Autism spectrum disorder risk detection system and method

Abstract

The invention discloses an autism spectrum disorder risk detection system and method, relates to the medical field of autism auxiliary diagnosis, and solves the problems that a traditional ASD risk detection method is weak in edge information extraction and analysis, the value of an edge serving as a potential biomarker cannot be fully excavated, and the risk of the autism spectrum disorder is affected. The accuracy and the reliability of autism risk detection based on the brain network are limited. The method comprises the following steps: collecting detected fMRI data, analyzing an average time sequence of the preprocessed fMRI data, and further constructing a functional connection matrix of a brain interval; extracting features associated with ASD from the functional connection matrix to obtain an fMRI brain network feature set; the ASD risk of the detected fMRI data is identified by a classification model. The accuracy of autism spectrum disorder risk detection is realized by focusing edge dynamic features and combining an advanced machine learning classifier and an optimization strategy.

Description

Technical Field

[0001] The present invention belongs to the field of autism-assisted diagnosis and medicine, and specifically relates to an autism spectrum disorder risk detection system and method. Background Art

[0002] ASD is a severe neurodevelopmental disorder, and its main features include social communication and language communication disorders, restricted interests and repetitive behaviors, etc.

[0003] Currently, the risk detection of ASD mainly focuses on the node-centered functional connectivity matrix, which is used to explore the abnormal changes in the connection strength between the brain regions of patients, ignoring the importance of the edge-centered functional connectivity network in the high-order information flow and interaction between brain regions. As the key carrier of information transmission between brain regions, the edge contains interaction features. However, the traditional methods for extracting and analyzing edge information are relatively weak, and the value of the edge as a potential biomarker has not been fully explored, limiting the accuracy and reliability of autism risk detection based on brain networks.

[0004] Therefore, the present invention proposes an autism spectrum disorder risk detection method and system. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention proposes an autism spectrum disorder risk detection system and method, which is used to solve the technical problems that the traditional ASD risk detection method is relatively weak in extracting and analyzing edge information, fails to fully explore the value of the edge as a potential biomarker, and limits the accuracy and reliability of autism risk detection based on brain networks.

[0006] To achieve the above object, the first aspect of the present invention provides an autism spectrum disorder risk detection method, including the following steps:

[0007] Collect the detected fMRI data, perform preprocessing, and analyze the average time series of the preprocessed fMRI data; wherein, the fMRI data includes public dataset or individual sample data obtained through clinical experiments;

[0008] Based on the average time series of the fMRI data, construct a functional connectivity matrix in the brain network centered on the edge; wherein, the edge is the functional connection edge between brain regions in the brain functional network;

[0009] Extract the features associated with ASD from the functional connectivity matrix to obtain the fMRI brain network feature set;

[0010] Input the fMRI brain network feature set into the classification model to identify the ASD risk of the detected fMRI data.

[0011] Preferably, the preprocessing process of the fMRI data includes:

[0012] Performing skull stripping, slice timing correction, motion correction, global mean intensity normalization, nuisance signal regression, and band-pass filtering on the fMRI data in sequence, and registering the processed fMRI data to the standard anatomical space.

[0013] Preferably, the analysis of the average time series of the preprocessed fMRI data includes:

[0014] Dividing the preprocessed fMRI data into several brain regions, and extracting all voxels of each brain region from the preprocessed fMRI data according to the AAL atlas;

[0015] Calculating the average value of the time signals of all voxels in each brain region at each time point:

[0016] TS j (t) = (1 / N) × Σ i Vij(t);

[0017] where i is the serial number of the brain region, j is the serial number of the voxel, i = 0, 1, …, N, j = 0, 1, …, M, N and M are positive integers, t is the serial number of the time point of the brain region, and TS j (t) is the average time series of the jth brain region, and Vij(t) is the signal value of the jth voxel in the ith brain region at the tth time point.

[0018] Preferably, calculating and constructing the functional connectivity matrix between brain regions based on the average time series of the fMRI data includes:

[0019] Calculating the product of the average time series of the blood oxygen levels of any two brain regions to obtain several edge time series;

[0020] Calculating the cosine similarity between any two edge time series to obtain the functional connectivity strength between several edge time series; constructing the functional connectivity matrix based on several functional connectivity strengths.

