A Cognitive and Brain Imaging Data Integration and Evaluation Method for Alzheimer's Disease

Abstract

The present invention relates to the field of online assisted assessment, and particularly to a method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease. According to the method of the present invention, first, multi-domain screening and evaluation of cognitive functions are carried out to achieve the purpose of rapid screening and detailed evaluation of cognitive functions. Then, multi-modal magnetic resonance imaging data analysis is performed, including: extracting hippocampal volume and shape analysis, calculating the functional connectivity values of cognitive-related brain functional networks, and using them as indicators in the aspect of neuroimaging to evaluate brain functions. Next, Bayesian joint modeling of cognition and nerves is carried out, and finally, supervised learning classification is used to identify specific change patterns of Alzheimer's disease.

Description

Technical Field

[0001] The present invention relates to the field of online assisted evaluation, and particularly relates to a method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease. Background Art

[0002] Alzheimer's Disease (AD) is a neurological degenerative disease with insidious onset and irreversibility.

[0003] Mild Cognitive Impairment (MCI) is a prodromal state of Alzheimer's disease proposed by predecessors, and is an intermediate state between normal aging and dementia. Mild Cognitive Impairment can be used as a "predictor" of Alzheimer's disease. If this state can be detected early and appropriate intervention and treatment are given, the progression of Alzheimer's disease can be delayed. Therefore, the correct diagnosis of Alzheimer's disease, especially the correct diagnosis of its early stage of Mild Cognitive Impairment, is crucial for the prevention, early detection and treatment intervention of Alzheimer's disease.

[0004] Screening and evaluation in multiple domains of cognitive function can effectively identify patients with cognitive dysfunction. Common early signs of Alzheimer's disease may include the following: (1) Memory impairment: often manifested as "forgetting things", "forgetting what was just said", and repeatedly asking the same question. (2) Visuospatial skill disorder: unable to accurately judge the position of objects, unable to find one's own room or bed, and unable to distinguish the left and right, front and back of clothes. (3) Language disorder: although talkative, the listener cannot understand what he is saying, stuttering and vague. (4) Writing difficulty: the content written is incoherent, and even unable to write one's own name. (5) Apraxia and agnosia: originally able to ride a bike or swim proficiently, but unable to do so after the illness, and unable to recognize the faces of one's own relatives and familiar friends. (6) Calculation disorder: unable to calculate when shopping, and even unable to do the simplest addition and subtraction in severe cases. (7) Mental function disorder: often accompanied by mania, hallucinations, personality changes, etc. (8) Movement disorder: walking back and forth aimlessly, opening and closing doors everywhere, incontinence, etc.

[0005] Meanwhile, the development of neuroimaging techniques has also provided a basis and assistance for the early detection of AD / MCI patients from the perspective of changes in brain structure and function. Magnetic Resonance Imaging (MRI) has the highest soft tissue contrast resolution among all medical imaging methods, can clearly distinguish between gray and white matter in the brain, shows brain atrophy or ventricular enlargement more clearly and sensitively than CT, and can measure the volumes of structures such as the entire temporal lobe or hippocampus and amygdala, which is of great significance for the early diagnosis of AD. Functional Magnetic Resonance Imaging (fMRI) enables us to "see" how the brain is functioning and can isolate different brain functional networks: the Default Mode Network (DMN), the Salience Network (SN), the frontoparietal network (FPN), the Dorsal Attention Network (DAN), etc. These brain networks are highly correlated with memory, attention, executive function, etc. The patterns of functional connectivity changes within and between these networks can be used as imaging indicators that are earlier than brain structure changes and assist in the early diagnosis of AD. Summary of the Invention

[0006] The object of the present invention is to provide a method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease.

[0007] The method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease according to the present invention includes the following steps:

[0008] Step I: Obtain the behavioral data of the multi-domain screening and evaluation of the cognitive function of the subject;

[0009] Step II: Analyze the multi-modal magnetic resonance imaging data, including the following steps:

[0010] 1. Quantitatively analyze the volume and shape of the hippocampus of the subject based on the fixed-point analysis method,

[0011] 2. Calculate the functional connectivity within and between the cognitive-related brain functional networks

[0012] 2.1 Preprocess the resting-state functional magnetic resonance imaging data to correct the biases existing in the original resting-state functional magnetic resonance imaging data,

[0013] 2.2 Based on the preprocessed resting-state functional magnetic resonance imaging data, extract the cognitive-related brain functional networks and calculate the functional connectivity values within each brain functional network,

[0014] 2.3 Calculate the effective connectivity between networks

[0015] Based on the localized brain functional networks related to higher cognitive functions, the main nodes of each brain functional network are selected as regions of interest, and the effective connectivity between pairwise brain functional networks is calculated.

