A method for selecting biomarkers of brain topological networks

Abstract

This invention discloses a method for selecting biometric features of brain topological networks, comprising the following steps: 1. Acquiring brain region biosignals and inter-brain region biosignals, using brain region biosignals as node features and inter-brain region biosignals as edge features; 2. Inputting node features and edge features into a feature selector to obtain initial weight values ​​corresponding to node features and edge features; 3. Updating the weight value of node feature i; 4. Sort all updated weight values ​​of node features and initial weight values ​​of edge features from largest to smallest, deleting the feature with the smallest weight value, and retaining the remaining node features and edge features; 5. Continuing step 2, iterating repeatedly until a total of K node features and edge features remain; 6. Using these node features and edge features as typical features of the brain network topology to assist in the discovery and diagnosis of mental illnesses. This invention is used to find typical feature data for subsequent highly reliable diagnostic classification.

Description

Technical Field

[0001] This invention relates to a method for selecting biometric features of brain topological networks, belonging to the category of human brain topological network feature extraction methods. Background Technology

[0002] Intelligent diagnostic classification of mental illnesses, as an artificial intelligence-based assisted diagnostic method, is actively developing towards higher accuracy, interpretability, and practicality. In particular, intelligent diagnostic methods using magnetic resonance imaging (MRI) technology are of great significance in exploring abnormalities in diseased brain regions, thereby enabling early detection, early diagnosis, and rapid treatment.

[0003] This invention addresses the diagnosis and classification of ADHD. Using low-frequency amplitude wave (ALFF) and functional connectivity (FC) data from the publicly available ADHD-200 database, brain topology structures were constructed in brain regions corresponding to the limbic network. This invention was then used to screen typical brain regions and connections within the limbic network associated with ADHD. Subsequently, the selected typical features were input into a binary hypothesis classification framework (Tang, Y., et al., ADHD classification using auto-encoding neural network and binary hypothesis testing, 2022), achieving an ADHD classification accuracy of over 95%. Furthermore, multiple medical studies have demonstrated that the limbic network is a major abnormal network in the development of ADHD (Wang, R., et al., Lifespan associations of resting-state brain functional networks with ADHD symptoms, 2022). The superior classification accuracy achieved by implementing the feature selection method of this invention on the limbic network further confirms the reliability of this feature selection method. Summary of the Invention

[0004] To address the issue that existing feature selection algorithms neglect the topological properties of brain networks, resulting in poor interpretability and low classification accuracy of features selected according to classification requirements, we introduce ALFF data as node features and FC data as edge features to construct a network topology and implement a classification algorithm based on biological features of brain topology networks.

[0005] The brain topological network biometric classification method of this invention can efficiently utilize various biological data calculated from MRI acquisition data, including feature data of brain regions such as volumetric density (VBM), low-frequency amplitude wave (ALFF), homomorphism (Reho), and degree centrality (DC), and feature data of brain regions such as functional connectivity (FC) and effector connectivity, to describe and analyze the patient's brain functional state from multiple perspectives. Because it uses multiple data sources, this invention can reliably identify typical feature data from these data for subsequent highly reliable diagnostic classification.

[0006] Simultaneously, by employing a brain topological structure representation using node and edge features, this invention allows for the priority removal of atypical edge features and the retention of node features, ensuring that the filtered brain topological structure does not contain isolated edge features (i.e., each filtered edge feature has at least one connected node feature). Since brain region feature data is essentially generated by calculating the biometric data of two corresponding connected brain regions according to a certain definition, the existence of a typical edge feature necessarily implies the existence of at least one corresponding typical node feature. Therefore, this invention conforms to this theory, thus effectively utilizing the characteristics of brain topological structures.

[0007] The technical solution of the present invention is as follows:

[0008] A method for selecting biometric features of brain topological networks includes the following steps:

[0009] Step 1: Obtain biological signals from brain regions and between brain regions. Use the biological signals from brain regions as signals on nodes of the brain topology network, i.e., node features, and use the biological signals from between brain regions as signals on edge connections of the brain topology network, i.e. edge features.