[0021] Preferably, the average time series of each brain region is a time series processed by standardization, including:

[0022] In the construction of functional connectivity, performing z-scores standardization processing on the extracted average time series as follows:

[0023]

[0024] where x i is the initial average time series, u i is the mean value of the average time series, and σ iis the standard deviation of the time series, z i is the average time series after standardization processing.

[0025] Preferably, extracting the features associated with ASD from the functional connectivity matrix to obtain the fMRI brain network feature set includes:

[0026] The random forest extracts the features associated with ASD from the functional connectivity matrix through its built-in feature importance evaluation mechanism to obtain the fMRI brain network feature set.

[0027] Preferably, the random forest through its built-in feature importance evaluation mechanism includes:

[0028] Statistically count the cumulative number of times each fMRI brain network feature is selected as a node splitting feature in the integrated decision tree model; calculate the reduction in the sample class information entropy within the node before and after splitting to obtain the purity improvement metric value of the split; perform a weighted sum of the purity improvement metric values of each fMRI brain network feature on all splitting nodes to obtain the importance score of each fMRI brain network feature.

[0029] Preferably, identifying the ASD risk of the detected fMRI data includes:

[0030] Input the fMRI brain network feature set into the classification model, and the classification model outputs a detection label; among them, the classification model is constructed based on SVM; the detection label includes a normal label and an abnormal label, usually represented by "0" and "1" respectively.

[0031] Preferably, the classification model is constructed based on SVM, including:

[0032] Obtain several fMRI brain network feature sets and detection labels from historical data;

[0033] Integrate the fMRI brain network feature set and abnormal fMRI data into several groups of training data and test data; use the training data to train the SVM; use the test data to test the trained SVM, and adjust the SVM according to the test results; finally obtain a classification model with the fMRI brain network feature set as the input and the detection label as the output.

[0034] Preferably, the second aspect of the present invention provides an autism spectrum disorder risk detection system, including a data processing module and a risk detection module;

[0035] Data processing module: used to collect the detected fMRI data, perform preprocessing, and analyze the average time series of the preprocessed fMRI data;

[0036] Based on the average time series of the fMRI data, calculate the functional connectivity matrix between brain regions in the brain network centered on the edge;

[0037] Risk detection module: used to extract features associated with ASD from the functional connectivity matrix to obtain the fMRI brain network feature set;

[0038] Input the fMRI brain network feature set into the classification model to identify the ASD risk of the detected fMRI data.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] The present invention proposes a brain network analysis method based on edge centers, breaking through the limitations of traditional brain network analysis methods, overcoming the deficiency of existing brain network analysis methods that overly rely on node connections, and shifting the focus of analysis to the edge as the core carrier of information transmission and interaction between brain regions. Since the edge plays an important role in the high-order information flow and interaction of multiple brain regions, by focusing on the high-dimensional and high-order information of the edge, the present invention can more accurately capture the abnormal patterns of brain functions in autistic patients. This improvement provides more biologically interpretable features for the early diagnosis of ASD, improving the accuracy and reliability of detection. Since brain network features often have a high-dimensional nature, direct classification is prone to overfitting problems. The present invention effectively suppresses overfitting and enhances the generalization ability of the model while ensuring classification performance by introducing advanced machine learning classifiers and combining feature selection and model optimization strategies. And the classification model constructed based on eFCN features shows better classification accuracy and sensitivity than the traditional nFCN method in experimental verification. This indicates that more key features can be obtained through edge functional connectivity analysis, effectively improving the ASD detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the descriptions of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0042] Figure 1 It is a schematic flowchart of the method of the present invention;

[0043] Figure 2 It is a comparison diagram of the brain network connection matrices of the present invention;

[0044] Figure 3 It is a performance comparison diagram of nFCN and eFCN of the present invention;

[0045] Figure 4 It is a diagram showing the change of the classification performance of the model of the present invention with the number of features;

[0046] Figure 5 This is the performance evaluation diagram of the SVM model of the present invention;

[0047] Figure 6 This is the comparison diagram of the performance evaluation of the classifier of the present invention;

[0048] Figure 7 This is the ranking diagram of the number of occurrences of brain regions involved in the present invention;

