Anxiety and depression comorbidity quantitative evaluation method based on bias label learning using brain images

Abstract

This invention discloses a method for quantifying anxiety and depression comorbidities using partial label learning based on brain imaging, belonging to the fields of medical image processing and artificial intelligence. The method is based on a dataset constructed using brain imaging measurements. It selects representative typical anxiety and depression samples with distinctive brain imaging features from clinically diagnosed groups and constructs a sparse nearest neighbor graph by comprehensively considering the relationships between each sample and its nearest ordinary and typical neighbor samples. Partial label learning modeling and optimization are performed through an iterative label propagation mechanism, ultimately outputting label confidence matrices for clinical anxiety patients, depression patients, and patients with comorbid anxiety and depression. This enables a quantitative assessment of the anxiety and depression levels in patients with comorbid anxiety and depression and completes the classification of comorbidity subtypes. This invention improves the accuracy and objectivity of anxiety and depression comorbidity assessment, providing auxiliary support for the clinical diagnosis, pathological mechanism research, and intervention target screening of anxiety and depression comorbidities.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence technology, specifically relating to an assessment method for quantifying anxiety and depression comorbidities using brain imaging based on partial labeling learning. Background Technology

[0002] Anxiety-depression comorbidity refers to a common mental disorder that simultaneously meets the diagnostic criteria for anxiety disorder and depression. It has a high incidence, complex clinical symptoms, and is difficult to treat, seriously affecting patients' physical and mental health and social functioning. Currently, the clinical assessment of anxiety-depression comorbidity mainly relies on the subjective judgment of clinicians and traditional scale scoring, which has the following technical limitations:

[0003] 1. The problem of ambiguous labeling is prominent: In clinical practice, the symptoms of most patients with comorbidities are mixed and difficult to be clearly defined as pure anxiety, pure depression or a certain type of comorbidity. This leads to ambiguity in assessment labels. Traditional assessment methods cannot effectively handle this biased labeling scenario and are prone to assessment bias.

[0004] 2. Lack of quantitative assessment indicators: Existing assessment methods are mostly based on qualitative descriptions or rating scales, which cannot output accurate quantitative assessment values ​​and are difficult to objectively reflect the severity of comorbidities and individual differences.

[0005] 3. Insufficient utilization of features: Traditional assessments often rely on single clinical scale features, ignoring the value of objective biological features such as brain imaging, failing to fully capture the neurobiological basis of comorbidities, resulting in low assessment accuracy;

[0006] 4. Poor individual heterogeneity and adaptability: There is significant neurobiological heterogeneity among comorbid patients. The pathological mechanisms and symptom manifestations of different subtypes of comorbid patients vary greatly. Traditional assessment methods use uniform assessment standards, which cannot be adapted to the individual characteristics of different subtypes and make it difficult to achieve individualized assessment.

[0007] Partial labeling learning, as an important branch of machine learning, can effectively handle problems such as label ambiguity and label noise. By mining the latent feature associations of samples, it can correct fuzzy labels and achieve accurate modeling. Applying partial labeling learning to the quantitative assessment of anxiety and depression comorbidities can effectively address the shortcomings of existing assessment methods. By combining multi-dimensional features of brain imaging, it can construct accurate and interpretable quantitative assessment models, providing support for the clinical diagnosis of comorbidities.

[0008] Therefore, developing a method for assessing the degree of anxiety and depression in patients with comorbid anxiety and depression based on partial labeling learning and combined with multidimensional brain imaging features to achieve accurate quantitative assessment has become an urgent need in the field of mental illness assessment. Summary of the Invention

[0009] To address the current shortcomings in assessing comorbid anxiety and depression, such as vague labels, lack of precise quantitative indicators, insufficient utilization of brain imaging features, and poor fit due to individual heterogeneity, this invention adopts the following technical solution:

[0010] A method for quantitative assessment of anxiety and depression comorbidity using brain imaging based on partial labeling learning includes the following steps:

[0011] Step 1, Dataset Construction: The data acquisition and brain imaging feature extraction methods used to construct the anxiety and depression comorbidity training dataset are as follows:

[0012] Step 1.1, Data Source: Collect brain imaging data from subjects who meet the clinical diagnostic criteria for anxiety and depression; based on the clinical diagnosis results, divide all subjects into anxiety group, depression group, and anxiety and depression comorbidity group;

[0013] Step 1.2, Brain image feature extraction: Calculate brain image measures from multiple levels of brain structure and brain function;

[0014] In terms of brain structural magnetic resonance imaging, voxel-level brain structural measurements and brain structural covariation networks were calculated for each subject to achieve a detailed characterization of the features of brain gray matter tissue and its co-variation patterns.

[0015] In terms of functional magnetic resonance imaging of the brain, local brain functional activity indicators were calculated for each subject, as well as brain function measures such as steady-state functional connectivity and dynamic functional connectivity at the level of brain regions and brain networks, to systematically characterize the brain functional activity patterns and the dynamic organizational features of its networks.