[0016] Step III: Based on the behavioral data from the multi-domain screening assessment of the cognitive functions of the subjects, the quantitative analysis data of the hippocampal volume and shape of the subjects, the within-network connectivity values of the brain functional networks related to cognition, and the effective connectivity between brain functional networks, the model distribution parameters of the behavioral and neural data are input into the same joint model. The hierarchical Bayesian method is used to fit the joint model to the relevant data, establish a hierarchical Bayesian joint model, obtain the central tendency and dispersion degree of the hyperparameter set, generate the joint posterior distribution of the parameters, and determine the degree and direction of the association of the cognitive and neural model parameters for mild cognitive impairment and Alzheimer's disease.

[0017] Step IV: According to the obtained joint posterior distribution of the parameters, a support vector machine (SVM) is used to train a specific change model for mild cognitive impairment and Alzheimer's disease.

[0018] According to the method for integrating and evaluating the cognition and brain imaging data of Alzheimer's disease of the present invention, the steps for preprocessing the resting-state functional magnetic resonance imaging data are as follows:

[0019] 1) Convert the image data file format into a format recognizable by the software for processing.

[0020] 2) Exclude the data of the first 10 time points of the image data to eliminate the error effects caused by the uneven magnetic field at the start of the scan and the discomfort of the subject on the results.

[0021] 3) Temporal correction processing: Temporal correction processing is performed on the remaining 230 data, and the different scan time points within the same TR are corrected to the same time point by a mathematical method for subsequent processing.

[0022] 4) Head correction: Each frame of the image in an experimental sequence is aligned with the first frame of this sequence according to a certain algorithm to correct the large error effects caused by the head movement of the subject.

[0023] 5) Spatial normalization: Map to the standard brain. The horizontal head movement and rotational head movement maps of the subject can be obtained through head correction. The images of the subjects with translation less than 1.5 mm and rotation less than 1.5° are mapped to the standard brain [-90, -126, -72; 90, 90, 108], and the voxel size is 3 * 3 * 3 mm.

[0024] 6) Smooth the image data with a smoothing kernel of [6 6 6] to improve the signal-to-noise ratio of the image signal after spatial normalization.

[0025] 7) Perform image detrending to remove the linear effects caused by machine temperature, subject adaptability, etc.

[0026] 8) Perform image filtering with a frequency band of 0.01 - 0.1 Hz to remove high-frequency signals.

[0027] 9) Extract covariates (head movement, whole brain, cerebrospinal fluid, white matter signal) and remove the covariates.

[0028] According to the method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease of the present invention, based on ICA, a brain functional network related to cognition is extracted.

[0029] According to the method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease of the present invention, the main nodes of each located brain functional network are selected as regions of interest, and the effective connectivity between pairwise brain functional networks is calculated.

[0030] According to the method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease of the present invention, a neural and behavioral data distribution is established, including behavioral data of multi-domain screening and evaluation of the cognitive function of the subject, quantitative analysis data of the hippocampal volume and shape of the subject, within-connection values within the brain functional network related to cognition, and effective connectivity data between brain functional networks. The distribution parameters of the two data models are input into the same joint model, and the hierarchical Bayesian method is used to fit the joint model to the relevant data to obtain the central tendency and dispersion degree of the hyperparameter set, and generate the joint posterior distribution of the parameters.

[0031] According to the method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease of the present invention, an SVM algorithm is used to determine a specific change model for mild cognitive impairment and Alzheimer's disease, specifically including: using 90% of the hyperparameter set data of the hierarchical Bayesian joint model of the obtained AD patients and healthy controls as the training set, and the remaining 10% as the test set. When the classification prediction accuracy rate on the test set is above 80%, it is considered that the SVM model classification prediction is reasonable.

[0032] According to the method for integrating and evaluating cognitive and brain imaging data of Alzheimer's disease of the present invention, the parameters of the specific change model for mild cognitive impairment and Alzheimer's disease are evaluated through cross-validation, specifically including the following steps:

[0033] The first step is to randomly divide the original data into 10 non-repeated samples.

[0034] The second step is to select 1 of them as the test set each time, and the remaining 9 as the training set for model training. After training on each training set, a model is obtained, and this model is used to test on the corresponding test set, and the evaluation index of the model: classification accuracy rate is calculated and saved.

[0035] In the third step, repeat the second step 10 times so that each portion of the original data has a chance to be used as the test set, and the remaining chances are used as the training set.

[0036] In the fourth step, calculate the average value of the 10 test results as an estimate of the model accuracy and use it as the performance indicator (true classification rate) of the current model.