[0049] Figure 8 This is the localization of the top four brain regions with the highest importance ranking of the present invention;

[0050] Figure 9 This is the brain region connection diagram of the present invention. Detailed implementation manners

[0051] Next, the technical solution of the present invention will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] Please refer to Figure 1 , an embodiment of the first aspect of the present invention provides a method for detecting the risk of autism spectrum disorder, including the following steps:

[0053] Collect the fMRI data to be detected, perform preprocessing, and analyze the average time series of the preprocessed fMRI data; wherein, the fMRI data includes public data sets or individual sample data obtained through clinical experiments;

[0054] A series of standard preprocessing steps are performed on the fMRI data obtained in step 1. The specific steps include skull stripping, slice timing correction, motion correction, global mean intensity normalization, nuisance signal regression, and band-pass filtering (frequency range: 0.01 - 0.1 Hz). The processed fMRI data is registered to the standard anatomical space (MNI152) to support cross-comparison and result verification of the research.

[0055] Based on the average time series of the fMRI data, calculate the functional connection matrix between brain regions in the brain network centered on the edges; wherein, the edges are the functional connection edges between brain regions in the brain functional network;

[0056] Divide the preprocessed fMRI data into several brain regions, and extract all voxels of each brain region from the preprocessed fMRI data according to the AAL atlas;

[0057] Calculate the average value of the time signals of all voxels in each brain region at each time point:

[0058] TSj V(t) = (1 / N) × Σ i Vij(t);

[0059] Where i is the serial number of the brain region, j is the voxel number, i = 0, 1, …, N, j = 0, 1, …, M, N and M are positive integers, t is the serial number of the time point of the brain region, t ∈ [1, T], TS j V(t) is the average time series of the j-th brain region, and Vij(t) is the signal value of the j-th voxel in the i-th brain region at the t-th time point.

[0060] In functional connectivity construction, z-scores normalization is performed on the average time series extracted from all detected brain regions to measure the difference of sample points relative to the overall mean and achieve data normalization. Specifically, calculate its mean and standard deviation, and then perform the following conversion:

[0061]

[0062] Where x i is the initial average time series, u i is the mean of the average time series, σ is the standard deviation of the time series, and z i is the average time series after normalization processing.

[0063] Element-wise multiplication of the average time series extracted from any two brain regions is performed at each time point, generating a new set of time series, labeled as the edge time series;

[0064] c 01 = [z0(1) × z1(1), …, z0(T) × z1(T)];

[0065] c 23 = [z2(1) × z3(1), …, z2(T) × z3(T)];

[0066] Where z0, z1, z2, and z3 respectively represent the time series after z-scores normalization of the four brain regions with serial numbers 0, 1, 2, and 3, and c 01 and c 23 respectively represent the edge time series of brain region pairs 0, 1 and brain region pairs 2, 3; according to this calculation method, several edge time series are obtained;

[0067] Calculate the cosine similarity between any two edge time series to obtain the functional connectivity strength between different edge time series, as follows:

[0068]

[0069] Based on several functional connectivity strengths, construct a functional connectivity matrix, that is, the eFCN matrix.

[0070] The present invention constructs an eFCN using an AAL atlas (excluding the cerebellum) to obtain a K*K symmetric matrix with a dimension of K = 4005. Similar to the traditional nFCN, the eFCN only uses the elements in the lower triangular part of the matrix as features for subsequent analysis. The eFCN is significantly extended based on the traditional nFCN, and the number of dimensions of the matrix is much larger than that of the nFCN matrix. Its additional high-dimensional features can provide richer supplementary information for studying the relationships between brain behaviors. On the one hand, the final averaging step in the calculation of the traditional Pearson correlation coefficient is omitted during the construction of the eFCN, reducing information loss. In fMRI data processing, the time series of each pair of brain regions is retained on the time scale of a single frame, and the significantly improved time resolution more accurately reflects the communication activities of brain regions over time. On the other hand, the eFCN represents the correlation in the interaction between two pairs of brain regions, thus forming a higher-order brain network based on edges. The larger the dimension of the eFCN matrix, the richer the information that can be obtained from the higher-order feature representation of brain connections, thereby characterizing the complex interaction patterns between multiple brain regions and the correlations across brain regions.