[0010] Step 2: Input the node features and edge features into the feature selector (the feature selector is an existing selector, such as the SVM-RFE feature selector, Lasso feature selector, etc.) to obtain the contribution of the node features and edge features to the target mental illness classification, and output them in the form of weights, thereby generating the initial weight values ​​corresponding to the node features and edge features.

[0011] Step 3: For the i-th node feature, update the weight value of node feature i based on the initial weight values ​​of all its connected edges and the other node feature k connected by the edges;

[0012] Step 4: Sort all updated node feature weights and initial edge feature weights from largest to smallest, delete the feature with the smallest weight, and keep the remaining node and edge features;

[0013] Step 5: Continue from step 2, iterating repeatedly until a total of K node features and edge features remain;

[0014] Step 6: Use these node features and edge features as typical features of brain network topology to assist in the discovery and diagnosis of mental illnesses.

[0015] Preferably, the specific steps for updating the weight value of node feature i in step 3 above are as follows:

[0016] Step 3-1: Construct the candidate update weight set S for node feature i i And initialize the weights w of node feature i. i Add to weight set S i Find all edges that are connected to node i;

[0017] Step 3-2: If node feature i has no edge connection, no operation is performed;

[0018] Step 3-3: If node feature i has an edge connection, and there is no other node feature connected to the edge, no operation is performed;

[0019] Steps 3-4: If node feature i is connected by an edge, and another node feature k exists on the edge connection, then compare the initial weight values ​​of the two nodes; if the weight w of node feature i is... i The weight w greater than the node feature k k Then the sum of the weights, w i +w k Add candidate update weight set S i ;

[0020] Steps 3-5: Traverse all edges connecting node feature i to obtain the weight set S. i All elements in;

[0021] Steps 3-6: Obtain the weight set S i The element with the largest value is used as the weight value after updating node feature i.

[0022] The beneficial effects of this invention are:

[0023] This invention proposes a method for selecting biometric features of brain topological networks. It employs a brain topological structure represented by node features and edge features. This method prioritizes the removal of atypical edge features and the retention of node features, ensuring that the selected brain topological structure does not contain isolated edge features (i.e., each selected edge feature has at least one connected node feature). Since brain region feature data is essentially calculated from the biometric data of two corresponding connected brain regions according to a certain definition, the existence of a typical edge feature necessarily implies the existence of at least one corresponding typical node feature. Therefore, this invention conforms to this theory, effectively utilizing the characteristics of brain topological structures. Furthermore, using ALFF and FC data on the publicly available ADHD-200 database, brain topological structures were constructed on brain regions corresponding to the limbic network. This method was then used to screen typical brain regions and connections within the limbic network associated with ADHD. Subsequently, the selected typical features were input into a binary hypothesis classification framework, achieving an ADHD classification accuracy of over 95%. Attached Figure Description

[0024] Figure 1 A diagram illustrating the binary hypothesis-based intelligent assisted diagnosis framework for ADHD based on the selection of biometric features from brain topological networks.

[0025] Figure 2 To select typical brain connectivity and brain region topologies using brain topological network biometrics, typical brain connectivity and brain region topologies were selected from the NYU (New York University Medical Center) site data in the ADHD-200 public database, utilizing the generated FC and ALFF biometric data. In the figures, brain regions with underlines are typical brain regions selected based on ALFF data, while brain regions without underlines are atypical brain regions. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0027] A method for selecting biometric features of brain topological networks includes the following steps:

[0028] Step 1: Obtain biological signals from brain regions and between brain regions. Use the biological signals from brain regions as signals on nodes of the brain topology network, i.e., node features, and use the biological signals from between brain regions as signals on edge connections of the brain topology network, i.e. edge features.