[0037] According to the technical solution of the present invention, first, multi-domain screening and evaluation of cognitive function are carried out to achieve the purpose of rapid screening and detailed evaluation of cognitive function. Then, multi-modal magnetic resonance imaging data analysis is performed, including: extracting hippocampal volume and shape analysis, calculating functional connectivity values of cognitive-related brain functional networks (default network, salience network, fronto-parietal network, dorsal attention network, etc.), and using these as neuroimaging indicators to evaluate brain function. Then, Bayesian joint modeling of cognition and nerves is carried out, and finally, supervised learning classification is used to identify specific change patterns of Alzheimer's disease.

[0038] The technical solution according to the present invention has the following advantages:

[0039] 1. The combination of multi-modal data indicators is more accurate. The multi-modal data indicators of the present invention include general information (such as gender, age, education level, etc.), behavioral data from multi-domain screening and evaluation of cognitive function (such as scale scores, task-based assessment scores, etc.), and neuroimaging indicators (such as hippocampal volume, functional connectivity within and between brain functional networks, etc.).

[0040] 2. Based on the research results and technologies of cutting-edge neural networks, multi-modal MRI brain image data is analyzed.

[0041] 3. Advanced machine learning algorithms are used to deeply mine the AD-specific change patterns in multi-modal data. Brief Description of the Drawings

[0042] Figure 1 is a flowchart of the method for integrated evaluation of cognition and brain image data of Alzheimer's disease according to the present invention;

[0043] Figure 2 is a framework diagram for establishing a joint model by hierarchical Bayesian. Detailed Embodiments

[0044] The technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0045] As Figure 1 shown, the method for integrated evaluation of cognition and brain image data of Alzheimer's disease according to the present invention includes the following steps:

[0046] Step I: Obtain the behavioral data from multi-domain screening and evaluation of the cognitive function of the subject;

[0047] Step II: Analyze multi-modal magnetic resonance imaging data

[0048] 1. Quantitatively analyze the hippocampal volume and shape of the subject based on the fixed-point analysis method.

[0049] 2. Calculate the functional connectivity within and between brain functional networks

[0050] 2.1 Preprocess the resting-state fMRI (functional magnetic resonance imaging) data to correct possible biases in the original fMRI data;

[0051] 2.2 Extract brain functional networks and calculate the functional connectivity within brain functional networks

[0052] Based on the preprocessed fMRI data, localize the brain functional networks related to higher cognitive functions, and calculate the average functional connectivity z-value within each brain functional network, representing the functional connectivity within the brain functional network.

[0053] 2.3 Calculate the effective connectivity between networks

[0054] Based on the localized brain functional networks related to higher cognitive functions, select the main nodes of each brain functional network as the regions of interest, and calculate the effective connectivity between pairwise brain functional networks, aiming to explore the organizational relationship between each brain functional network;

[0055] Step III: Based on the general information (such as gender, age, education level, etc.) of AD and MCI patients, the behavioral data (such as scale scores, task-based assessment scores, etc.) from multi-domain screening assessments of cognitive functions, and neuroimaging indicators (such as hippocampal volume, functional connectivity within and between brain functional networks, etc.), input the model distribution parameters of behavioral and neural data into the same joint model, use the hierarchical Bayesian method to fit the joint model to the relevant data, obtain the central tendency and dispersion of the hyperparameter set, generate the joint posterior distribution of the parameters, and determine the degree and direction of the association between the cognitive and neural model parameters of mild cognitive impairment and Alzheimer's disease.

[0056] Step IV: According to the joint posterior distribution of the parameters obtained from the hierarchical Bayesian joint model, use the support vector machine (SVM) to train the specific change models for mild cognitive impairment and Alzheimer's disease.

[0057] I. Steps for obtaining the behavioral data from multi-domain screening assessments of the cognitive functions of the subject

[0058] The screening and assessment of multiple domains of cognitive function formally consist of two parts: classical neuropsychological assessment scales (which have been computerized); and computer-aided cognitive psychology behavioral assessment tasks. At the same time, according to different usage scenarios and applicable populations, different assessment tools can be combined to complete the rapid screening and detailed assessment of cognitive function.

[0059] 1. Classical neuropsychological assessment scales:

[0060] Clinically, the screening of various cognitive impairments often relies on neuropsychological scales. Commonly used screening scales include the Clock Drawing Task (CDT), Mini Mental State Examination (MMSE), and Montreal Cognitive Assessment (MOCA). In addition, researchers can use more detailed functional assessment scales for each sub-domain of cognition, including memory, visuospatial ability, executive function, attention, and activities of daily living. In terms of differentiating cognitive impairments caused by other reasons, neuropsychiatric questionnaires, Hamilton Depression Scale, Frontal Function Questionnaire, and Hachinski Ischemia Scale are often used. The above assessment scales have been computerized to facilitate clinical operation assessment and data storage.