[0071] Extract the features associated with ASD from the functional connectivity matrix to obtain an fMRI brain network feature set;

[0072] Specifically, count the cumulative number of times each fMRI brain network feature is selected as a node splitting feature in the integrated decision tree model; calculate the reduction in the sample class information entropy within the node before and after splitting to obtain a measure of the purity improvement of the split; perform a weighted sum of the purity improvement measures of each fMRI brain network feature on all split nodes to obtain the importance score of each fMRI brain network feature.

[0073] It should be noted that as one of the multivariate feature selection methods, the random forest can explore the mutual relationships between features through feature importance scores and shows strong stability in dealing with noisy data and missing values. To achieve a fair evaluation of the nFCN and eFCN subsequently, the research reduces the number of eFCN features to the same as that of the nFCN for dimension matching. For example, the top 4000 features ranked by importance are selected for a comparative study of feature selection methods. In the experiment, the random forest method showed better accuracy performance and could effectively process the high-dimensional brain network data of the eFCN. When performing feature selection, the random forest ranks the features according to their importance and selects the features that are most important for the classification task, reducing feature redundancy and improving the stability and accuracy of the model. Therefore, the random forest is selected as the feature selection method in the experiment.

[0074] For example, the lower triangular part of the eFCN matrix constructed by each subject was subjected to feature extraction, and the element feature set was randomly sampled 1000 times, with 10,000 features sampled each time, and no repeated selection was performed during the sampling process. Through this large-scale sampling strategy, the study achieved a relatively comprehensive and uniform data distribution, ensuring that the features of each brain region were fully covered. The sampled feature set was then used for classification to obtain the importance ranking of each feature. After 1000 samplings, all important features were included in the final feature set for subsequent classification. To further optimize feature selection, after completing the initial run of all features, the study screened out the features with the highest relative importance metrics from all nearly 16 million features. This screening process reduces the number of features in high-dimensional data while ensuring that the number of selected features is large enough to ensure that the feature subset is highly discriminative.

[0075] The fMRI brain network feature set is input into the classification model to identify the ASD risk of the detected fMRI data;

[0076] Specifically, the fMRI brain network feature set is input into the classification model, and the classification model outputs a detection label; wherein the classification model is constructed based on SVM; the detection label includes a normal label and an abnormal label, which are usually represented by "0" and "1" respectively.

[0077] Among them, the classification model is built based on SVM, including the following steps:

[0078] Obtain several fMRI brain network feature sets and detection labels from historical data;

[0079] The fMRI brain network feature set and abnormal fMRI data are integrated into several groups of training data and test data; the training data is used to train the SVM; the test data is used to test the trained SVM, and the SVM is adjusted according to the test results; finally, a classification model is obtained with the input being the fMRI brain network feature set and the output being the detection label.

[0080] In order to verify the effectiveness of the edge-centered functional connectivity method proposed in the present invention in ASD detection, the present invention uses the heterogeneous multi-site dataset ABIDE I released in August 2012 for experimental verification. This dataset brings together data from 17 different international collection sites and publicly shares the neuroimaging and phenotypic data of 1,112 subjects, including 539 ASD patients and 573 typical healthy controls.

[0081] 871 subjects meeting the experimental quality requirements were selected from those meeting the criteria of imaging quality and phenotypic information. The final selected dataset included 403 ASD patients and 468 healthy controls. Table 1 describes the sample information of the ASD patients and healthy control group in the present invention.

[0082] Table 1 Sample Information of ASD Patients and Healthy Control Group

[0083]

[0084] Please refer to Figure 2 , a comparative evaluation was carried out on two methods of representing brain network features, nFCN and eFCN. By quantitatively analyzing the accuracy and sensitivity indicators of their classification, the differences and advantages of the two methods of representing brain network features in classification were evaluated.

