A method and system for evaluating cognitive function of depression based on macroscopic brain dynamics characteristics

Abstract

The application discloses a kind of depression cognitive function evaluation methods based on macroscopic brain dynamics characteristics, belong to medical image processing and mental illness auxiliary evaluation technical field.The method is based on resting-state functional magnetic resonance imaging data, first to original data is standardized and preprocessed, obtain the BOLD time series of brain region level;Then system extracts multidimensional dynamics characteristics, constructs macroscopic brain dynamics index by principal component analysis, for representing brain core dynamic mode;Introduce cross-subject pattern alignment technology, eliminate principal component direction uncertainty, realize the unity of characteristic space;On this basis, construct multi-scale brain dynamics characteristics from brain region level and functional network level;Finally, establish regression model to predict the score of cognitive function defect scale (PDQ-D).Experimental results show that the method can realize the objective quantitative evaluation of depression cognitive function.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and auxiliary assessment technology for mental illness, specifically to a method and system for assessing cognitive function in depression based on macroscopic brain dynamics characteristics. Background Technology

[0002] Depression is a common and serious mental disorder, mainly characterized by symptoms such as depressed mood, loss of interest, and cognitive decline. It severely impacts patients' quality of life and places a heavy burden on society and the healthcare system. Therefore, establishing an objective and reliable assessment method for depression is of great significance.

[0003] Currently, clinical assessment of depression primarily relies on rating scales, which are subjective to some extent and lack stable objective biological indicators. Cognitive deficits (including attention, memory, and executive function impairments) are a crucial core dimension of depression and are closely related to patients' functional recovery and disease prognosis. These are typically assessed using patient self-assessment scales (such as the PDQ-D), but such methods are easily influenced by individual subjective factors, resulting in limited stability and consistency. Therefore, it is necessary to introduce objective neuroimaging indicators to quantitatively characterize cognitive deficits.

[0004] In recent years, functional magnetic resonance imaging (fMRI) has been widely used in mental illness research. Existing techniques have attempted to utilize fMRI time series to extract dynamic features to assess brain function in depression. For example, Wu et al. proposed a depression analysis method based on dynamic region homogeneity (dReHo), extracting dynamic region homogeneity from resting-state fMRI data as a single dynamic indicator and combining it with support vector machines to construct a classification model to distinguish between depressed patients and healthy controls. This method relies on a single source of dynamic features, focusing only on local consistency changes in the time series of local brain regions, failing to systematically characterize the dynamic characteristics of BOLD signals from multiple dimensions such as statistics, frequency domain, and nonlinearity. Another example is An et al., who used co-activation mode analysis to characterize the spatiotemporal dynamics of the brain network in depression, extracting a few predefined dynamic attributes such as dwell time, incidence, transition probability, and Markov trajectory entropy, and combining them with support vector machines to achieve patient classification diagnosis. Although this method introduces multiple dynamic indicators, the indicators used are still a limited number of features pre-selected by the researchers, limited by selection biases of prior knowledge, making it difficult to comprehensively capture the complex and ever-changing dynamic patterns of brain signals over time.

[0005] The common shortcomings of the aforementioned existing technologies are as follows: First, the dynamic features used are all based on a small number of predefined indicators, with limited feature dimensions, making it difficult to comprehensively characterize the complex dynamic changes of the BOLD signal; second, due to the limited number of features and reliance on prior assumptions, their ability to represent the overall dynamic structure of brain function is insufficient; and third, the established assessment models mostly focus on auxiliary diagnostic classification of depression, lacking the ability to quantitatively assess cognitive function levels. Therefore, existing technologies still struggle to achieve an objective and quantitative assessment of an individual's cognitive function status in depression.

[0006] To address the aforementioned technical problems, this invention provides a method for assessing cognitive function in depression based on macroscopic brain dynamics characteristics. By systematically extracting multidimensional dynamic features of brain region time series, constructing low-dimensional macroscopic brain dynamics indicators, and establishing a quantitative regression model with cognitive function scale scores, an objective assessment of an individual's cognitive function in depression can be achieved. Summary of the Invention

[0007] The purpose of this invention is to provide a method for assessing cognitive function in depression based on macroscopic brain dynamics characteristics, in order to solve the problem that the assessment of cognitive function deficits in the prior art mainly relies on clinical scales and lacks objective neuroimaging biomarkers.

[0008] This invention provides a method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics, comprising the following steps: Step 1: Acquire resting-state functional magnetic resonance imaging data and preprocess them to obtain the average time series of 100 brain regions; Step 2: Extract multiple dynamic features from the time series of each brain region and construct a brain region dynamic feature matrix; Step 3: Standardize the dynamic features and select those that are effective in all subjects; Step 4: The selected features are centered and principal component analysis is used to extract the first principal component and the second principal component as macroscopic brain dynamics indicators. The projection values ​​of each brain region on the first principal component and the second principal component are calculated. Step 5: Using the average first principal component spatial pattern and the average second principal component spatial pattern of healthy control subjects as reference templates, Procrustes analysis was used to align the kinetic indices of each subject. Step 6: Based on the aligned dynamic indicators, brain regions with significant differences are selected through inter-group statistical comparison to construct brain region hierarchical features, and networks with significant differences are selected through functional network segmentation to construct network hierarchical features. The brain region hierarchical features and network hierarchical features are then concatenated to form a multi-scale brain dynamic feature vector. Step 7: Using the multi-scale brain dynamics feature vector as the input variable and the cognitive function scale score as the output variable, a regression model is used to model and assess the cognitive function of people with depression.

[0009] Further, in step 1, the preprocessing includes: converting DICOM format data to NIfTI format, removing the first 5 time points of each subject's functional data, retaining 250 time points, performing time-level correction and head movement correction, and removing subjects with head movements exceeding 3 mm or 3 degrees, using an EPI template-based standardization method to standardize the data to the MNI space and resample to a 3×3×3 cubic millimeter voxel resolution, removing linear drift in the time series, using the CompCor method to extract principal components from white matter and cerebrospinal fluid, and performing joint regression with head movement parameters and global signals to remove noise, performing 0.01 to 0.1 Hz frequency band filtering, and extracting the average time series of 100 brain regions based on the Schaefer100 template.

[0010] Furthermore, in steps 2 and 3, the extraction of multiple dynamic features uses a high-time-series analysis toolbox to calculate 7770 dynamic features, which include statistical features, time-dependent features, frequency domain features, and nonlinear features; the screening of features that are effective in all subjects includes using the Sigmoid function to perform nonlinear mapping on the features, eliminating features with invalid values ​​or infinite values, and performing consistency screening in 285 subjects to retain 6452 features.