[0061] 2. Computer-aided cognitive psychology behavioral assessment tasks

[0062] At the same time, for the above-mentioned cognitive function measurement items, whether one or several scales are selected to test patients, it is quite labor-intensive, material-consuming and time-consuming, and the accuracy of the assessment largely depends on the operation process and subjective experience of the tester, which is not conducive to the standardized promotion of the assessment. Based on the cognitive paradigms of classical cognitive neuroscience research, computer-aided cognitive psychology behavioral assessment tasks can enable the tested person to complete the assessment tasks independently under the standard guidance of the computer program and effectively evaluate the individual's sensory perception, memory, attention, agility, flexibility, logical thinking, and language ability. The advantage of task-based assessment is that it reduces the influence of subjective factors such as the tester on the assessment results, the indicators are more objective, the ability to evaluate and distinguish the cognitive function of patients is better; it is easy to operate; and it improves work efficiency. Therefore, this kind of computer-aided cognitive psychology behavioral assessment task can be used for the rapid screening of the elderly population or the detailed assessment of multi-domain cognitive abilities.

[0063] II. Multimodal magnetic resonance imaging data analysis

[0064] 1. Calculation of hippocampal volume and shape analysis

[0065] The volume of the hippocampus and shape analysis were calculated using the FIRST tool (FSL-integrated registration and segmentation toolbox, https: / / fsl.fmrib.ox.ac.uk / fsl / fslwiki / FIRST) in the FSL software package. First, the hippocampus was segmented from the whole-brain structural T1 image, and the volume of the bilateral hippocampi was calculated. Then, shape analysis of the hippocampus was performed, that is, the curved surface of the hippocampus was meshed, the normalized intensity of the mesh surface was sampled and modeled based on the multivariate Gaussian hypothesis, and then the shape was represented as the degree of variation of the average mode, so that quantitative analysis of the surface shape of the hippocampus could be carried out.

[0066] 2. Calculate the functional connectivity within and between brain functional networks

[0067] 2.1 Preprocessing of resting-state fMRI (functional magnetic resonance imaging) data

[0068] The purpose of the preprocessing process is to correct several possible biases in the original fMRI data to reduce or even eliminate the additional burden on subsequent analysis methods to correct or ignore these biases. Therefore, whether it is the resting-state fMRI data in the ADNI database or that of new subjects, preprocessing is required before a series of subsequent analyses can be carried out. The process of preprocessing resting-state fMRI data is as follows:

[0069] 1) Convert the image data from DICOM file format to NIFTI format recognized by the software for processing.

[0070] 2) Exclude the data of the first 10 time points of the image data to eliminate the error effects caused by uneven starting scanning magnetic fields and subject discomfort on the results.

[0071] 3) Temporal correction processing: Temporal correction processing was performed on the remaining 230 data (30 layers, scanning order [1:2:29, 2:2:30], reference layer is the 2nd or 29th layer), and the different scanning time points within the same TR were corrected to the same time point by mathematical methods for subsequent processing.

[0072] 4) Head correction, align each frame of the image in an experimental sequence with the first frame of this sequence according to a certain algorithm to correct the large error effects caused by subject head movement.

[0073] 5) Spatial normalization, mapping to the standard brain. The horizontal and rotational head movement maps of the subject can be obtained through head correction. The images of subjects with translation less than 1.5 mm and rotation less than 1.5° were mapped to the standard brain [-90, -126, -72; 90, 90, 108], and the voxel size was 3*3*3 mm.

[0074] 6) Smooth the image data with a smoothing kernel of [6 6 6] to improve the signal-to-noise ratio of the image signal after spatial normalization.

[0075] 7) Perform image detrending to remove the linear effects caused by machine temperature, subject adaptability, etc.

[0076] 8) Image filtering, with a frequency band of 0.01 - 0.1 Hz, to remove high-frequency signals.

[0077] 9) Extract covariates (head motion, whole brain, cerebrospinal fluid, white matter signal) and remove the covariates.

[0078] 2.2 Extract the brain functional network and calculate the functional connectivity within the network

[0079] The method of independent component analysis (ICA) can successfully extract the functional information of each local system of the brain only relying on the internal characteristics of the data and without the need to construct a model, etc., so it is often used in fMRI cognitive experiments and disease analysis. ICA can not only separate interference components such as heartbeat and respiration, but also successfully separate the components with spatial specificity in fMRI data (also known as "intrinsic connection network" or "resting state network").