[0085] In the experiment, due to the differences in magnetic resonance imaging protocols, participant selection criteria, and recruitment methods among different acquisition sites, there were significant challenges in classifying datasets from multiple institutions. The study adopted a 10-fold cross-validation strategy to effectively reduce the possibility of overfitting and enhance the adaptability of the model to new data. These validation methods could fully evaluate the stability of the model under different data distributions and ensure the reliability of the classification results. To quantitatively evaluate the diagnostic performance of all the methods mentioned in this article, the study reported the accuracy and sensitivity of different methods. The classifier used in this study was SVM, which is a widely used and recognized classifier in medical data classification, especially good at handling high-dimensional data and avoiding overfitting problems.

[0086] Since the feature subsets of eFCN were obtained by random sampling, the classification results of each sampling might vary significantly. To comprehensively evaluate the classification performance, the study calculated the average, maximum, and minimum values of the classification accuracy based on 1000 sampling results and compared them with nFCN, as shown in Table 2 Performance Evaluation of eFCN Random Sampling below.

[0087] Table 2 Performance Evaluation of eFCN Random Sampling

[0088]

[0089] As can be seen from Table 2, the classification accuracy of the sampling subsets of eFCN ranges from 0.40 to 0.75, with an average of 0.58. This large volatility may be due to the very high feature dimension of the original eFCN (4005*4005), resulting in a low probability of selecting meaningful features during the random sampling process, thus affecting the stability of the classification results. In some cases, random sampling of eFCN can achieve high accuracy and sensitivity, indicating its ability to capture important features, and random sampling can cover more feature combinations. In contrast, nFCN maintains relatively stable classification accuracy and sensitivity during the random sampling process, indicating its high reliability in the feature selection process, but it performs worse than eFCN in capturing some relatively subtle and highly discriminative features.

[0090] To compare the classification performance of nFCN and eFCN, the dimensions of nFCN and eFCN were matched with the same number of features. The performance comparison graph of the two is as Figure 3 shown. The results show that when the number of features is low, the accuracy of nFCN is higher than that of eFCN. As the number of features gradually increases, eFCN shows better classification ability under the same number of features, significantly higher than the traditional nFCN. This improvement in performance can be attributed to the advantages of eFCN in brain network modeling. Different from nFCN that directly calculates the correlation or covariance matrix based on time series, eFCN constructs a brain network using edge-level features, which can capture the interaction patterns between brain regions more carefully. Especially in the high-dimensional feature space, eFCN can more comprehensively represent complex brain connection features by additionally introducing the information of edge time series.

[0091] Based on the SVM method, the present invention classifies ASD patients and healthy controls using LOOCV. To optimize the classification performance, the study screened the top 4000 most discriminative features with the best effect from the high-dimensional feature set for the classification task. These features were screened from the original nearly 16 million-dimensional features through the random forest method and have high discriminative ability.

[0092] The experimental results are as Figure 4 shown. When classifying using the first 4000 features, the highest classification accuracy reached 72.42%. This result verifies the effectiveness of feature selection, indicating that the selected eFCN features have high discriminative ability in distinguishing ASD patients from healthy controls and have important classification value in the ASD diagnosis task.

[0093] Using LOOCV ensures that an independent test set is used in each round of validation, successfully avoiding overfitting, and improving the overall stability and adaptability of the model. By analyzing the performance of the classifier, it has a better performance in high-dimensional data classification, can effectively avoid the curse of dimensionality problem, and at the same time improves the generalization ability of the model by maximizing the classification margin. From Figure 5 it can be seen that the model performs well in the overall classification performance, but there is still room for improvement. When randomly selecting ASD patients and healthy controls, the model has a 73.7% probability of correctly identifying ASD patients, and the overall performance is better than random classification. In the case of unbalanced data between the ASD patient group and the healthy control group, the classifier achieves a good balance between precision and recall.

[0094] To verify the effectiveness of the method proposed in the present invention, the study compared and evaluated SVM with classical machine learning methods RF and ELM, as Figure 6 shown;

[0095] As Figures 4 - 6 shown, the study used the same feature subset for six performance evaluation comparisons. It can be seen from the figure that SVM performs relatively well under most evaluation metrics, especially in terms of accuracy, AUC, specificity, and F1-score, demonstrating strong feature processing capabilities. However, SVM lags slightly behind the other two types of machine learning classifiers in terms of the sensitivity index, probably because SVM is more sensitive to the imbalance of sample classes and tends to classify more obvious classes during optimization, thus affecting the recognition ability of minority classes.