[0011] Further, the feature is that, in step 6, the inter-group statistical comparison includes: performing a permutation test on 159 patients with depression and 126 healthy controls, with 5000 permutations, and setting a significance threshold of 0.01 with false positive correction, to screen out 5 significantly different brain regions corresponding to the first principal component and 2 significantly different brain regions corresponding to the second principal component; the functional network is divided into a Yeo seven-functional network, which includes a default mode network, a somatic motor network, a visual network, a dorsal attention network, a ventral attention network, a limbic system network, and a frontoparietal control network. Within each functional network, the dynamic indicators of the brain regions are averaged, and a permutation test combined with false discovery rate correction is used, with the false discovery rate threshold set to 0.05, to screen out the default mode network and visual network corresponding to the first principal component and the frontoparietal control network corresponding to the second principal component.

[0012] Furthermore, in step 7, the regression model is a ridge regression model, the cognitive function scale is a PDQ-D scale, and the modeling process uses cross-validation to perform covariate regression on the training set. The covariates include age, gender, and years of education, with the scale score residuals as output variables. In each fold cross-validation process, the regression coefficients corresponding to each input feature in the ridge regression model are extracted, the absolute values ​​of the regression coefficients are taken, and the absolute values ​​of the coefficients in all cross-validation folds are averaged to obtain the average contribution of each feature. All features are then ranked based on the average contribution.

[0013] This invention constructs a macroscopic brain time dynamics index system, starting from the overall dynamic structure of time series, to characterize the high-dimensional dynamic pattern of brain functional activities, and combines multi-scale feature modeling to achieve objective prediction and quantitative assessment of cognitive function levels in depression.

[0014] Its technical solution includes the following steps: A: Functional magnetic resonance imaging (fMRI) data acquisition and preprocessing A1: Convert the raw DICOM format data exported from the magnetic resonance scan into NIfTI format data; A2: Remove the first n time points to eliminate the effects of magnetization instability; A3: Perform time-layer correction to eliminate time differences between different scan layers; A4: Perform head motion correction processing and obtain rigid body motion parameters at each time point; A5: Perform spatial standardization to register individual brain images to a standard brain spatial template; A6: Remove linear trends from time series data; A7: Perform covariate regression denoising, including head motion parameters, white matter signals, cerebrospinal fluid signals, and global signals; A8: Perform bandpass filtering; A9: Brain regions are divided based on brain region templates and the average BOLD time series is extracted:

[0015] Where N is the number of brain regions.

[0016] B: Multidimensional dynamic characteristics of brain regions in time series B1: Time series for each brain region Calculate the set of multidimensional dynamic characteristic functions:

[0017] Where M is the total number of features.

[0018] B2: Calculate multiple time series dynamic features for each brain region time series, including statistical features, autocorrelation features, frequency domain features, and nonlinear dynamic features. The High Time Series Analysis Toolkit (hctsa) is preferred for feature calculation.

[0019] B3: The statistical characteristics include indicators that describe the distribution characteristics of time series, such as mean, variance, skewness, and kurtosis; The autocorrelation feature is used to characterize the correlation structure of time series at different time lags; The frequency domain features are used to describe the energy distribution of the time series at different frequency components; The nonlinear dynamic features are used to characterize complex dynamic behaviors of time series, including entropy features and complexity features.

[0020] B4: Obtain a set of dynamic feature vectors for each brain region, and combine the features of all brain regions to form a brain region dynamic feature matrix:

[0021] C: Dynamic characteristic standardization and reliable characteristic screening C1: Standardize the brain region dynamic feature matrix obtained in step B to eliminate dimensional differences between different features and improve the stability of feature distribution.

[0022] In a preferred embodiment, for each feature Normalization is performed using a nonlinear mapping method based on the Sigmoid function:

[0023] in: Indicates the first The first brain region One feature; These are parameters that are adaptively determined based on the feature distribution.

[0024] C2: To address the potential issues of missing features or computational failures among different participants, a validity label is established for each feature.

[0025] C3: Perform consistency screening across all participants, retaining only features that are effective across all participants, and construct a unified feature space:

[0026] in This represents the number of features after filtering.

[0027] D: Construction of Macroscopic Brain Dynamics Indicators D1: Center the standardized brain region dynamic feature matrix obtained in step C.

[0028] D2: Dimensionality reduction analysis is performed on high-dimensional dynamic features, with principal component analysis being the preferred method to construct a macroscopic dynamic representation of brain regions.

[0029] Where Z represents the low-dimensional dynamics; W is the projection matrix.

[0030] D3: Calculate the variance explained by each component. Select the top few components with the highest variance explained as macroscopic neurodynamic indicators.

[0031] D4: Obtain the projection values ​​of each brain region onto the macroscopic neurodynamic index.

[0032] E: Alignment of brain dynamics patterns across participants E1: The average dynamic pattern of the healthy control group is selected as the reference template.

[0033] E2: Calculate the spatial mapping relationship between each subject's macroscopic neurodynamic index and the target neurodynamic pattern. Procrustes analysis is preferably used to solve for the orthogonal transformation matrix R, minimizing the difference between the individual's macroscopic neurodynamic representation and the reference template, and satisfying the following conditions:

[0034] Where R represents the orthogonal transformation matrix; I represents the identity matrix.

[0035] E3: The aforementioned mapping is used to align the dynamic representations of each subject, ensuring that the dynamic patterns of different subjects have a consistent orientation in the same feature space.

[0036] in, This represents the aligned macroscopic brain dynamics representation matrix.

[0037] E4: By aligning the principal components of different subjects, the orientations are kept consistent, thus eliminating sign uncertainty and rotational differences in dimensionality reduction methods. This yields comparable macroscopic neurodynamic indices across subjects.

[0038] F: Construction of Multiscale Neurodynamic Features F1: Aligned macroscopic neurodynamic indicators were compared between groups at the brain region level to screen for brain regions with significant differences.

[0039] F2: Brain region features are integrated based on the functional networks to which the brain regions belong, constructing network-level features and calculating the average dynamic indices of each functional network to reflect the system-level brain functional state. Statistical tests are used to screen network features with significant differences.

[0040] F3: Integrating brain region-level and network-level features to construct multi-scale dynamic feature vectors:

[0041] in These are the significant brain region features selected in step F1. These are the significant network features selected in step F2.

[0042] G: A cognitive function assessment model for depression based on multi-scale neurodynamic characteristics G1: Pair each subject's multi-scale neurodynamic feature vector with its corresponding depression cognitive function scale score to construct a prediction dataset.

[0043] G2: The model is trained and tested using a K-fold cross-validation strategy. In each iteration, the data is divided into training and testing sets, and predictions are performed iteratively for all folds.