[0080] The purpose of using the ICA method for analysis is: (1) Locate the main functional networks of the brain, including the brain functional networks related to higher cognitive functions such as memory, attention, and executive control. (2) Calculate the average functional connectivity z-value within the range of each brain functional network, representing the functional connectivity within the network, and input it as a part of the feature values in the classifier training into the model.

[0081] Based on the preprocessed fMRI data, select the Infomax algorithm to complete the calculation process of independent component analysis ICA, and use the ICASSO function to run 100 times repeatedly to find the most stable result of independent component separation, and then select the following main brain functional networks:

[0082] 1) Default Mode Network (DMN): The main brain regions are the medial prefrontal cortex, anterior cingulate gyrus, posterior cingulate gyrus and precuneus, bilateral angular gyrus, etc. It is negatively activated when the brain is focused on tasks, but is activated instead in the resting state, so it is named the default network and is related to meditation, etc.

[0083] 2) Salience Network (SN): The main brain regions are the bilateral anterior insulas and the middle anterior cingulate, forming a triangular structure that mainly functions as a switch / trip, i.e., evaluating input stimuli, finding the most relevant and appropriate stimuli, and switching to the relevant processing system.

[0084] 3) Frontoparietal Network (FPN): Also known as the executive control network, it is mainly distributed in regions such as the bilateral dorsolateral prefrontal cortex and the superior parietal gyrus. When extracted by ICA, it is separated into two left and right networks (LFP, RFP), which are related to working memory, executive function, etc.

[0085] 4) Dorsal Attention Network (DAN): The main brain regions are the bilateral intraparietal sulcus, inferior parietal gyrus, and superior occipital gyrus, etc. Its main function is to provide top-down attention orientation.

[0086] 2.3 Calculate the effective connectivity between networks

[0087] Based on the results of the brain functional network located by ICA, the main nodes of each of the above brain functional networks are selected as regions of interest (ROI), and the Coefficient-base Multivariate Granger Causal Analysis (mGCA) method is used to calculate the effective connectivity (EC) between pairwise brain functional networks, aiming to explore the organizational relationship between each brain functional network.

[0088] The Granger causality test is a statistical method for hypothesis testing, which tests whether a set of time series x is the cause of another set of time series y. Its basis is the autoregressive model in regression analysis. In the case of time series, the Granger causality between two variables X and Y is defined as: if, under the condition of including the past information of variables X and Y, the prediction effect of variable Y is better than the prediction effect of only using the past information of Y alone to predict Y, that is, variable X helps to explain the future changes of variable Y, then variable X is considered to be the Granger cause leading to variable Y.

[0089] If there are n time series (Y1, Y2,..., Y n ), for example, in the present invention, the main nodes of multiple brain functional networks are selected, then the Granger causality analysis model of n time series is determined, where the coefficient matrix (i.e., the effective connectivity to be calculated) is:

[0090]

[0091] Among them is the effective connectivity of Granger causality from the j-th brain functional network to the k-th brain functional network (generally setting i = 1), and the effective connectivity has directionality

[0092] III. Joint Modeling of Behavioral and Neural Data

[0093] According to the method for integrated assessment of cognitive and brain imaging data of Alzheimer's disease of the present invention, the model distribution parameters of the behavioral data of the multi-domain screening assessment of the cognitive function of the subject, the quantitative analysis data of the hippocampal volume and shape of the subject, and the within-network connection values and between-network effective connection data of the brain functional networks related to cognition are input into the same joint model. A joint model is established through hierarchical Bayesian, and the joint model is fitted to the relevant data using the hierarchical Bayesian method to obtain the central tendency and dispersion degree of the hyperparameter set, and generate the joint posterior distribution of the parameters, and determine the degree and direction of the association of the cognitive and neural model parameters of mild cognitive impairment and Alzheimer's disease

[0094] Based on the general information (such as gender, age, education level, etc.) of AD and MCI patients, the behavioral data (such as scale scores, task-based assessment scores, etc.) of the multi-domain screening assessment of cognitive function, and the neuroimaging indicators (such as hippocampal volume, functional connectivity within and between brain functional networks, etc.), the behavioral and neural data are jointly modeled using the hierarchical Bayesian method

[0095] As Figure 2 shown, the left side represents the neural data and its corresponding model parameters, the right side represents the behavioral data and its corresponding model parameters, and the two middle parameters respectively represent the central tendency and dispersion degree of the hyperparameter set Ω. The parameters θ j and δ j are conditionally independent, and the dependence relationship between the two can be used to jointly limit the parameter estimation of the joint model