[0029] Step 2: Input the node features and edge features into the feature selector to obtain the degree of contribution of the node features and edge features to the target mental illness classification, and output them in the form of weights, thereby generating the initial weight values ​​corresponding to the node features and edge features;

[0030] Step 3: For the i-th node feature, update the weight value of node feature i based on the initial weight values ​​of all its connected edges and the other node feature k connected by the edges;

[0031] Step 3-1: Construct the candidate update weight set S for node feature i i And initialize the weights w of node feature i. i Add to weight set S i Find all edges that are connected to node i;

[0032] Step 3-2: If node feature i has no edge connection, no operation is performed;

[0033] Step 3-3: If node feature i has an edge connection, and there is no other node feature connected to the edge, no operation is performed;

[0034] Steps 3-4: If node feature i is connected by an edge, and another node feature k exists on the edge connection, then compare the initial weight values ​​of the two nodes; if the weight w of node feature i is... i The weight w greater than the node feature k k Then the sum of the weights, w i +w k Add candidate update weight set S i ;

[0035] Steps 3-5: Traverse all edges connecting node feature i to obtain the weight set S. i All elements in;

[0036] Steps 3-6: Obtain the weight set S i The element with the largest value is used as the weight value after updating node feature i.

[0037] Step 4: Sort all updated node feature weights and initial edge feature weights from largest to smallest, delete the feature with the smallest weight, and keep the remaining node and edge features;

[0038] Step 5: Continue from step 2, iterating repeatedly until a total of K node features and edge features remain;

[0039] Step 6: Use these node features and edge features as typical features of brain network topology to assist in the discovery and diagnosis of mental illnesses.

[0040] Step 6-1: Feature selection, such as Figure 1As shown, the test data are labeled as either healthy individuals (TD) or ADHD patients, and then fed into the feature selection module along with the training data. The brain topology network biometric feature selection method of this invention is used to obtain typical node and edge features of the training and test data under different test data labeling assumptions. In the training and test data, each individual contains their own ALFF and FC multimodal data.

[0041] Step 6-2: Feature extraction. Typical brain node and edge features of the training and test data under different test data label assumptions are used for feature dimensionality reduction via the feature extraction module to obtain high-level features of the training and test data. For example... Figure 1 As shown, feature extraction is divided into three parts: an encoding network, a decoding network, and a classification network. The encoding network is used to obtain high-level features; the decoding network requires that these high-level features can reconstruct the typical brain node features and edge features of the input of this module; the classification network is required to identify the labels of the training data and the hypothetical labels of the test data from the high-level features. The purpose of the above operations is to enhance the feature clustering ability of high-level features given the labels of the training data and the hypothetical labels of the test data.

[0042] Step 6-3: ADHD decision-making. Calculate the variability score (i.e., the ratio of intra-class divergence to inter-class divergence) of high-level features of training and test data under different test data label assumptions, compare the clustering performance, and select the hypothesis with the smaller score as the correct hypothesis.

[0043] As shown in Table 1, when testing the accuracy of ADHD classification and diagnosis, this invention uses the NYU (New York University Medical Center), KKI (Kennedy Krieger Institute), PU (Peking University), PU_1 (the first subset of the Peking University dataset), and NI (NeuroImage) site datasets from the ADHD-200 dataset to verify the effectiveness of the classification algorithm. Both the training and test datasets consist of low-frequency amplitude fluctuations (ALFF) data and functional connectivity (FC) data from the human limbic network (right hemisphere). One sample is selected from the obtained data as test data, and the remaining data are used as training data. The ADHD classification and diagnosis accuracy of this invention is verified using leave-one-out cross-validation. Meanwhile, comparisons were made with classic binary hypothesis methods, specifically the subspace projection-based binary hypothesis method (SP-BH) (Tang Y, et al., Identifying ADHD individuals from resting-state functional connectivity using subspaceclustering and binary hypothesis testing, 2021) and the auto-encoding neural network-based binary hypothesis method (AENet) (Tang Y, et al., ADHD classification using auto-encoding neural network and binary hypothesis testing, 2022), as shown in Table 1. Table 1 shows that SP-BH and AENet methods are based on feature selection algorithms for whole-brain connectivity, selecting 50 brain connectivity features for ADHD classification. SP-BH achieved a lower ADHD classification accuracy, with an average accuracy of 92.4%; AENet achieved a higher ADHD classification accuracy, with an average accuracy of 99.6%. In contrast, the method of this invention selectively selects features only from the edge network, using 35 features (9 brain region ALFF features and 26 brain connectivity features), achieving an average accuracy of 97.3%. Compared to AENet, this invention replaces the traditional SVM-RFE feature selection method in AENet with a brain topology network biometric selection method. It fully utilizes information from multimodal data, achieving the classification accuracy of AENet with 50 similar features using only 35 features, while reducing the number of features used by 30%, demonstrating the effectiveness of the biometric selection method in this invention. Therefore, Table 1 shows that the brain topology network biometric selection method of this invention can more efficiently identify abnormal features, thus better serving the classification of mental illnesses.