[0016] Brain structure measures, brain function measures, or a combination of both types of brain imaging measures, are used to construct each subject's [data / assessment / report]. 3D brain imaging feature vectors;

[0017] Step 1.3: The dataset is constructed by combining the brain imaging feature matrices of all participants' samples with their corresponding category labels (anxiety disorder, depression, and comorbid anxiety and depression); the total number of samples is defined as... .

[0018] Step 2, based on the datasets of participants with anxiety disorders, depression, and comorbid anxiety and depression, complete the modeling and optimization of the nearest neighbor propagation partial label learning algorithm, specifically including:

[0019] Step 2.1: Determine the initial typical anxiety subject sample set and the initial typical depression subject sample set, and construct the initial category center: Based on the feature matrix obtained in Step 1, within the samples of the anxiety group or depression group, calculate the Euclidean distance between the feature vector corresponding to each sample and the feature vectors corresponding to all other samples in the same category. Calculate the similarity of all Euclidean distances using the Gaussian kernel function and sum them up as the intra-class centrality index of the sample. The larger the index value, the closer the sample is to other samples in the same category as a whole, and the stronger its representativeness in the disease category.

[0020] The samples from the anxiety disorder group and the depression group were sorted in descending order according to the intraclass centrality index. The top 10% of the samples in each group were selected as typical anxiety disorder subjects and typical depression subjects, and the initial typical anxiety disorder sample set and the initial typical depression sample set were constructed.

[0021] The feature vector of the category center of the anxiety group is obtained by taking the mean of the feature vectors of all typical anxiety disorder subjects. Similarly, we obtain the feature vector of the category center of the depression group by taking the mean of the feature vectors of all typical depression subjects. Step 2.2 determines the samples to be propagated in this layer, performs partial label initialization of the samples to be propagated in this layer, and sets the candidate label set and initial confidence vector for the samples to be propagated in this layer;

[0022] Determine the samples to be propagated in this layer: First, combine the current set of typical samples with a set... Find the union of the sets, then extract the nearest neighbors of each sample in the union set, and then combine these nearest neighbors with the set. Find the union of the samples to be propagated in this layer. This indicates the number of samples to be propagated in this layer. (The last part, "when iterating round," appears to be unrelated and likely refers to a different topic.) =1, For an empty set, when When >1, the set This equals the set of samples to be propagated in this layer. When the set... The number of elements is equal to Then, perform the final iterations from steps 2.2 to 2.4.

[0023] Perform partial label initialization for the samples to be propagated in this layer: Let... Number the label. Corresponding to the anxiety tag, For the depression label, the candidate labels of typical samples are unique and the same as their initial labels. Therefore, the candidate label of typical anxiety samples is set as 1, the candidate label of typical depression samples is set as 2, and the candidate label set of the remaining samples includes 1 and 2.

[0024] Set the label confidence matrix of the samples to be propagated in this layer. , The initial label confidence matrix is ​​set as follows: , Indicates the number of categories, =2; Satisfy the following formula: ;

[0025] in, Indicates the first Each sample belongs to the label value The The initial likelihood of the class; For the first The tag value of each tag. For the first The candidate label set for each sample, then For the first The number of labels in the candidate label set for each sample. Through this process, the initial confidence vector of the typical anxiety disorder subject sample is set to... The initial confidence vector of a typical sample of depressed subjects is set as follows: The initial confidence vector for the remaining samples is set to Step 2.3: Construct the sample to be propagated in this layer. Nearest neighbor graph;

[0026] First, construct a fully connected graph. The subject samples to be propagated in this layer are used as nodes in the graph. Weighted edges between nodes The similarity is calculated using the Gaussian kernel function based on the feature vectors between samples.

[0027] in, For the first The sample and the first Similarity between samples The Euclidean distance between the feature vectors of the two brain images is given; the similarity between all nodes forms the weight matrix of the fully connected graph. ;

[0028] Secondly, construct the ordinary nearest neighbor weight matrix. Here, for the weight matrix After normalization, select the sample with the highest similarity. A normal nearest neighbor sample, Include each node and its nearest neighbor Similarity between nearest neighbor nodes It is obtained from the following formula:

[0029] in, For the first one sample A set of ordinary nearest neighbor samples;

[0030] Next, a typical nearest neighbor weight matrix is ​​constructed. Here, the sample with the highest similarity is selected. A typical sample, Include each node and its nearest neighbor Similarity between typical sample nodes It is obtained from the following formula: ;

[0031] in, For the first one sample A typical set of nearest neighbors;

[0032] Finally, build Nearest neighbor graph, which combines two matrices and By fusion Weight matrix corresponding to the nearest neighbor graph , ;here , It retains the correlation of local sample features while strengthening the guiding role of typical sample information;

[0033] Step 2.4: Perform iterative label propagation on the samples to be propagated in this layer to determine their label confidence matrix;

[0034] Step 2.4.1, Label Propagation Iteration: Based on Weight matrix corresponding to the nearest neighbor graph Perform the label propagation operation of the partial label learning algorithm.