[0044] G3: In each round of cross-validation, a covariate regression model is established based on the training set. The covariates include age, gender and years of education. Specifically, it includes: (1) using the covariates in the training set to perform linear regression on the PDQ-D score and extracting the residuals as the target variable; (2) performing covariate regression on each feature variable in the training set and extracting the residual features.

[0045] Subsequently, the covariate regression model obtained on the training set is applied to the test set, and the test set data is processed in the same way to avoid information leakage from the test data.

[0046] G4: After feature processing, the features of the training set are standardized, and the resulting parameters are applied to the test set. A ridge regression model is then used to build a prediction model and predict the test samples.

[0047] G5: After completing cross-validation predictions for all participants, the predictive ability of the model was evaluated by calculating the Pearson correlation coefficient between the predicted values ​​and the residuals of the actual PDQ-D scores. To further verify the reliability of the model results, a permutation test was used. By randomly shuffling the clinical score labels and repeating the model, a null distribution was constructed to assess the statistical significance of the actual prediction results.

[0048] G6: To improve the interpretability of the model, this invention quantitatively evaluates the contribution of neurodynamic features to cognitive function prediction. Specifically, in each fold of cross-validation, the regression coefficients of each feature in the ridge regression model are recorded, and their absolute values ​​are taken, ignoring their positive or negative directions. Subsequently, the absolute values ​​of the feature coefficients in all cross-validation folds are averaged to obtain the average contribution of each feature in the overall model. Based on this average contribution, all features are ranked, thereby identifying the key brain regions and functional network dynamic features that play a major role in cognitive function prediction.

[0049] The present invention provides a method and system for assessing cognitive function in depression based on macroscopic brain dynamics characteristics, which has the following beneficial effects and significant progress.

[0050] Compared with existing technologies, this invention systematically represents brain functional activity from the perspective of the time-series dynamics of brain signals. By integrating multidimensional time-series features through dimensionality reduction, it constructs macroscopic brain dynamics indicators that reflect the overall dynamic organizational pattern of brain function, effectively reducing redundancy of high-dimensional features while improving the model's stability and generalization ability. Simultaneously, this invention combines multi-scale brain dynamics features at the brain region and functional network levels to comprehensively characterize the spatial organizational patterns of brain functional abnormalities, thereby improving the representational ability and assessment consistency of cognitive function in depression. This invention provides a new technical solution for the objective neuroimaging representation of cognitive function levels and offers a new technical means for studying the dynamic mechanisms of brain functional abnormalities related to mental illness.

[0051] First, this invention achieves an objective and quantitative assessment of cognitive function in patients with depression. Traditional assessments rely on scales such as the PDQ-D, which are subject to subjective bias and recall bias. This invention, based on resting-state functional magnetic resonance imaging (fMRI) data, extracts 7770 dynamic features from brain region time series data. Principal component analysis is used to construct macroscopic brain dynamics indicators, and finally, a regression model is used to quantitatively predict an individual's cognitive function level. In Example 1, the correlation coefficient between the predicted score and the actual score residuals reached 0.312, and the permutation test p-value was 0.017, demonstrating that this method can statistically effectively assess cognitive function in patients with depression.

[0052] Second, a multi-scale brain dynamics feature space was constructed. This invention simultaneously extracts 7-dimensional features at the brain region level and 3-dimensional features at the functional network level, fusing them into a 10-dimensional feature vector. This not only characterizes local brain region abnormalities but also reflects system-level network features. The significantly different brain regions selected in Example 1 include the left and right inferior occipital gyri and the right postcentral gyrus, while the significantly different networks include the default mode network, the visual network, and the frontoparietal control network. These results are consistent with the conclusions of neuroimaging studies on depression, validating the effectiveness of the features.

[0053] Third, it solves the problem of comparability of cross-subject brain dynamics patterns. Addressing the directional uncertainty inherent in principal component analysis, this invention employs Procrustes analysis, using the average pattern of the healthy control group as a reference template for alignment, thereby reducing the impact of directional ambiguity. In Example 3, without alignment, the predicted correlation coefficient decreased from 0.312 to 0.244 and was no longer statistically significant, directly demonstrating the necessity and technical contribution of the alignment step.

[0054] Fourth, it has good potential for clinical application. This invention is based on resting-state functional magnetic resonance imaging data that is routinely available in clinical settings. The preprocessing process is standardized, feature extraction is automated, and the assessment results can be quickly output after model training. At the same time, by performing regression processing on covariates such as age, gender, and years of education, the influence of confounding factors is reduced, and the individualized accuracy of the assessment results is improved.

[0055] Fifth, it enhances the interpretability of the model. This invention introduces a feature contribution analysis method into the prediction model to rank the roles of brain region-level and functional network-level features in cognitive function prediction, thereby identifying key brain regions and functional networks and providing a basis for the neurobiological interpretation of the results.

[0056] Compared with the prior art, the present invention has the following significant advantages: First, it expands from single features to a multi-dimensional dynamic feature system. Existing methods mostly use single indicators such as low-frequency oscillation amplitude and local consistency, which have limited information representation capabilities. This invention uses a high-time-series analysis toolbox to calculate 7770 features, and after screening, retains 6452 effective features, covering statistical, time-series dependent, frequency domain and nonlinear features, comprehensively characterizing the dynamic patterns of BOLD signals.

[0057] Secondly, it expands from a single spatial scale to a dual scale of brain regions and networks. Existing technologies typically focus only on the brain region level or the whole-brain network level, making it difficult to take into account both local and overall information. This invention simultaneously constructs 7-dimensional features at the brain region level and 3-dimensional features at the network level, and the multi-scale fusion significantly improves predictive performance.

[0058] Third, a pattern alignment mechanism is introduced to reduce the impact of principal component orientation uncertainty. To address the potential orientation ambiguity in cross-subject comparisons during principal component analysis, this invention introduces the Procrustes alignment method to align macroscopic neurodynamic indicators, thereby improving comparability between different individuals. Example 3 results show that this step has a crucial impact on model performance.

[0059] Fourth, the predictive performance reaches a good application level. In Example 1, the correlation coefficient is 0.312 and the p-value is 0.017. In Example 2, after feature selection and optimization, the correlation coefficient is improved to 0.323 and the p-value is 0.009. Overall, it is better than the common level of most cognitive function prediction methods based on resting-state functional magnetic resonance imaging, showing good application potential.

[0060] This invention has achieved significant technological advancements in feature extraction comprehensiveness, multi-scale fusion strategy, cross-subject alignment mechanism, and predictive performance, and has clear clinical translational value. Attached Figure Description

[0061] Figure 1This is a flowchart of a cognitive function assessment method for depression based on macroscopic brain dynamics characteristics provided by an embodiment of the present invention.