[0096] The study used the ABIDE I dataset for experiments and compared the results with existing studies. The comparison objects were classical classification studies based on nFCN and learning methods that used the ABIDE I dataset and detected ASD. The experimental results in Table 3 show that there are certain differences in the performance of different classification methods in ASD detection. The model (ours) of the present invention is the most prominent in terms of accuracy and sensitivity. Although its specificity is slightly lower than that of other models, it still outperforms other methods in terms of overall performance. In contrast, the traditional machine learning methods RF and ELM based on nFCN brain network features have good performance in terms of accuracy and sensitivity, but are slightly lacking in terms of specificity. The deep neural network DNN also shows good sensitivity and accuracy, close to our model. Generally speaking, although each method has differences in different metrics, our model has achieved the best performance in balancing sensitivity and accuracy, showing its potential in ASD detection.

[0097] Table 3 Comparison of classification performance of different methods in ASD detection

[0098]

[0099] When constructing the eFCN feature matrix and performing feature selection, the top 4000 feature subsets ranked by importance are used for brain region involvement statistics. Since eFCN represents the edges, that is, the connection strength between brain regions, and the features involve multiple brain regions, to discuss the abnormal connection between ASD pathology and brain regions, we select the top 20 features for display, as shown in Table 4:

[0100] Table 4 Comparison of classification performance of different methods in ASD detection

[0101]

[0102]

[0103] The above-mentioned brain regions involved are separately plotted as Figure 7 , and through the visualization graph, it can be obtained that the top 4 brain regions are the fusiform gyrus, superior frontal gyrus, middle temporal gyrus, and inferior temporal gyrus. Previous studies have shown that these brain regions are closely related to ASD pathology. Please refer to Figures 8 - 9 .

[0104] The fusiform gyrus is a key region in visual processing, especially playing an important role in face recognition and object recognition. Autistic patients often have difficulties in face recognition and face processing in social situations. The dysfunction of the fusiform gyrus in ASD may be related to the difficulty in face recognition. Therefore, the abnormality of the fusiform gyrus is considered an important neurobiological feature in ASD diagnosis.

[0105] The middle temporal gyrus plays an important role in functions such as language processing, face recognition, and social cognition in the brain. Research has shown that the abnormal activity of the middle temporal gyrus is closely related to the social and language communication disorders of autistic patients. For ASD individuals, the middle temporal gyrus may not be able to effectively process social information related to others, resulting in communication disorders. Therefore, the abnormal functional connection of the middle temporal gyrus may be a biomarker of ASD.

[0106] The influence of the superior frontal gyrus in ASD patients is an important research area. The superior frontal gyrus is located in the frontal lobe of the brain and is a brain region closely related to various higher cognitive functions such as cognitive control, emotion regulation, decision-making, and social behavior. Research has shown that there may be abnormalities in the structure and function of the superior frontal gyrus in ASD patients, and these abnormalities may be closely related to core symptoms such as social disorders, repetitive behaviors, and communication difficulties. Specifically, some studies have found that there may be a reduction in gray matter or abnormal functional connection in the superior frontal gyrus of ASD patients, which may lead to difficulties in their executive control tasks, emotion regulation, and social interaction. The abnormal function of the superior frontal gyrus may explain the challenges of ASD patients in aspects such as flexibility, planning, and decision-making, thus affecting their social adaptability.

[0107] The inferior temporal gyrus is closely related to object recognition, facial expression recognition, and other higher-order visual processing. Individuals with autism often have difficulty in recognizing facial expressions, which is related to the dysfunction of the inferior temporal gyrus. The study also found that the inferior temporal gyrus may show abnormal functional connectivity in the neural network of ASD individuals, indicating that there are obstacles in processing social visual information in this region.

[0108] The second aspect of the present invention provides an autism spectrum disorder risk detection system, including a data processing module and a risk detection module;

[0109] Data processing module: used to collect the detected fMRI data, perform preprocessing, and analyze the average time series of the preprocessed fMRI data;

[0110] Based on the average time series of the fMRI data, calculate the functional connectivity matrix between brain regions in the brain network centered on the edge;

[0111] Risk detection module: used to extract the features associated with ASD from the functional connectivity matrix to obtain the fMRI brain network feature set;

[0112] Input the fMRI brain network feature set into the classification model to identify the ASD risk of the detected fMRI data.