[0035] No. Temporary label confidence matrix The calculation method is as follows: , The parameter is used to balance the first... The updated label confidence matrix after the second propagation and the initial label confidence matrix For the The impact of the round confidence matrix The larger the value, the stronger the influence of the tag information from the previous round of propagation on the current round;

[0036] The calculated temporary label confidence matrix Perform normalization to obtain the first... Wheel label confidence matrix The normalization calculation formula is:

[0037] ;

[0038] in, Indicates the first Normalized label confidence matrix The Middle The sample belongs to the first Confidence values ​​for each label; Indicates the first Each sample in its candidate label set The sum of all temporary confidence values ​​within the range.

[0039] Step 2.4.2: Calculate the similarity between each atypical anxiety sample and the anxiety category center based on the Gaussian kernel function; sort the atypical anxiety samples in descending order of similarity to the anxiety category center, select the top 15% of samples, and calculate the union of these samples with the existing typical anxiety sample set as the updated typical anxiety samples. The update of the typical depression sample set is consistent with the above. Calculate the updated category centers based on the updated typical sample set.

[0040] Step 2.4.3: When the conditions are met, the iteration terminates, and the final label confidence matrix is ​​output, along with the category labels determined by the model for subjects clinically diagnosed as anxious and clinically diagnosed as depressed.

[0041] When the Wheel and the first The difference in the confidence matrix of the round label satisfies The iteration stops when the preset maximum number of iterations is reached. Otherwise, proceed to step 2.2.

[0042] Step 2.4.4: Output the final label confidence matrix. ;

[0043] Based on the final label confidence matrix The category labels for subjects clinically diagnosed with anxiety and those clinically diagnosed with depression are determined based on the anxiety confidence level of the sample. Greater than the confidence level of depression When the condition is met, the model will classify the sample as anxiety disorder; otherwise, it will classify the sample as depression.

[0044] Step 3, under different parameters and Take value Construction of the nearest neighbor graph and label propagation, determination of optimal parameters and their corresponding models, and the results obtained based on the optimal model. Quantitative analysis of the degree of depression and anxiety in patients with comorbid anxiety and depression;

[0045] Step 3.1, with different parameters and Under the given value, proceed according to step 2. The nearest neighbor graph is constructed and labels are propagated, outputting the label confidence matrix under different parameters. And the category labels determined by the model for subjects clinically diagnosed with anxiety and subjects clinically diagnosed with depression;

[0046] Step 3.2, determine the optimal parameters and their corresponding models: compare the disease category labels of anxiety and depression patients based on clinical indicators with the predicted labels output by the models under different parameters, and calculate the classification accuracy respectively; select the parameters that make the classification accuracy of anxiety and depression patients reach the highest as the optimal parameters, and determine the corresponding model as the optimal model;

[0047] Step 3.3: Quantify the degree of anxiety and depression in patients with comorbid anxiety and depression based on the optimal model;

[0048] For patients with comorbid anxiety and depression, the optimal model was obtained Output and As a quantitative measure of their anxiety and depression levels; if their If so, the subject is determined to be a patient with comorbid anxiety and depression with depressive tendencies; if If so, the subject is determined to be a patient with anxiety-prone comorbid depression.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] 1. This invention effectively overcomes the inherent limitations of traditional classification and diagnosis by using a nearest-neighbor propagation partial label learning algorithm, solving the technical challenges of ambiguous comorbidity labels and low diagnostic consistency. Existing technologies use traditional classification and diagnosis systems such as the Diagnostic and Statistical Manual of Mental Disorders (5th Edition) and the International Classification of Diseases (11th Revision), employing a discrete category division logic of "either / or," which cannot adapt to the spectrumal nature of anxiety and depression comorbidity samples, characterized by "continuous symptom distribution and ambiguous boundaries," leading to ambiguous comorbidity label definitions. This invention, through a nearest-neighbor propagation partial label learning algorithm, assigns differentiated initial partial labels to depression, anxiety, and anxiety-depression comorbidity samples based on brain imaging. This objectively reflects the label ambiguity of comorbidity samples and can accurately distinguish different subtypes such as comorbidity with a predominantly depressive or predominantly anxious nature, completely breaking through the constraints of traditional classification and diagnosis frameworks and significantly improving the objectivity, consistency, and accuracy of comorbidity assessment and diagnosis.