[0096] For the neural data (N j ) and behavioral data (B j ) of the j-th user, assuming that the neural data follows the distribution N j ~Neural(δ j ), and the behavioral data follows the distribution B j ~Behav(θ j ), the parameters of the two data model distributions are written into a joint model, and this model can be expressed as (δ j , θ j) to M(Ω), where Ω represents the set of hyperparameters, and Ω may include a series of hyper-mean parameters Φ and hyper-dispersion parameters Σ. Therefore, Ω = {Φ, Σ}. The hierarchical Bayesian method is used to fit the relevant data to this joint model. According to the previous model definition, the joint posterior distribution of the joint model parameters can be written as:

[0097]

[0098] where p() represents the probability distribution; Behav(a|b) and Neural(a|b) represent the density functions of data a under the condition of given parameter b in the behavioral or neural model; M((a,b)|c) represents the density function of the parameter set (a, b) under the condition of given parameter c in the joint model.

[0099] Assume that the joint distribution form of the model (δ j , θ j ) follows a multivariate normal distribution, that is, (δ j , θ j ) ~ N p (Φ, Σ), where N p (a, b) represents the multivariate normal distribution of the mean vector a and the variance-covariance matrix b in the p-th dimension; the mean vector parameter Φ includes the means at all group levels. Therefore, Φ = {δ μ , θ μ}; the variance-covariance matrix includes the variance parameters at the group level, where ρ is the matrix of correlation coefficients of all model parameters of interest.

[0100]

[0101] Partition the variance-covariance matrix Σ to reflect that it is a mixture of a diagonal matrix and a full matrix when there are multiple parameters in the neural and behavioral vectors. Assume that there are 3 parameters in each of the neural (δ j ) and behavioral / cognitive (θ j ) models of subject j. When all the parameters of the neural and cognitive models are incorporated into the variance-covariance matrix, it can be written in the following form, where δ σ,1 represents the hyper-standard deviation in the first neural model parameter set, and θ σ,1 represents the hyper-standard deviation in the first cognitive model parameter set:

[0102]

[0103] The correlation coefficient parameter ρ reflects the degree and direction of the association between a pair of model parameters, and can directly infer to what extent a certain cognitive model parameter is related to a certain neural model parameter. The multivariate normal distribution is selected for the distribution of M(Ω) to adapt to the support of various parameter spaces, which is convenient for evaluating the relationship between neural model parameters and cognitive model parameters.

[0104] IV. Classification and prediction of AD-specific change patterns

[0105] Using the supervised learning support vector machine classification algorithm, specific change patterns of AD and MCI are found for classifying the population and for early detection and diagnosis of AD / MCI.

[0106] Using the SVM algorithm for data classification, the support vector machine SVM model constructed in the present invention is as follows: 90% of the hyperparameter set data of the hierarchical Bayesian joint model of AD patients and healthy controls in the ADNI database is used as the training set, and the remaining 10% is used as the test set. The hyperparameter set data of the hierarchical Bayesian joint model obtained after analyzing the training set data is used as the eigenvalue to input into the model. In addition, the kernel function type is adjusted, and a penalty function part is added to solve the non-linearly inseparable situation. When the classification prediction accuracy rate on the test set is above 80%, the SVM model classification prediction is considered reasonable, and then cross-validation is carried out.

[0107] The SVM algorithm uses a hypothesis space in the form of a linear function and is trained using an algorithm based on optimization theory. The best hyperplane for distinguishing between two types of data, AD patients and healthy controls, in the input space is obtained by measuring the margin and finding the point with the largest geometric margin.

[0108] Specifically, the hyperparameter set data of each user is represented as where represents the real number field, i = 1, 2,..., N, N represents the total amount of all user data, and each user data and the classification label (patient / healthy) of the user can be represented as y i ∈{-1, +1}.

[0109] In the D-dimensional feature space, the formula for the best hyperplane for distinguishing between two types of data is: wx i + b = 0. Where, w is the normal vector, and b is the position (intercept) of the plane relative to the data center. When the data x i satisfies the inequality wx i + b <= -1, the data is classified as an AD patient, and when it satisfies wx i + b >= 1, it is classified as a healthy population.

[0110] The best margin value is obtained by maximizing the closest distance 1 / (||w||) between the hyperplane and the data pattern. The following Lagrange multiplier is used to determine the maximum or minimum relative value of the function restricted by the constraint conditions, α i is the Lagrange multiplier corresponding to the data x i .