[0062] Figure 2 This is a structural block diagram of a cognitive function assessment system for depression based on macroscopic brain dynamics characteristics provided in an embodiment of the present invention.

[0063] Figure 3 This is an overall flowchart provided by an embodiment of the present invention.

[0064] Figure 4 This is a schematic diagram illustrating the results of predicting cognitive function in depression using the method provided in this embodiment of the invention.

[0065] Figure 5 This is a schematic diagram showing the ranking of feature contributions of the evaluation model in an embodiment of the present invention. Detailed Implementation

[0066] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics, comprising the following steps: S101: Acquire resting-state functional magnetic resonance imaging data and preprocess them to obtain the average time series of 100 brain regions; S102, extract multiple dynamic features from the time series of each brain region and construct a brain region dynamic feature matrix; S103, Standardize the dynamic features and screen out features that are effective in all subjects; S104, The selected features are centered and principal component analysis is used to extract the first principal component and the second principal component as macroscopic brain dynamics indicators, and the projection value of each brain region on the first principal component and the second principal component is calculated. S105, using the average first principal component spatial pattern and the average second principal component spatial pattern of healthy control subjects as reference templates, and employing Procrustes analysis to align the kinetic indices of each subject; S106, Based on the aligned dynamic indicators, brain regions with significant differences are selected through inter-group statistical comparison to construct brain region hierarchical features, and network hierarchical features are selected through functional network segmentation to construct network hierarchical features. The brain region hierarchical features and network hierarchical features are then concatenated into a multi-scale brain dynamic feature vector. S107. Using the multi-scale neurodynamic feature vector as the input variable and the cognitive function scale score as the output variable, a regression model is used to model and assess the cognitive function of people with depression.

[0067] The preprocessing provided in this embodiment includes: converting DICOM format data to NIfTI format, removing the first 5 time points of each subject's functional data and retaining 250 time points, performing time-level correction, head movement correction and removing subjects with head movements exceeding 3 mm or 3 degrees, using an EPI template-based standardization method to standardize the data to the MNI space and resample to 3×3×3 cubic millimeter voxel resolution, removing linear drift in the time series, using the CompCor method to extract principal components from white matter and cerebrospinal fluid and perform joint regression with head movement parameters and global signals to remove noise, performing 0.01 to 0.1 Hz frequency band filtering, and extracting the average time series of 100 brain regions based on the Schaefer100 template.

[0068] The present invention provides the extraction of multiple dynamic features using a high-time-series analysis toolbox to calculate 7770 dynamic features, including statistical features, time-dependent features, frequency domain features, and nonlinear features. The specific steps for selecting features that are effective in all subjects are as follows: using the Sigmoid function to perform nonlinear mapping on the features, eliminating features with invalid values ​​or infinite values, and performing consistency screening on all 285 subjects to retain 6452 features.

[0069] The intergroup statistical comparison provided in this embodiment of the invention includes: performing a permutation test on 159 patients with depression and 126 healthy controls, with 5000 permutations, and setting a significance threshold of 0.01 with false positive correction, to screen out 5 significantly different brain regions corresponding to the first principal component and 2 significantly different brain regions corresponding to the second principal component; the functional network is divided into a Yeo seven-functional network, including a default mode network, a somatic motor network, a visual network, a dorsal attention network, a ventral attention network, a limbic system network, and a frontoparietal control network. Within each functional network, the dynamic indicators of the corresponding brain regions are averaged, and a permutation test combined with false discovery rate correction is used, with a false discovery rate threshold set to 0.05, to screen out the default mode network and visual network corresponding to the first principal component and the frontoparietal control network corresponding to the second principal component.

[0070] The regression model provided in this embodiment of the invention is a ridge regression model; the cognitive function scale is the PDQ-D scale; the modeling process uses cross-validation to perform covariate regression on the training set, and the covariates include age, gender and years of education, with the scale score residuals as output variables.

[0071] like Figure 2 As shown in the figure, an embodiment of the present invention provides a cognitive function assessment system for depression based on macroscopic brain dynamics characteristics, comprising: The data acquisition and preprocessing module is used to acquire resting-state functional magnetic resonance imaging data and perform preprocessing to obtain the average time series of 100 brain regions; The dynamic feature extraction module is used to extract multiple dynamic features from the time series of each brain region and construct a brain region dynamic feature matrix. The feature selection module is used to standardize the dynamic features and select features that are effective in all subjects; The macro-indicator construction module is used to centralize the screened features and use principal component analysis to extract the first principal component and the second principal component as macro-brain dynamics indicators, and calculate the projection value of each brain region on the first principal component and the second principal component. The alignment module is used to align the kinetic parameters of each subject using Procrustes analysis, with the mean first principal component space pattern and the mean second principal component space pattern of healthy control subjects as reference templates. The multi-scale feature construction module is used to select significantly different brain regions based on aligned dynamic indicators through inter-group statistical comparison to construct brain region hierarchical features, and to select significantly different networks through functional network segmentation to construct network hierarchical features. The brain region hierarchical features and network hierarchical features are then concatenated to form a multi-scale brain dynamic feature vector. The assessment module is used to model the cognitive function of people with depression by using the multi-scale neurodynamic feature vector as input variable and the cognitive function scale score as output variable, and by using a regression model.

[0072] The data acquisition and preprocessing module provided in this embodiment of the invention performs the following operations: converting DICOM format data to NIfTI format, removing the first 5 time points of each subject's functional data and retaining 250 time points, performing time-level correction, head movement correction and removing subjects with head movements exceeding 3 mm or 3 degrees, using an EPI template-based standardization method to standardize the data to the MNI space and resample to 3×3×3 cubic millimeter voxel resolution, removing linear drift in the time series, using the CompCor method to extract principal components from white matter and cerebrospinal fluid and perform joint regression with head movement parameters and global signals to remove noise, performing 0.01 to 0.1 Hz frequency band filtering, and extracting the average time series of 100 brain regions based on the Schaefer100 template.

[0073] The dynamic feature extraction module provided in this embodiment of the invention uses a high-time-series analysis toolbox to calculate 7770 dynamic features; the feature screening module uses the Sigmoid function for nonlinear mapping to remove features with invalid or infinite values, and performs consistency screening on all 285 subjects, retaining 6452 features.

[0074] In the multi-scale feature construction module provided in this embodiment of the invention, the inter-group statistical comparison includes performing a permutation test on 159 patients with depression and 126 healthy controls, with 5000 permutations. Combined with false positive correction, the significance threshold is set to 0.01. The left inferior orbitofrontal gyrus, right superior dorsolateral frontal gyrus, left inferior occipital gyrus, right inferior occipital gyrus, and right postcentral gyrus corresponding to the first principal component, and the right angular gyrus and right inferior temporal gyrus corresponding to the second principal component are selected. The functional network is divided into a Yeo seven-functional network, and the default mode network and visual network corresponding to the first principal component and the frontoparietal control network corresponding to the second principal component are selected.