[0044] like Figure 2 As shown, when testing the accuracy of ADHD classification diagnosis, this invention validates the classification algorithm on the NYU site dataset in the ADHD-200 dataset. It utilizes the low-frequency amplitude fluctuation (ALFF) data and functional connectivity (FC) data of the human brain limbic network (right hemisphere) to obtain the brain network topology map through the biometric selection method of this invention, under the assumption of correct labeling of the test data. Figure 2 As can be seen, there is no isolated edge feature, which indicates that if there is a typical edge feature, there must be at least one corresponding typical node feature.

[0045] Table 1: Comparison of accuracy performance (%) between different ADHD classification and diagnostic methods and the method of this invention.

[0046] NYU PU PU_1 KKI NI Average SP-BH 96.2 95.8 91.7 86.7 91.6 92.4 AENet 99.8 99.6 99.6 99.8 99.3 99.6 Method of the present invention 98.4 94.1 98.6 99.0 96.2 97.3

[0047] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for selecting biometric features of brain topological networks, characterized in that... Includes the following steps: Step 1: Obtain biological signals from brain regions and between brain regions. Use the biological signals from brain regions as signals on nodes of the brain topology network, i.e., node features, and use the biological signals from between brain regions as signals on edge connections of the brain topology network, i.e. edge features. Step 2: Input the node features and edge features into the feature selector to obtain the degree of contribution of the node features and edge features to the target mental illness classification, and output them in the form of weights, thereby generating the initial weight values ​​corresponding to the node features and edge features; Step 3: For the i-th node feature, update the weight value of node feature i based on the initial weight values ​​of all its connected edges and the other node feature k connected by the edges; Step 4: Sort all updated node feature weights and initial edge feature weights from largest to smallest, delete the feature with the smallest weight, and keep the remaining node and edge features; Step 5: Continue from step 2, iterating repeatedly until a total of K node features and edge features remain; Step 6: Use these node features and edge features as typical features of brain network topology to assist in the discovery and diagnosis of mental illnesses.

2. The method for selecting biometric features of brain topological networks according to claim 1, characterized in that, The specific steps for updating the weight value of node feature i in step 3 are as follows: Step 3-1: Construct the candidate update weight set S for node feature i i And initialize the weights w of node feature i. i Add to weight set S i Find all edges that are connected to node i; Step 3-2: If node feature i has no edge connection, no operation is performed; Step 3-3: If node feature i has an edge connection, and there is no other node feature connected to the edge, no operation is performed; Steps 3-4: If node feature i is connected by an edge, and another node feature k exists on the edge connection, then compare the initial weight values ​​of the two nodes; if the weight w of node feature i is... i The weight w greater than the node feature k k Then the sum of the weights, w i +w k Add candidate update weight set S i ; Steps 3-5: Traverse all edges connecting node feature i to obtain the weight set S. i All elements in; Steps 3-6: Obtain the weight set S i The element with the largest value is used as the weight value after updating node feature i.

Images