[0113] Some of the data in the above formula are calculated by removing the dimension and taking their numerical values. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the real situation; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0114] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for detecting the risk of autism spectrum disorder, characterized in that, It includes the following steps: Collect the detected fMRI data, perform preprocessing, and analyze the average time series of the preprocessed fMRI data; Based on the average time series of the fMRI data, construct a functional connectivity matrix in the brain network centered on edges; wherein, the edges are the functional connection edges between brain regions in the brain functional network; Extract the features associated with ASD from the functional connectivity matrix to obtain an fMRI brain network feature set; Input the fMRI brain network feature set into a classification model to identify the ASD risk of the detected fMRI data.

2. The autism spectrum disorder risk detection method according to claim 1, characterized in that, The preprocessing process of the fMRI data includes: Perform skull stripping, slice timing correction, motion correction, global mean intensity normalization, interference signal regression, and band-pass filtering on the fMRI data in sequence, and register the processed fMRI data to the standard anatomical space.

3. The autism spectrum disorder risk detection method according to claim 1, wherein The analysis of the average time series of the preprocessed fMRI data includes: Divide the preprocessed fMRI data into several brain regions, extract all voxels of each brain region from the preprocessed fMRI data according to the AAL atlas, and calculate the average value of the time signals of all voxels in each brain region at each time point.

4. The autism spectrum disorder risk detection method according to claim 3, characterized in that The calculation of the functional connectivity matrix between brain regions in the brain network centered on edges based on the average time series of the fMRI data includes: Calculate the product of the average time series of the blood oxygen levels of any two brain regions to obtain several edge time series; Calculate the cosine similarity between any two edge time series to obtain the functional connection strength between several edge time series; based on several functional connection strengths, construct a functional connectivity matrix.

5. The autism spectrum disorder risk detection method according to claim 4, wherein The average time series of each brain region is a time series that has been standardized.

6. The autism spectrum disorder risk detection method according to claim 1, wherein, The extraction of the features associated with ASD from the functional connectivity matrix includes: The random forest extracts the features associated with ASD from the functional connectivity matrix through its built-in feature importance evaluation mechanism to obtain an fMRI brain network feature set.

7. The autism spectrum disorder risk detection method according to claim 6, wherein The feature importance evaluation mechanism includes: Statistically count the cumulative number of times each fMRI brain network feature is selected as a node splitting feature in the integrated decision tree model; calculate the reduction in the sample class information entropy within the node before and after splitting to obtain the purity improvement metric value of the splitting; perform a weighted sum of the purity improvement metric values of each fMRI brain network feature on all splitting nodes to obtain the importance score of each fMRI brain network feature.

8. A method for detecting the risk of autism spectrum disorder according to claim 1, characterized in that, The identification of the ASD risk of the detected fMRI data includes: Input the fMRI brain network feature set into a classification model, and the classification model outputs a detection label; wherein, the classification model is constructed based on SVM.

9. The autism spectrum disorder risk detection method according to claim 1, wherein, The construction of the classification model based on SVM includes: Obtain several fMRI brain network feature sets and detection labels from historical data; Integrate the fMRI brain network feature sets and abnormal fMRI data into several groups of training data and test data; use the training data to train the SVM; use the test data to test the trained SVM, and adjust the SVM according to the test results; finally, obtain a classification model with the fMRI brain network feature set as the input and the detection label as the output.

10. A risk detection system for autism spectrum disorder, operating based on a risk detection method for autism spectrum disorder according to any one of claims 1-9, characterized in that, It includes a data processing module and a risk detection module; The data processing module: is used to collect the detected fMRI data, perform preprocessing, and analyze the average time series of the preprocessed fMRI data; Based on the average time series of the fMRI data, calculate the functional connectivity matrix between brain regions in the brain network centered on the edge; The risk detection module: is used to extract the features associated with ASD from the functional connectivity matrix to obtain the fMRI brain network feature set; Input the fMRI brain network feature set into the classification model to identify the ASD risk of the detected fMRI data.

Images