[0111]

[0112] By adding constraint conditions and solving the following equation for the maximum point, the support vector parameters are obtained:

[0113]

[0114]

[0115] For the non - linear distribution problem of neural and cognitive posterior joint parameters, further adjust the kernel function type (including linear, polynomial, radial basis, sigmoid function); map the data to a high - dimensional space and construct an optimal separating hyperplane in the high - dimensional feature space. Use grid search to find the optimal kernel parameters; use the optimal parameters to train the classification model. For the problem that some sample points do not satisfy the constraint condition that the functional margin is greater than or equal to 1, add a penalty function part on the basis of the linearly separable problem:

[0116]

[0117] s.t.y i (w T x i +b)≥1 - ξ i

[0118] ξ i ≥0

[0119] i = 1, 2,..., N

[0120] Then, calculate the accuracy of the prediction model according to the leave - one - out cross - validation method, and take the higher one as the better, and determine the final prediction model and the function expression.

[0121] The process of evaluating by leave - one - out cross - validation: Repeatedly divide the obtained sample data, combine them into different training sets and test sets, use the training set to train the model, and use the test set to evaluate the quality of the model prediction. On this basis, multiple different training sets and test sets can be obtained. A certain sample in a certain training set may become a sample in the test set next time, that is, the so - called "cross". The specific 10 - fold cross - validation includes the following steps:

[0122] The first step is to randomly divide the original data into 10 non - repeating samples.

[0123] The second step is to select one of them as the test set each time, and the remaining 9 as the training set for model training. After training on each training set, a model is obtained, and this model is used to test on the corresponding test set, and calculate and save the evaluation index of the model: classification accuracy.

[0124] The third step is to repeat the second step 10 times, so that each piece of original data has a chance to be used as the test set once, and the rest of the opportunities as the training set.

[0125] In the fourth step, calculate the average value of the 10 test results as an estimate of the model accuracy and use it as the performance metric (true classification rate) of the current model. Screen the SVM classification prediction algorithm with the highest classification rate.

[0126] Further perform model correction: First, perform the same analysis and processing on the brain imaging data of the newly added subjects, jointly establish a model with cognitive and neural data, and generate the posterior distribution of hyperparameters. Input the hyperparameters into the classification model trained based on the ADNI dataset to perform AD disease classification prediction on this newly added subject to assist doctors in diagnosis. When the doctor obtains the diagnosis result by synthesizing multi-dimensional information, the true diagnosis result will be recorded into the model, continuously correct the prediction accuracy of the model, make the model closer and closer to various situations in the real environment, and improve the sensitivity and accuracy of the model.

[0127] The above are only embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A cognitive and brain imaging data integration and evaluation method for Alzheimer's disease, characterized in that The method includes the following steps: Step I: Obtain the behavioral data of the multi-domain screening assessment of the cognitive function of the subject; Step II: Analyze the multi-modal magnetic resonance imaging data, including: (1) Perform magnetic resonance imaging analysis on the brain structure, and quantitatively analyze the hippocampal volume and shape of the subject based on the fixed-point analysis method; (2) Preprocess the resting-state functional magnetic resonance imaging data to correct the deviation existing in the original resting-state functional magnetic resonance imaging data; (3) Based on the preprocessed resting-state functional magnetic resonance imaging data, localize the brain functional networks related to cognition, calculate the functional connection values within each brain functional network, and based on the localized brain functional networks related to higher cognitive functions, select the main nodes of each brain functional network as the regions of interest, and calculate the effective connection between pairwise brain functional networks; Step III: Input the model distribution parameters of the behavioral data of the multi-domain screening assessment of the cognitive function of the subject, the quantitative analysis data of the hippocampal volume and shape of the subject, and the intra-connectivity values within the cognitive-related brain functional network and the effective connectivity data between brain functional networks into the same joint model. Use the hierarchical Bayesian method to fit the joint model to the relevant data, establish a hierarchical Bayesian joint model, obtain the central tendency and dispersion degree of the hyperparameter set, generate the joint posterior distribution of the parameters, and determine the degree and direction of the association of the cognitive and neural model parameters of mild cognitive impairment and Alzheimer's disease; In Step III, for the neural data N of the j-th user j and the behavioral data B j , assume that the neural data follows the distribution N j ~Neural(δ j ), and the behavioral data follows the distribution B j ~Behav(θ j ). Write the parameters of the two data model distributions into a joint model, which is expressed as (δ j , θ j )~M(Ω), where Ω represents the hyperparameter set, and Ω includes a series of hyper-mean parameters Φ and hyper-dispersion parameters Σ. Therefore, Ω = {Φ, Σ}; Use the hierarchical Bayesian method to fit the joint model (δ j , θ j )~M(Ω) to the relevant data. Among them, the joint posterior distribution of the joint model parameters is written as: , where p() represents a probability distribution, Behav(a|b) and Neural(a|b) represent the density functions of data a under the condition of given parameter b in a behavioral or neural model, and M((a,b)|c) represents the density function of the parameter set (a, b) under the condition of given parameter c in a joint model; assume that the joint distribution form of the model (δ j , θ j ) follows a multivariate normal distribution, that is, (δ j , θ j ) ~ Np(Φ, Σ), where Np(a, b) represents the multivariate normal distribution of the mean vector a and the variance-covariance matrix b in the p-th dimension. The mean vector parameter Φ includes the means at all group levels. Therefore, Φ = {δ μ , θ μ}; the variance-covariance matrix includes the variance parameters at the group level, where ρ is a matrix including the correlation coefficients of all model parameters of interest, , the variance-covariance matrix Σ is partitioned to reflect that it is a mixture of a diagonal matrix and a full matrix when multiple parameters are included in the neural and behavioral vectors. ρ reflects the degree and direction of the association between a pair of model parameters, directly inferring to what extent a certain cognitive model parameter is related to a certain neural model parameter; for the distribution of M(Ω), a multivariate normal distribution is chosen to accommodate the support of various parameter spaces, facilitating the evaluation of the relationship between neural model parameters and cognitive model parameters; Step IV: According to the joint posterior distribution of the obtained parameters, use a support vector machine to train a specific change model for mild cognitive impairment and Alzheimer's disease; in Step IV, represent the hyperparameter set data of each user as , where, represents the real number field, i = 1, 2, …, N, N represents the total amount of all user data, and each user data and the user's classification label "patient / healthy" are represented as y i ∈{-1, +1}. In the D-dimensional feature space, the formula for the optimal hyperplane that distinguishes patient or healthy data is: wx i +b = 0, where, w is the normal vector and b is the position of this plane relative to the data center. When the data x i satisfies the inequality , this data is classified as the data of Alzheimer's patients, and when it satisfies , it is classified as the data of healthy people.