[0051] 2. Achieving individualized and precise quantification of the severity of anxiety and depression comorbidity, solving the challenge of quantifying comorbidity in clinical practice. Existing technologies mostly focus only on "whether comorbidity exists," failing to quantify the symptom weights and severity of comorbidity in individuals, and thus cannot adapt to the clinical reality of overlapping comorbidity symptoms and significant individual heterogeneity. This invention relies on an iterative label propagation mechanism based on a nearest neighbor propagation partial label learning algorithm. Through typical sample guidance and iterative confidence calibration, it outputs individualized confidence indices for depression and anxiety, accurately characterizing the symptom tendency subtypes of individuals with comorbidity. Simultaneously, it constructs a "brain imaging features-comorbidity symptom weight" correlation model to achieve precise quantification of comorbidity severity. This technically solves the quantitative dilemma caused by overlapping comorbidity symptoms in clinical practice, providing a reliable quantitative basis for individualized assessment of comorbidity.

[0052] 3. Adapting to the heterogeneity of comorbid individuals, accurately capturing the specific brain network characteristics of different comorbid subtypes. Existing studies on anxiety and depression comorbidity mostly adopt population analysis methods, oversimplifying the complexity of comorbid neurophenotypes and masking the specific brain feature changes of different comorbid individuals. This invention constructs a " A number of ordinary nearest neighbors + The sparse nearest neighbor graph of "typical nearest neighbors" not only preserves the feature associations of local samples, but also strengthens the reliable label guidance role of typical samples. It can effectively capture the specific brain features of different comorbidity subtypes (comorbidity with anxiety, comorbidity with depression), avoid the feature masking problem caused by the group average effect, and achieve precise adaptation to the heterogeneity of comorbidity individuals.

[0053] 4. An innovative algorithm architecture fills the gap in the application of existing partial labeling algorithms for comorbidity quantification. Existing partial labeling propagation algorithms are not specifically designed for the comorbidity quantification of anxiety and depression. The proposed algorithm, based on nearest neighbor propagation partial labeling learning, uses typical anxiety and depression samples for screening... The nearest neighbor graph fusion construction and iterative optimization mechanism is the first to apply the partial label propagation technology to the problem of quantifying the degree of anxiety and depression comorbidity in individuals based on brain imaging, filling the technical gap of existing algorithms in this scenario. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0056] Figure 1 This is an overall flowchart of the method for quantitative assessment of anxiety and depression comorbidity based on partial labeling learning using brain imaging provided in this embodiment of the invention;

[0057] Figure 2 This is a diagram showing the classification results of the optimal parameter selection and optimal model in this embodiment of the invention. Figure 2a. This demonstrates the impact of different K and N parameters on the model's classification performance; Figure 2 b shows the confusion matrix of the model's classification results for the test set samples under the optimal parameter configuration.

[0058] Figure 3 This is the process of the anxiety and depression confidence matrix of a portion of the samples from initialization to its steady state. Detailed Implementation

[0060] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0061] A method for quantitative assessment of anxiety and depression comorbidity using brain imaging based on partial labeling learning includes the following steps:

[0062] Step 1, Dataset Construction:

[0063] The data acquisition and brain imaging feature extraction methods used to construct the anxiety and depression comorbidity dataset are as follows:

[0064] Step 1.1, Data Source:

[0065] Brain imaging data were collected from 925 participants aged 18-65 years who met the diagnostic criteria of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Based on clinical diagnoses, all participants were strictly divided into three groups: a clinically diagnosed depression group (502 cases), a clinically diagnosed anxiety group (246 cases), and a comorbid anxiety and depression group (177 cases). The total sample size was defined as [missing data]. .

[0066] Step 1.2, Brain Imaging Feature Extraction

[0067] Brain imaging measures are calculated from a functional perspective, as follows:

[0068] Brain region segmentation framework: Two complementary templates are integrated to define 116 brain regions. The cortical region uses the Schaefer 100 region template to define 100 brain regions, and the subcortical region uses the Melbourne Subcortex Atlas to define 16 nuclei.

[0069] Functional connectivity calculation: For time-series data of 116 brain regions for each subject, the Pearson correlation coefficient of each pair of brain region time series was calculated to generate a 116×116 symmetric functional connectivity matrix (the matrix element values ​​range from [-1,1], positive values ​​indicate functional synergy between brain regions, and negative values ​​reflect functional inhibition).

[0070] Feature vector generation: Extract the upper triangular elements of the 116×116 functional connectivity matrix to form a one-dimensional feature vector, which serves as the brain functional imaging feature vector of the subject.

[0071] Based on the above brain functional characteristics, a D-dimensional brain imaging feature vector is constructed for each subject, where (The number of upper triangular elements in a 116×116 matrix).

[0072] Step 1.3: Construct the brain imaging feature matrix and dataset.

[0073] The feature matrix is ​​composed of the D-dimensional brain imaging features of all 925 subjects. The characteristic matrix of the first Behavior No. Feature vector of each sample ; the brain imaging feature matrix The data is matched with the category labels corresponding to the subjects (subjects clinically diagnosed with depression, subjects clinically diagnosed with anxiety, and subjects with comorbid anxiety and depression) to form a dataset for the study of comorbid anxiety and depression.