[0075] The assessment module provided in this embodiment of the invention uses a ridge regression model and a PDQ-D cognitive function scale. The training set is processed by cross-validation, and the covariates include age, gender, and years of education. The scale score residuals are used as output variables.

[0076] This embodiment provides a method for assessing cognitive function in depression based on macroscopic brain dynamics characteristics. This method systematically processes resting-state functional magnetic resonance imaging (fMRI) data to extract macroscopic dynamic characteristics at the brain region and functional network levels, and constructs a predictive model to achieve a quantitative assessment of an individual's cognitive function in depression.

[0077] Figure 3 This is an overall flowchart of the method of the present invention, which mainly includes brain region time series extraction, high-dimensional dynamic feature calculation, macroscopic brain dynamic index construction, multi-scale feature fusion, and cognitive function assessment and model evaluation based on the features.

[0078] Figure 4 This diagram illustrates the results of predicting cognitive function in patients with depression using the method of this invention. The horizontal axis represents the actual cognitive function score of the subjects (Observed), and the vertical axis represents the model's predicted score (Predicted). Each scatter point in the diagram represents a subject, the solid line represents the linear regression fit between the predicted and actual values, and the dashed line represents the regression confidence interval range.

[0079] Figure 5 This is a schematic diagram showing the ranking of feature contributions of the evaluation model in this embodiment of the invention. The horizontal axis represents the standardized feature contribution, and the vertical axis represents the names of features in each brain region and network. The data used in this embodiment comes from the Fourth People's Hospital of Chengdu, and a total of 285 subjects were included, including 126 healthy controls and 159 patients with major depressive disorder.

[0080] The specific steps are as follows: A: Functional magnetic resonance imaging (fMRI) data acquisition and preprocessing The functional magnetic resonance data preprocessing procedure in this embodiment is implemented based on the DPABI toolbox. A1: Convert the raw DICOM format data exported from the magnetic resonance scanner to NIfTI format; A2: The first 5 time points of functional data for each subject were removed to eliminate the effects of magnetic field instability and the subject's adaptation process, and 250 time points were retained for subsequent analysis; A3: Time-level correction; A4: Head movement correction. In this embodiment, head movements exceeding 3mm or... are excluded. The participants were ultimately all 285 subjects included in the subsequent analysis; A5: A standardization method based on EPI templates is used to standardize the data to the MNI space and uniformly resample it. voxel resolution; A6: Remove linear drift from time series; A7: The CompCor method was used to extract principal components from white matter and cerebrospinal fluid, and then combined with head movement parameters and global signals for regression to remove physiological and motion noise. A8: Performs 0.01–0.1Hz frequency band filtering to retain low-frequency neural activity signals; A9: The average time series of brain regions was extracted based on the Schaefer100 template, resulting in BOLD signals for 100 brain regions at 250 time points each.

[0081] B: Multidimensional dynamic characteristics of brain regions in time series This step aims to extract a rich variety of dynamic features from the time series of each brain region to comprehensively characterize the dynamic properties of the BOLD signal. For each brain region time series of each subject, this embodiment used the High Time Series Analysis Toolbox (hctsa) to calculate 7770 dynamic features, including statistical features, time-dependent features, frequency domain features, and nonlinear features.

[0082] In this embodiment, each subject ultimately forms a 100×7770 brain region dynamic feature matrix, which is used to characterize the complex dynamic patterns of brain signals in the time dimension.

[0083] C: Standardization and Screening of Dynamic Characteristics Because different features have significantly different dimensions and numerical ranges, directly performing dimensionality reduction analysis may lead to the results being dominated by features with larger dimensions. This step standardizes the features and selects those that are stable and reliable across all participants.

[0084] C1: The Sigmoid function is used to perform non-linear mapping on the features, so that different feature distributions are mapped to a stable interval; C2: Establish a validity label for each feature, identify and remove features that have invalid values ​​(such as NaN) or infinite values ​​(Inf).

[0085] C3: Consistency screening was performed across all 285 participants, retaining only features that were valid across all participants. In this embodiment, 6452 features that were valid across all participants were ultimately retained.

[0086] This step constructs a unified and stable feature space, providing reliable input for subsequent analysis.

[0087] D: Construction of Macroscopic Brain Dynamics Indicators This step uses principal component analysis to extract low-dimensional macroscopic indicators that can summarize the main dynamic change patterns of the brain.

[0088] D1-D2: The screened features are centered and principal component analysis is used to extract the main dynamic modes; D3: The results in this embodiment indicate that: (1) Healthy control group: PC1 explained 29.73%, PC2 explained 17.97%; (2) Depression group: PC1 explained 29.91%, PC2 explained 18.34%.

[0089] The cumulative variance explained by the first two principal components in both groups was close to 50%, and the explained rates were highly consistent, indicating that PC1 and PC2 can stably represent the core change patterns of the macroscopic intrinsic dynamics of the brain. Therefore, PC1 and PC2 were selected as macroscopic brain dynamics indicators.

[0090] D4: Calculate the projection values ​​of 100 brain regions for each subject on PC1 and PC2 to obtain the hierarchical dynamic representation of brain regions.

[0091] E: Alignment of brain dynamics patterns across participants Principal component analysis (PC1) introduces orientation uncertainty, as the PC1 / PC2 orientations of different participants may be opposite or exhibit rotational differences. This step eliminates this uncertainty through pattern alignment.

[0092] E1: In this embodiment, the average PC1 and PC2 spatial patterns of 126 healthy control subjects are used as a reference template.

[0093] E2: For each subject, their brain region dynamics representation is matched with a reference pattern, and Procrustes analysis is used to calculate the optimal rotation transformation that minimizes the difference between the transformed individual dynamics pattern and the reference pattern.

[0094] E3: After obtaining this transformation relationship, the original kinetic indices of each subject are aligned so that they are expressed in the same reference space.

[0095] This step eliminates the problem of inconsistent principal component orientations and improves the accuracy of inter-group comparisons.

[0096] F: Construction of Multiscale Neurodynamic Features To characterize dynamic anomalies at different spatial scales, brain region-level and network-level features were constructed.

[0097] F1: Brain Region Hierarchy Analysis In this embodiment, a statistical comparison was performed between 159 patients with depression and 126 healthy controls based on aligned brain region dynamics indices (PC1 and PC2).