2. The cognitive and brain imaging data integration and evaluation method for Alzheimer's disease according to claim 1, characterized in that The steps of preprocessing the resting-state functional magnetic resonance imaging data are as follows: 1) Convert the image data file format into a format recognizable by the software for processing; 2) Eliminate the data of the first 10 time points of the image data to exclude the error effects caused by the uneven magnetic field at the beginning of the scan and the discomfort of the subject on the results; 3) Time correction processing: Perform time correction processing on the remaining 230 data, and correct different scan time points within the same TR to the same time point by mathematical methods for subsequent processing; 4) Head correction, align each frame of the image in an experimental sequence with the first frame of this sequence according to a certain algorithm to correct the error effects caused by the head movement of the subject; 5) Spatial normalization, mapping to the standard brain. Head correction is performed to obtain the horizontal head movement and rotational head movement images of the subject. The images of subjects with translation less than 1.5 mm and rotation less than 1.5° are mapped onto the standard brain [-90, -126, -72; 90, 90, 108], with a voxel size of 3 3 3 mm; 6) Smooth the image data, and the smoothing kernel is [6, 6, 6], which is used to improve the signal-to-noise ratio of the image signal after spatial normalization; 7) Perform image detrending to remove the linear effects caused by the machine temperature and the adaptability of the subject; 8) Image filtering, and the frequency band is 0.01 - 0.1 Hz to remove high-frequency signals; 9) Extract covariates and remove the covariates.

3. The method for integrated evaluation of cognition and brain imaging data of Alzheimer's disease according to claim 1, wherein Locate the brain functional networks related to cognition based on ICA.

4. The cognitive and brain imaging data integration and evaluation method for Alzheimer's disease according to claim 1, wherein Use the SVM algorithm in the supervised learning algorithm to determine the specific change model of mild cognitive impairment and Alzheimer's disease. Among them, 90% of the hyperparameter set data of the Alzheimer's disease patients and healthy controls obtained to establish a hierarchical Bayesian joint model is used as the training set, and the remaining 10% is used as the test set. When the classification prediction accuracy rate on the test set is above 80%, it is considered that the SVM model classification prediction is reasonable.

5. The cognitive and brain imaging data integration and evaluation method for Alzheimer's disease according to claim 1, characterized in that Evaluate the parameters of the specific change model of mild cognitive impairment and Alzheimer's disease through cross-validation, specifically including the following steps: The first step is to randomly divide the original data into 10 parts without repeated sampling; The second step is to select 1 part of them as the test set each time, and the remaining 9 parts are used as the training set for model training. After training on each training set, a model is obtained, and this model is used to test on the corresponding test set, and the classification accuracy rate of the model is calculated and saved; The third step is to repeat the second step 10 times, so that each part of the original data has a chance to be used as the test set, and the rest of the opportunities are used as the training set; The fourth step is to calculate the average value of the 10 test results as an estimate of the model accuracy and use it as the performance index of the current model.

Images