[0074] Step 2: Modeling and Optimization of the Nearest Neighbor Propagation Partial Label Learning Algorithm

[0075] Based on brain imaging datasets of subjects clinically diagnosed with anxiety, subjects clinically diagnosed with depression, and subjects with comorbid anxiety and depression, modeling and optimization of a nearest neighbor propagation partial label learning algorithm were completed, specifically including:

[0076] Step 2.1, Initial Typical Sample Selection and Class Center Construction

[0077] Based on the feature matrix obtained in step 1 Perform the following operations:

[0078] Within-class centrality index calculation: Within the sample groups of subjects clinically diagnosed with anxiety and subjects clinically diagnosed with depression, the Euclidean distance between the feature vector corresponding to each sample and the feature vectors corresponding to all other samples in the same class is calculated. The similarity of all Euclidean distances is calculated using a Gaussian kernel function, and then the sums are obtained to obtain the within-class centrality index of the sample. The larger the index value, the closer the sample is to other samples in the same class as a whole, and the stronger its representativeness in the disease category.

[0079] Initial typical sample screening: The samples of the clinically diagnosed anxiety sample group and the clinically diagnosed depression sample group were sorted in descending order according to the intraclass centrality index. The top 10% of the samples in each group were selected as typical anxiety sample and typical depression sample, respectively, to construct the initial typical anxiety sample set and the initial typical depression sample set. The typical samples retained their original category labels and did not participate in the subsequent fuzzy label assignment.

[0080] Initial category center generation: The feature vectors of all typical anxious subjects are averaged to obtain the category center feature vectors of the anxious subject sample group; similarly, the feature vectors of all typical depressed subjects are averaged to obtain the category center feature vectors of the depressed subject sample group.

[0081] Step 2.2: Determine the samples to be propagated in this layer and initialize the partial labels.

[0082] Determine the samples to be propagated in this layer: First, combine the current set of typical samples with the set... Find the union of the sets, then extract the nearest neighbors of each sample in the union set, and then combine these nearest neighbors with the set. The union of these unions yields the set of samples to be propagated in this layer. Let the number of samples to be propagated in this layer be denoted as . When the iteration round At that time, the set It is an empty set; when At that time, the set It equals the set of samples to be propagated in the previous round of this layer; when the set The number of elements equals the total number of samples. Then, perform the final iteration from step 2.2 to 2.4.

[0083] Partial label initialization: Initialize the partial labels of the samples to be propagated in this layer: Let... Number the label. Corresponding to the anxiety tag, For the depression label, the candidate labels of typical samples are unique and the same as their initial labels. Therefore, the candidate label of typical anxiety samples is set as 1, the candidate label of typical depression samples is set as 2, and the candidate label set of the remaining samples includes 1 and 2.

[0084] Set the label confidence matrix of the samples to be propagated in this layer. , The initial label confidence matrix is ​​set as follows: , Indicates the number of categories, ;

[0085] Satisfy the following formula:

[0086] ;

[0087] in, Indicates the first Each sample belongs to the label value The The initial likelihood of the class; For the first The tag value of each tag. For the first The candidate label set for each sample, then For the first The number of labels in the candidate label set for each sample. Through this process, the initial confidence vector of the typical anxiety disorder subject sample is set to... The initial confidence vector of a typical sample of depressed subjects is set as follows: The initial confidence vector for the remaining samples is set to .

[0088] Step 2.3, Nearest Neighbor Graph Construction

[0089] Constructing for the samples to be propagated in this layer The process for a nearest neighbor graph is as follows:

[0090] Fully Connected Graph Construction: Constructing a fully connected graph The subject samples to be propagated in this layer are used as nodes in the graph. Weighted edges between nodes The similarity is calculated using the Gaussian kernel function based on the feature vectors between samples.

[0091] ;

[0092] in, For the first The sample and the first Similarity between samples The Euclidean distance between the feature vectors of the two brain images is given; the similarity between all nodes forms the weight matrix of the fully connected graph. ;

[0093] Ordinary nearest neighbor weight matrix Construction: For the weight matrix After normalization, select the sample with the highest similarity. A normal nearest neighbor sample, Include each node and its nearest neighbor Similarity between nearest neighbor nodes It is obtained from the following formula: .

[0094] in, For the first one sample A set of ordinary nearest neighbor samples.

[0095] Typical nearest neighbor weight matrix Construction: Filter the samples with the highest similarity to each other. A typical sample, Include each node and its nearest neighbor Similarity between typical sample nodes It is obtained from the following formula:

[0096] ;

[0097] in, For the first one sample A typical set of nearest neighbor samples.

[0098] Finally, build Nearest neighbor graph, which combines two matrices and By fusion Weight matrix corresponding to the nearest neighbor graph , ;here , It retains the correlation of local sample features while strengthening the guiding role of typical sample information.

