic and sparse coding disclosed in embodiment 1, or / and the training method of the identification model, are executed.

[0128] When applying the method of Example 1, it can be applied in the form of software, such as by designing it as a program that can run independently on a computer-readable storage medium. The computer-readable storage medium can be a USB flash drive, designed as a USB security token, and the program can be designed to start the entire method through an external trigger.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for fMRI brain signal recognition based on dynamics and sparse coding, characterized in that, Includes the following steps: To build a recognition model, first construct a feature extraction template, then train the recognition model to obtain a trained recognition model; Get G fMRI brain signal data of a patient to be identified; G ≥1; Use the trained recognition model to G The fMRI brain signal data of the patients to be identified were processed to obtain G The identification results; among them, the first... g The recognition result corresponds to the first... g fMRI brain signal data of a patient to be identified; g ∈[1, G ]; The recognition model includes: The feature extraction unit is used to extract the first feature. g fMRI brain signal data of the patients to be identified were processed to obtain Q The feature extraction unit comprises: a sliding window processing subunit, a vector transformation subunit, a feature selection subunit, and a sparse autoencoder subunit; the sliding window processing subunit is used to extract the first potential feature in a sliding window manner. g fMRI brain signal data of the patients to be identified were divided into n Each data window is converted accordingly. n A functional connection matrix; the vector transformation subunit is used to select... n The upper triangle of the first functional connection matrix, and the first e The upper triangular expansion of the functional connection matrix S Each one-dimensional feature vector outputs a total of [number] outputs. n*S One-dimensional feature vector; e ∈[1, n The feature selection subunit is used to select features from the first feature extraction template based on the first feature extraction template. e Each functional connection matrix corresponds to S Extracting the top-ranking F-score features from one-dimensional feature vectors P 1 high-value feature vector, outputting a total of 10 high-value feature vectors, with a total output of 1000 high-value feature vectors. n*P The 1st high-value feature vector; the sparse autoencoder subunit is used to process the 1st high-value feature vector; e Each functional connection matrix corresponds to P We obtain the feature dimension reduction of the high-value feature vectors. Q 1 potential feature, with a total output n*Q One potential feature; Feature enhancement section, which is used for the first e Each functional connection matrix corresponds to Q Feature enhancement is performed on the latent features to obtain Q One enhanced feature, with a total output n*Q One enhanced feature; and The classification prediction unit, which is used for unidirectional LSTM pairs n*Q The enhanced features are used for feature prediction to obtain the recognition result.

2. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The sliding window size is one-eighth of the time series length, and the step size is one-sixteenth of the time series length; the time series length is the [missing information]. g fMRI brain signal data acquisition time for each patient to be identified T .

3. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1 or 2, characterized in that, The formula for calculating the functional connectivity matrix is: ； In the formula, Represents the functional connection matrix; Indicates the total length is T brain region time series U The Middle t One element; Indicates the total length is T brain region time series V The Middle t One element; Indicates the total length is T brain region time series U The mean of the elements; Indicates the total length is T brain region time series V The element mean.

4. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The sparse autoencoder submodule is connected through a fully connected layer. P Encode and decode the high-value feature vectors to obtain Q One potential feature; KL divergence is used in the encoding process.

5. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The feature enhancement unit is based on a multi-head attention mechanism. Q Process each potential feature to make Q The latent features have small changes in mean and large changes in variance, resulting in... Q One enhanced feature.

6. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The classification prediction unit first uses a one-way LSTM to... Q The enhanced features are used to predict the predicted features, and then the predicted features are binary-classified using a Softmax classifier to obtain probability 1 and probability 2; where probability 1 is the first... g The probability that the first patient to be identified is a normal person; probability two is the probability of the first... g The probability that a patient to be identified is an ASD patient.

7. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The method for constructing the feature extraction template includes the following steps: S1.1 Obtain the sample dataset and randomly select four-fifths of the data in the sample dataset as reference data; S1.2, the reference data is processed through the sliding window processing subunit to obtain... T*n Each function is connected to a reference matrix; T This indicates the number of participants included in the reference data; T*n Each functional connection reference matrix is ​​then processed by a vector transformation subunit to obtain... T*n*S One-dimensional feature reference vector; Then to T*n*S The data is averaged using one-dimensional feature reference vectors to obtain... T*S One-dimensional feature average vector; S1.3, according to T*S Calculate the average vector of each one-dimensional feature vector. S 1 F-score, and retain the value size in the top 10. P position P One F-score; in accordance with P The F-scores are in S By establishing a positional mapping relationship among the positions in each F-score, the feature extraction template is obtained.

8. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 1, characterized in that, The training method for the recognition model includes the following steps: S2.1, Obtain the sample dataset; S2.2, Divide the sample dataset into a training set and a test set; S2.3, Train the recognition model using the training set, and introduce a loss function. Perform backpropagation to adjust the model parameters; The trained recognition model was evaluated using the test set. Loss function If the value is below the preset threshold or the recognition model converges, a well-trained recognition model is obtained.

9. The fMRI brain signal recognition method based on dynamics and sparse coding according to claim 8, characterized in that, Before S1.2, the sample dataset is preprocessed to remove problematic data from the sample dataset; Or / and, in S1.2, the sample dataset is divided into five folds and then divided into a training set and a test set using five-fold cross-validation.

10. An autism identification system based on dynamics and sparse coding, characterized in that, It uses the fMRI brain signal recognition method based on dynamics and sparse coding as described in any one of claims 1-9; The data acquisition module is used to acquire... G fMRI brain signal data of a patient to be identified; G ≥1; as well as The recognition model module loads a pre-trained recognition model for use in... G The fMRI brain signal data of the patients to be identified were processed to obtain G The identification results; among them, the first... g The recognition result corresponds to the first... g fMRI brain signal data of a patient to be identified; g ∈[1, G ]; The recognition model module includes: a feature extraction module, a feature enhancement module, and a classification prediction module; the feature extraction module is used to perform the following on the first feature extraction module: [The first feature extraction module is used to extract the first feature extraction module, and the second feature enhancement module is used to extract the first feature extraction module, and the third feature enhancement module is used to perform the following on the second feature extraction module, and the fourth feature enhancement module is used to perform the following on the second feature extraction module, and the fifth feature enhancement module is used to perform the following on the third feature extraction g fMRI brain signal data of the patients to be identified were processed to obtain n*Q One potential feature; the feature enhancement module is used for... n*Q Feature enhancement is performed on each latent feature to obtain n*Q One enhanced feature; the classification prediction module is used for the unidirectional LSTM pair n*Q The enhanced features are used for feature prediction to obtain the recognition result.

Images