[0098] The significance threshold was set to p<0.01 (i.e., 1 / 100) using a permutation test (5000 permutations) and false positive correction. Screening results showed that for PC1, patients with depression had significantly decreased levels in both inferior occipital gyri and the right postcentral gyrus, and significantly increased levels in the left inferior orbitofrontal gyrus and the right superior dorsolateral frontal gyrus; for PC2, patients with depression had significantly decreased levels in the right angular gyrus and significantly increased levels in the right inferior temporal gyrus.

[0099] The dynamic indicators corresponding to the above brain regions are extracted and combined to construct brain region hierarchical feature vectors, which are used to characterize local abnormal brain function dynamics patterns.

[0100] F2: Network Hierarchy Analysis To further characterize the brain's functional organization at the system level, this embodiment groups brain regions based on the seven functional networks proposed by Yeo et al. (Yeo7). Specifically, these seven functional networks include: Default Mode Network (DMN), Somatic Motor Network (SMN), Visual Network (VIS), Dorsal Attention Network (DAN), Ventral Attention Network (VAN), Limbic System Network (LIM), and Frontoparietal Control Network (FPN).

[0101] In this embodiment, each brain region in the Schaefer100 brain region template is first mapped to the aforementioned functional network. Then, for each subject, within each functional network, the dynamic indices (PC1 or PC2) of all brain regions included in that network are averaged to obtain the overall dynamic level of the network. After obtaining each network index, a permutation test (5000 permutation) combined with false discovery rate correction (FDR, q<0.05) is further performed to screen for network features that show significant differences between patients with depression and healthy controls.

[0102] The results showed that for PC1, the default mode network increased while the visual network decreased; for PC2, the frontoparietal control network increased.

[0103] The above three significant network dynamics indicators were selected as network hierarchical features.

[0104] F3: Multi-scale feature fusion By concatenating brain region-level features with network-level features, a multi-scale brain dynamics feature vector is constructed.

[0105] In this embodiment, the brain region-level features are 7-dimensional, and the network-level features are 3-dimensional, ultimately forming a 10-dimensional feature vector. This multi-scale feature simultaneously includes local brain region information and overall network information, enabling a more comprehensive reflection of brain function abnormalities associated with depression.

[0106] G: Cognitive Function Assessment Model for Depression G1: In this embodiment, 113 out of 159 patients with depression completed the PDQ-D scale assessment. For each patient, the multi-scale neurodynamic feature vector constructed in step F was extracted and paired with the corresponding PDQ-D score to construct a supervised learning dataset.

[0107] G2: This embodiment uses a 5-fold cross-validation method for model evaluation. Specifically, all samples are divided into 5 subsets. Each time, one subset is selected as the test set, and the remaining 4 subsets are used as the training set. This process is repeated until all subsets have been used as the test set once.

[0108] G3: In each round of cross-validation, the training set is first subjected to covariate regression processing. The covariates include age, gender, and years of education, specifically including: (1) Use the covariates in the training set to perform linear regression on the PDQ-D score and extract the residuals as the target variable; (2) Perform covariate regression on each feature variable separately and extract residual features; The trained regression model is then applied to the test set, and the test set data is processed in the same way to avoid leakage of test information.

[0109] G4: In this embodiment, a ridge regression model is used for modeling (in other embodiments, linear regression, support vector regression, or other regression models may also be used). This model uses multi-scale neurodynamic features as input variables and PDQ-D score residuals as output variables. By fitting the relationship between features and scores on training samples, it can predict cognitive function levels.

[0110] G5: After obtaining the prediction results for all patients, the model performance is quantitatively evaluated.

[0111] First, the Pearson correlation coefficient between the predicted score and the actual PDQ-D score residuals was calculated. In this embodiment, the correlation coefficient r = 0.312, indicating that the constructed model can reflect the trend of cognitive function changes in depression to a certain extent. Second, to further verify the statistical significance of the prediction results, a permutation test method was used. Specifically, by randomly shuffling the PDQ-D score labels and repeating the complete modeling process, a null distribution was constructed. After 5000 permutations, this embodiment obtained a significance level of p = 0.017.

[0112] G6: This embodiment quantitatively evaluates the feature contribution during the prediction process. In each round of the 5-fold cross-validation, the standardized regression coefficients of the ridge regression model corresponding to the 10 input features were recorded, and their absolute values ​​were taken. The absolute values ​​of the coefficients obtained from all folds were averaged to obtain the average contribution of each feature in the overall model. The results show that among the 10 features, the right postcentral gyrus (PC1) dynamics index has the highest contribution, followed by the right inferior occipital gyrus (PC1) and the frontoparietal control network (PC2). The visual network (PC1), default mode network (PC1), and right dorsolateral superior frontal gyrus (PC1) also showed high predictive contributions. The above feature contribution ranking confirms the potential importance of the occipital visual cortex, the postcentral gyrus somatosensory area, and the frontoparietal control network in cognitive impairment in depression, thereby enhancing the clinical interpretability of the model.

[0113] The above results demonstrate that the method of the present invention can effectively predict the level of cognitive function in depression in a statistical sense, verifying the effectiveness of the constructed multi-scale brain dynamics features.