[0099] Step 2.4, Iterative Label Propagation and Result Output

[0100] Step 2.4.1, Label Propagation Iteration

[0101] based on Weight matrix corresponding to the nearest neighbor graph Perform the label propagation operation of the partial label learning algorithm:

[0102] No. Temporary label confidence matrix The calculation method is as follows: , The parameter is used to balance the first... The updated label confidence matrix after the second propagation and the initial label confidence matrix For the The impact of the round confidence matrix The larger the value, the stronger the influence of the tag information from the previous round of propagation on the current round;

[0103] The calculated temporary label confidence matrix Perform normalization to obtain the first... Wheel label confidence matrix The normalization calculation formula is:

[0104] ;

[0105] in, Indicates the first Normalized label confidence matrix The Middle The sample belongs to the first Confidence values ​​for each label; Indicates the first Each sample in its candidate label set The sum of all temporary confidence values ​​within the range.

[0106] Step 2.4.2, updating typical samples and category centers

[0107] The similarity between each atypical anxiety sample and the anxiety category center is calculated using a Gaussian kernel function. The atypical anxiety samples are then sorted in descending order of their similarity to the anxiety category center. The top 15% of these samples are then joined with the existing typical anxiety sample set to form the updated typical anxiety sample set. The update method for the typical depression sample set is the same. Based on the updated typical sample set, the updated feature vectors of the category centers for the anxiety and depression subject sample groups are calculated respectively.

[0108] Step 2.4.3, Iteration Termination Condition

[0109] When the conditions are met, the iteration terminates, and the final label confidence matrix is ​​output, along with the category labels determined by the model for subjects clinically diagnosed as anxious and those clinically diagnosed as depressed.

[0110] When the Wheel and the first The difference in the confidence matrix of the round label satisfies The iteration stops when the preset maximum number of iterations is reached. Otherwise, proceed to step 2.2.

[0111] Step 2.4.4: Output the final label confidence matrix. ;

[0112] Based on the final label confidence matrix The category labels for subjects clinically diagnosed with anxiety and those clinically diagnosed with depression are determined based on the anxiety confidence level of the sample. Greater than the confidence level of depression If the condition is met, the model will classify the sample as anxiety disorder; otherwise, it will classify the sample as depression.

[0113] Step 3, under different parameters and Take value Construction of the nearest neighbor graph and label propagation, determination of optimal parameters and their corresponding models, and the results obtained based on the optimal model. A quantitative analysis of the degree of depression and anxiety was conducted on patients with comorbid anxiety and depression.

[0114] Step 3.1, with different parameters and Under the given value, proceed according to step 2. The nearest neighbor graph is constructed and labels are propagated, outputting the label confidence matrix under different parameters. And the category labels determined by the model for subjects clinically diagnosed with anxiety and subjects clinically diagnosed with depression;

[0115] Step 3.2, determine the optimal parameters and their corresponding models: compare the disease category labels of anxiety and depression patients based on clinical indicators with the predicted labels output by the models under different parameters, and calculate the classification accuracy respectively; select the parameters that make the classification accuracy of anxiety and depression patients reach the highest as the optimal parameters, and determine the corresponding model as the optimal model;

[0116] Step 3.3: Quantify the degree of anxiety and depression in patients with comorbid anxiety and depression based on the optimal model;

[0117] For patients with comorbid anxiety and depression, the optimal model was obtained Output and As a quantitative measure of their anxiety and depression levels; if their If so, the subject is determined to be a patient with comorbid anxiety and depression with depressive tendencies; if If so, the subject is determined to be a patient with anxiety-prone comorbid depression.

[0118] Through different Value and The calculation is performed under different parameter combinations. Clinical labels for patients with clinical anxiety and clinical depression are compared with labels predicted by the model based on images. Accuracy, precision, recall, and F1 score are calculated to select the parameter combination with optimal classification performance. The impact of different parameter combinations on classification performance is as follows: Figure 2 As shown in figure a. Experimental results show that the optimal parameter combination is: , The classification confusion matrix obtained under this parameter configuration is as follows: Figure 2As shown in b, the vertical axis represents pure anxiety, pure depression, 3-1 comorbid anxiety, and 3-2 comorbid depression. The horizontal axis represents clinical anxiety, clinical depression, and clinical comorbidity. The confusion matrix results show that in the classification of pure anxiety and pure depression, 38 clinical anxiety samples were classified as pure depression under the aid of imaging analysis, 4 clinical depression samples were classified as pure anxiety, and the imaging classification results of most other anxiety and depression samples were consistent with the clinical diagnosis. For 177 patients with clinical comorbidity, the model classified 23 cases as comorbid anxiety and 154 cases as comorbid depression based on imaging features. The confidence matrix when the optimal parameters finally reach a stable state is shown in the figure. Figure 3 The first and second columns of the matrix serve as quantitative measures of anxiety and depression levels in patients with comorbid anxiety and depression, respectively, with values ​​ranging from [0,1]. These values ​​reflect the model's predictive confidence levels for the severity of individual anxiety and depressive symptoms; higher values ​​indicate a more significant tendency towards the corresponding symptoms. The results show that for patients with clinical anxiety and clinical depression, the model's predictions are generally consistent with clinical diagnoses, indicating a high degree of consistency between imaging and clinical judgment. Furthermore, for patients with comorbidities, the model can further assess the relative tendency of anxiety and depression within the comorbidity, thus providing support for more precise clinical auxiliary diagnosis and treatment.

[0119] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments described above. For those skilled in the art, various modifications made within the spirit and scope of the invention as defined and determined by the appended claims are readily apparent, and all inventions utilizing the concept of this invention should be included within the scope of protection of this invention.

Claims

1. A method for anxiety-depression comorbidity quantitative assessment based on bias label learning using brain images, characterized in that, Includes the following steps: Step 1, Dataset Construction: The data acquisition and brain imaging feature extraction methods used to construct the anxiety and depression comorbidity training dataset are as follows: Step 1.1, Data Source: Collect brain imaging data from subjects who meet the clinical diagnostic criteria for anxiety and depression; based on the clinical diagnosis results, divide all subjects into anxiety group, depression group, and anxiety and depression comorbidity group; Step 1.2, Brain image feature extraction: Calculate brain image measures from multiple levels of brain structure and brain function; In terms of brain structural magnetic resonance imaging, voxel-level brain structural measurements and brain structural covariation networks were calculated for each subject to achieve a detailed characterization of the features of brain gray matter tissue and its co-variation patterns. In terms of functional magnetic resonance imaging of the brain, local brain functional activity indicators were calculated for each subject, as well as brain functional measures based on steady-state and dynamic functional connectivity at the level of brain regions and brain networks, to systematically characterize the brain functional activity patterns and the dynamic organizational features of its networks. Brain structure measures, brain function measures, or a combination of both types of brain imaging measures, are used to construct each subject's [data / assessment / report]. 3D brain imaging feature vectors; Step 1.3: The dataset is constructed by combining the brain imaging feature matrices of all subjects' samples with the corresponding category labels of the subjects; the total number of samples is defined as... ; Step 2, based on the datasets of participants with anxiety disorders, depression, and comorbid anxiety and depression, complete the modeling and optimization of the nearest neighbor propagation partial label learning algorithm, specifically including: Step 2.1: Determine the initial typical anxiety subject sample set and the initial typical depression subject sample set, and construct the initial category center: Based on the feature matrix obtained in Step 1, within the samples of the anxiety group or depression group, calculate the Euclidean distance between the feature vector corresponding to each sample and the feature vectors corresponding to all other samples in the same category. Calculate the similarity of all Euclidean distances using the Gaussian kernel function and sum them up as the intra-class centrality index of the sample. The larger the index value, the closer the sample is to other samples in the same category as a whole, and the stronger its representativeness in the disease category. The samples from the anxiety disorder group and the depression group were sorted in descending order according to the intraclass centrality index. The top 10% of the samples in each group were selected as typical anxiety disorder subjects and typical depression subjects, and the initial typical anxiety disorder sample set and the initial typical depression sample set were constructed. The feature vector of the category center of the anxiety group is obtained by taking the mean of the feature vectors of all typical anxiety disorder subjects; similarly, the feature vector of the category center of the depression group is obtained by taking the mean of the feature vectors of all typical depression subjects. Step 2.2: Determine the samples to be propagated in this layer, perform partial label initialization of the samples to be propagated in this layer, and set the candidate label set and initial confidence vector for the samples to be propagated in this layer. Determine the samples to be propagated in this layer: First, combine the current set of typical samples with a set... Find the union of the sets, then extract the nearest neighbors of each sample in the union set, and then combine these nearest neighbors with the set. Find the union of the sets of samples to be propagated in this layer. This indicates the number of samples to be propagated in this layer; when the iteration round... =1, For an empty set, when When >1, the set It equals the set of samples to be propagated in this layer; when the set The number of elements is equal to Then, perform the final iteration from steps 2.2 to 2.4; Perform partial label initialization for the samples to be propagated in this layer: Let... Number the label. Corresponding to the anxiety tag, For the depression label, the candidate labels of typical samples are unique and the same as their initial labels. Therefore, the candidate label of typical anxiety samples is set as 1, the candidate label of typical depression samples is set as 2, and the candidate label set of the remaining samples includes 1 and 2. Set the label confidence matrix of the samples to be propagated in this layer. , The initial label confidence matrix is ​​set as follows: , Indicates the number of categories, =2; Satisfy the following formula: ； in, Indicates the first Each sample belongs to the label value The The initial likelihood of the class; For the first The tag value of each tag. For the first The candidate label set for each sample, then For the first The number of labels in the candidate label set for each sample; through this process, the initial confidence vector of the typical anxiety disorder subject sample is set to The initial confidence vector of a typical sample of depressed subjects is set as follows: The initial confidence vector for the remaining samples is set to Step 2.3: Construct the sample to be propagated in this layer. Nearest neighbor graph; First, construct a fully connected graph. The subject samples to be propagated in this layer are used as nodes in the graph. Weighted edges between nodes The similarity is calculated using the Gaussian kernel function based on the feature vectors between samples. ； in, For the first The sample and the first Similarity between samples The Euclidean distance between the feature vectors of the two brain images is given; the similarity between all nodes forms the weight matrix of the fully connected graph. ; Secondly, construct the ordinary nearest neighbor weight matrix. Here, for the weight matrix After normalization, select the sample with the highest similarity. A normal nearest neighbor sample, Include each node and its nearest neighbor Similarity between nearest neighbor nodes It is obtained from the following formula: ； in, For the first one sample A set of ordinary nearest neighbor samples; Next, a typical nearest neighbor weight matrix is ​​constructed. Here, the sample with the highest similarity is selected. A typical sample, Include each node and its nearest neighbor Similarity between typical sample nodes It is obtained from the following formula: ； in, For the first one sample A typical set of nearest neighbors; Finally, build Nearest neighbor graph, which combines two matrices and By fusion Weight matrix corresponding to the nearest neighbor graph , ;here , It retains the correlation of local sample features while strengthening the guiding role of typical sample information; Step 2.4: Perform iterative label propagation on the samples to be propagated in this layer to determine their label confidence matrix; Step 2.4.1, Tag Propagation Iteration: Based on Weight matrix corresponding to the nearest neighbor graph Perform the label propagation operation of the partial label learning algorithm. No. Temporary label confidence matrix The calculation method is as follows: , The parameter is used to balance the first... The updated label confidence matrix after the second propagation and the initial label confidence matrix For the first The impact of the round confidence matrix The larger the value, the stronger the influence of the tag information from the previous round of propagation on the current round; The calculated temporary label confidence matrix Perform normalization to obtain the first... Wheel label confidence matrix The normalization calculation formula is: ； in, Indicates the first Normalized label confidence matrix The Middle The sample belongs to the first Confidence values ​​for each label; Indicates the first Each sample in its candidate label set The sum of all temporary confidence values ​​within the range; Step 2.4.2: Calculate the similarity between each atypical anxiety sample and the anxiety category center based on the Gaussian kernel function; sort the atypical anxiety samples in descending order of similarity to the anxiety category center, select the top 15% of samples and calculate the union with the existing typical anxiety sample set as the updated typical anxiety samples, and update the typical depression sample set in the same way as above; calculate the updated category center based on the updated typical sample set. Step 2.4.3: When the condition is met, the iteration terminates, and the final label confidence matrix is ​​output, along with the category labels determined by the model for subjects clinically diagnosed as anxiety and subjects clinically diagnosed as depression. When the Wheel and the first The difference in the confidence matrix of the round label satisfies The iteration stops when the number of iterations reaches the preset maximum number of iterations; otherwise, proceed to step 2.