[0114] Example 1 This embodiment uses data from 285 subjects collected from the Fourth People's Hospital of Chengdu to perform the assessment method, including 126 healthy controls and 159 patients with major depressive disorder. In the data preprocessing stage, the original DICOM format data was converted to NIfTI format, and the first 5 time points of each subject's functional data were removed, retaining 250 time points for subsequent analysis. Temporal correction and head movement correction were performed sequentially, and subjects with head movements exceeding 3 mm or 3 degrees were excluded; all 285 subjects met the requirements. The data were normalized to the MNI space using an EPI template-based normalization method and resampled to a 3x3x3 cubic millimeter voxel resolution. After removing linear drift in the time series, principal components were extracted from white matter and cerebrospinal fluid using the CompCor method, and jointly regressed with head movement parameters and global signals to remove physiological and motion noise, followed by filtering in the 0.01 to 0.1 Hz frequency band. The average time series of 100 brain regions was extracted based on the Schaefer100 template, with each region containing 250 time points of BOLD signal. For each brain region time series, 7770 dynamic features were calculated using the High Time Series Analysis Toolkit, forming a 100x7770 brain region dynamic feature matrix. A sigmoid function was used to perform nonlinear mapping on the features, eliminating features with invalid or infinite values, retaining 6452 valid features from all 285 subjects. The selected features were centered and principal component analysis was performed to extract the first and second principal components, calculating the projection values ​​of each brain region onto the two principal components. Using the average first and second principal component spatial patterns of 126 healthy controls as reference templates, Procrustes analysis was used to align each subject. Intergroup comparisons employed a 5000-permutation test combined with false positive correction, with a significance threshold set at 0.01. Five brain regions corresponding to the first principal component (left inferior orbitofrontal gyrus, right superior dorsolateral frontal gyrus, left inferior occipital gyrus, right inferior occipital gyrus, and right postcentral gyrus) and two brain regions corresponding to the second principal component (right angular gyrus and right inferior temporal gyrus) were selected, and dynamic indices from these brain regions were extracted to form a 7-dimensional feature matrix. Based on the Yeo seven-function network partitioning, 100 brain regions were mapped to the default mode network, somatic motor network, visual network, dorsal attention network, ventral attention network, limbic system network, and frontoparietal control network. Within each network, the dynamic indices of the corresponding brain regions were averaged. After 5000 permutation tests combined with error detection rate correction (with an error detection rate threshold of 0.05), the default mode network and visual network corresponding to the first principal component, and the frontoparietal control network corresponding to the second principal component, were selected. Indicators from these three networks were extracted to form 3D features. The 7-dimensional brain region hierarchical features were concatenated with the 3-dimensional network hierarchical features to form a 10-dimensional multi-scale brain dynamic feature vector.In 113 patients with depression who completed the PDQ-D scale assessment, using 10-dimensional features as input and PDQ-D score residuals as output, a ridge regression model combined with 5-fold cross-validation was employed. The Pearson correlation coefficient between the predicted and actual score residuals was 0.312, and the significance level was 0.017 after 5000 permutations. In the above predictive modeling process, a feature contribution evaluation method based on the average of cross-validation coefficients was used to obtain the average contribution of each feature. The results showed that the features with higher contributions were, in descending order: right postcentral gyrus (PC1), right inferior occipital gyrus (PC1), frontoparietal control network (PC2), visual network (PC1), default mode network (PC1), and right dorsolateral superior frontal gyrus (PC1). This confirms that the method can effectively assess cognitive function in depression and has clinical interpretability.

[0115] Example 2 This embodiment further verifies the impact of feature selection stability on predictive performance based on Embodiment 1. Using the same data from 285 subjects, in the feature standardization and selection step C, the consistency selection threshold was adjusted from retaining features effective in all 285 subjects to retaining features effective in at least 90% of subjects. This adjustment increased the number of retained features from 6452 to 7103. Subsequent principal component analysis showed that the first principal component explained 29.68% and the second principal component explained 17.92% in the healthy control group; the first principal component explained 29.85% and the second principal component explained 18.28% in the depression group, highly consistent with the results of Embodiment 1. After Procrustes alignment, the significantly different brain regions and networks identified in the inter-group comparisons were identical to those in Embodiment 1, but the effect sizes of some brain regions changed. In the predictive model, using the same ridge regression and 5-fold cross-validation, the correlation coefficient between the predicted score and the actual PDQ-D score residuals increased to 0.323, and the permutation test significance level was 0.009. The results show that appropriately relaxing the feature validity threshold can increase feature diversity and further improve prediction performance, verifying the robustness and adjustability of the feature selection step in this method.

[0116] Example 3 This embodiment verifies the necessity of the Procrustes alignment step for cross-subject comparisons. In the entire processing flow of Example 1, the Procrustes alignment operation in step E is omitted, and the unaligned projection values ​​of the first and second principal components are directly used for inter-group comparisons and feature construction. The results show that, under the unaligned condition, inter-group statistical comparisons revealed a significant decrease in PC1 in the inferior occipital gyrus and right postcentral gyrus, a significant decrease in PC1 in the visual network, and a significant increase in PC2 in the frontoparietal control network; while the significant changes in the left inferior orbitofrontal gyrus, right superior dorsolateral frontal gyrus, right angular gyrus, and right inferior temporal gyrus in Example 1 became insignificant. In the prediction model, using the unaligned multi-scale features as input, the correlation coefficient between the predicted score and the actual PDQ-D score residuals decreased to 0.244, and the permutation test p-value was 0.077, which did not reach statistical significance. This comparative embodiment demonstrates that the Procrustes alignment step eliminates the uncertainty of principal component orientation and is a key technical mechanism to ensure the comparability of cross-subject dynamic patterns; the absence of this step will lead to the failure of the evaluation model.

[0117] Example 4 Extended Scheme This embodiment, based on existing single-modal resting-state functional magnetic resonance imaging (fMRI) data, introduces gray matter volume features from structural magnetic resonance imaging (SMRI) data as supplementary information. For the 113 patients with depression who completed the PDQ-D scale assessment in Example 1, their T1-weighted structural image data were used. The total gray matter volume of the whole brain was calculated using a voxel-based morphological analysis method, and the total intracranial volume of the whole brain was used as a covariate for correction. The whole-brain gray matter volume features were concatenated with the existing multi-scale neurodynamic features to construct an 11-dimensional feature vector. In the ridge regression model, using the same 5-fold cross-validation and covariate regression processing, the correlation coefficient between the predicted score and the actual PDQ-D score residuals increased to 0.326, and the permutation test p-value was 0.028. This multimodal extension scheme further improves the accuracy of cognitive function assessment and provides a new technical path for studying the mechanism of abnormal coupling between brain structure and function in depression.

[0118] Example 5 Extended Scheme This embodiment expands the original cross-sectional assessment into a longitudinal prediction model. It includes patients with depression who have completed baseline assessments, of whom 38 received two weeks of repetitive transcranial magnetic stimulation (rTMS) and completed the PDQ-D scale assessment again after treatment. Using 10-dimensional multi-scale macroscopic neurodynamic characteristics extracted from baseline as input variables and the change in PDQ-D score after two weeks of treatment (baseline score minus post-treatment score) as the output variable, a ridge regression model consistent with Example 1 was used for modeling. After 5-fold cross-validation and covariate regression, the Pearson correlation coefficient between the predicted and actual score changes was 0.281, and the permutation test p-value was 0.039. This prediction scheme indicates that baseline macroscopic neurodynamic characteristics are associated with changes in cognitive function after short-term treatment and can predict individual treatment responses to a certain extent, possessing significant clinical application value and serving as a powerful supplement to existing cross-sectional assessment methods.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics, characterized in that, Includes the following steps: Step 1: Acquire resting-state functional magnetic resonance imaging data and preprocess them to obtain the average time series of 100 brain regions; Step 2: Extract multiple dynamic features from the time series of each brain region and construct a brain region dynamic feature matrix; Step 3: Standardize the dynamic features and select those that are effective in all subjects; Step 4: The selected features are centered and principal component analysis is used to extract the first principal component and the second principal component as macroscopic brain dynamics indicators. The projection values ​​of each brain region on the first principal component and the second principal component are calculated. Step 5: Using the average first principal component spatial pattern and the average second principal component spatial pattern of healthy control subjects as reference templates, Procrustes analysis was used to align the kinetic indices of each subject. Step 6: Based on the aligned dynamic indicators, brain regions with significant differences are selected through inter-group statistical comparison to construct brain region hierarchical features, and networks with significant differences are selected through functional network segmentation to construct network hierarchical features. The brain region hierarchical features and network hierarchical features are then concatenated to form a multi-scale brain dynamic feature vector. Step 7: Using the multi-scale brain dynamics feature vector as the input variable and the cognitive function scale score as the output variable, a regression model is used to model and assess the cognitive function of people with depression.