2. Step 2.4.4: Output the final label confidence matrix. ; Based on the final label confidence matrix The category labels for subjects clinically diagnosed with anxiety and those clinically diagnosed with depression are determined based on the anxiety confidence level of the sample. Greater than the confidence level of depression When the condition is met, the model will classify the sample as anxiety disorder; otherwise, it will classify the sample as depression. Step 3, under different parameters and Take value Construction of the nearest neighbor graph and label propagation, determination of optimal parameters and their corresponding models, and the results obtained based on the optimal model. Quantitative analysis of the degree of depression and anxiety in patients with comorbid anxiety and depression; Step 3.1, with different parameters and Under the given value, proceed according to step 2. The nearest neighbor graph is constructed and labels are propagated, outputting the label confidence matrix under different parameters. And the category labels determined by the model for subjects clinically diagnosed with anxiety and subjects clinically diagnosed with depression; Step 3.2, determine the optimal parameters and their corresponding models: compare the disease category labels of anxiety and depression patients based on clinical indicators with the predicted labels output by the models under different parameters, and calculate the classification accuracy respectively; select the parameters that make the classification accuracy of anxiety and depression patients reach the highest as the optimal parameters, and determine the corresponding model as the optimal model; Step 3.3: Quantify the degree of anxiety and depression in patients with comorbid anxiety and depression based on the optimal model; For patients with comorbid anxiety and depression, the optimal model was obtained Output and As a quantitative measure of their anxiety and depression levels; if their If so, the subject is determined to be a patient with comorbid anxiety and depression with depressive tendencies; if If so, the subject is determined to be a patient with anxiety-prone comorbid depression.

Images