2. The method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics according to claim 1, characterized in that, In step 1, the preprocessing includes: converting DICOM format data to NIfTI format, removing the first 5 time points of each subject's functional data, retaining 250 time points, performing time-level correction and head movement correction, and removing subjects with head movements exceeding 3 mm or 3 degrees. The data is standardized to MNI space using an EPI template-based standardization method and resampled to 3 x 3 x 3 cubic millimeter voxel resolution to remove linear drift in the time series. Principal components are extracted from white matter and cerebrospinal fluid using the CompCor method and combined with head movement parameters and global signals for joint regression to remove noise. Filtering is performed in the 0.01 to 0.1 Hz frequency band. The average time series of 100 brain regions is extracted based on the Schaefer100 template.

3. The method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics according to claim 1, characterized in that, In steps 2 and 3, the extraction of multiple dynamic features involves calculating 7770 dynamic features using the High Time Series Analysis Toolbox. These dynamic features include statistical features, time-dependent features, frequency domain features, and nonlinear features. The process of selecting features that are effective across all subjects includes using the Sigmoid function to perform non-linear mapping on the features, eliminating features with invalid or infinite values, and performing consistency screening on 285 subjects to retain 6452 features.

4. The method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics according to claim 1, characterized in that, In step 6, the intergroup statistical comparison includes: performing a permutation test on 159 patients with depression and 126 healthy controls, with 5000 permutations. A significance threshold of 0.01 was set using false positive correction to identify 5 significantly different brain regions corresponding to the first principal component and 2 significantly different brain regions corresponding to the second principal component. The functional network is divided into a Yeo seven-functional network, which includes a default mode network, a somatic motor network, a visual network, a dorsal attention network, a ventral attention network, a limbic system network, and a frontoparietal control network. Within each functional network, the dynamic indices of the corresponding brain regions are averaged. A permutation test combined with false discovery rate correction is used, with a false discovery rate threshold of 0.05, to identify the default mode network and visual network corresponding to the first principal component, and the frontoparietal control network corresponding to the second principal component.

5. The method for assessing cognitive function in depression based on macroscopic neurodynamic characteristics according to claim 1, characterized in that, In step 7, the regression model is a ridge regression model, the cognitive function scale is a PDQ-D scale, and the modeling process uses cross-validation to perform covariate regression on the training set. The covariates include age, gender, and years of education, and the scale score residuals are used as output variables. Furthermore, in each fold cross-validation process, the regression coefficients corresponding to each input feature in the ridge regression model are extracted, the absolute values ​​of the regression coefficients are taken, and the absolute values ​​of the coefficients in all cross-validation folds are averaged to obtain the average contribution of each feature. All features are then ranked based on the average contribution.

6. A cognitive function assessment system for depression based on macroscopic neurodynamic characteristics, characterized in that, The system assesses cognitive function in patients with depression using the method for assessing cognitive function in patients with depression based on macroscopic neurodynamic characteristics as described in any one of claims 1-5; the system comprises: The data acquisition and preprocessing module is used to acquire resting-state functional magnetic resonance imaging data and perform preprocessing to obtain the average time series of 100 brain regions; The dynamic feature extraction module is used to extract multiple dynamic features from the time series of each brain region and construct a brain region dynamic feature matrix. The feature selection module is used to standardize the dynamic features and select features that are effective in all subjects; The macro-indicator construction module is used to centralize the screened features and use principal component analysis to extract the first principal component and the second principal component as macro-brain dynamics indicators, and calculate the projection value of each brain region on the first principal component and the second principal component. The alignment module is used to align the kinetic parameters of each subject using Procrustes analysis, with the mean first principal component space pattern and the mean second principal component space pattern of healthy control subjects as reference templates. The multi-scale feature construction module is used to select significantly different brain regions based on aligned dynamic indicators through inter-group statistical comparison to construct brain region hierarchical features, and to select significantly different networks through functional network segmentation to construct network hierarchical features. The brain region hierarchical features and network hierarchical features are then concatenated to form a multi-scale brain dynamic feature vector. The assessment module is used to model the cognitive function of people with depression by using the multi-scale neurodynamic feature vector as input variable and the cognitive function scale score as output variable, and by using a regression model.

7. The system according to claim 6, characterized in that, The data acquisition and preprocessing module performs the following operations: converting DICOM format data to NIfTI format, removing the first 5 time points of each subject's functional data and retaining 250 time points, performing time-level correction, head movement correction and removing subjects with head movements exceeding 3 mm or 3 degrees, using an EPI template-based standardization method to standardize the data to MNI space and resample to 3×3×3 cubic millimeter voxel resolution, removing linear drift in the time series, using the CompCor method to extract principal components from white matter and cerebrospinal fluid and perform joint regression with head movement parameters and global signals to remove noise, performing 0.01 to 0.1 Hz frequency band filtering, and extracting the average time series of 100 brain regions based on the Schaefer100 template.

8. The system according to claim 6, characterized in that, The dynamic feature extraction module uses the high-time-series analysis toolbox to calculate 7770 dynamic features; the feature screening module uses the Sigmoid function for nonlinear mapping to remove features with invalid or infinite values, and performs consistency screening on all 285 subjects, retaining 6452 features.

9. The system according to claim 6, characterized in that, In the multi-scale feature construction module, the inter-group statistical comparison includes performing a permutation test on 159 patients with depression and 126 healthy controls, with 5000 permutations. Combined with false positive correction, the significance threshold is set to 0.

01. The left inferior orbitofrontal gyrus, right superior dorsolateral frontal gyrus, left inferior occipital gyrus, right inferior occipital gyrus, and right postcentral gyrus corresponding to the first principal component, and the right angular gyrus and right inferior temporal gyrus corresponding to the second principal component are selected. The functional network is divided into a Yeo seven-functional network, and the default mode network and visual network corresponding to the first principal component and the frontoparietal control network corresponding to the second principal component are selected.

10. The system according to claim 6, characterized in that, The assessment module uses a ridge regression model and a PDQ-D cognitive function scale. Cross-validation is used to perform covariate regression on the training set. The covariates include age, gender, and years of education. The scale score residuals are used as output variables.

Images