KR102599756B1 - Providing method of diagnostic information on alzheimer's disease using structural, diffusion, and functional neuroimaging data and the APOE genotype - Google Patents

Abstract

The present invention relates to a method of providing diagnostic information for classifying Alzheimer's disease or mild cognitive impairment from normal people. To form a multi-modal system for distinguishing Alzheimer's disease and to increase classification accuracy, various neuroimaging techniques (sMRI, FDG-PET, AV45, -PET, DTI, and rs-fMRI) relates to a method of providing diagnostic information by combining apolipoprotein-E genotype.

Description

{Providing method of diagnostic information on Alzheimer's disease using structural, diffusion, and functional neuroimaging data and the APOE genotype}

The present invention relates to a method of providing diagnostic information for Alzheimer's disease using multimodal features including structural, diffusion, and functional neuroimaging and APOE genotype.

Alzheimer's disease (AD) is a neurodegenerative disorder characterized by chronic cortical atrophy (e.g. posterior cingulate atrophy and medial temporal atrophy) and progressive decline in cognitive function. Alzheimer's disease is usually diagnosed in people over 65 years of age. As the population lives longer, the prevalence of Alzheimer's disease and its cost to society are also increasing. Therefore, detection of Alzheimer's disease or its precursor form, mild cognitive impairment (MCI), is an important goal of biomedical research to provide new treatments to help slow the progression of Alzheimer's disease. Mild cognitive impairment is a transitional stage (meaning an intermediate stage of functional and cognitive decline between normal aging and dementia patients), a broad stage followed by memory impairment in the absence of dementia. Mild cognitive impairment, which is cognitive impairment in multiple domains until the disability threshold is reached, is called systemic Alzheimer's disease; The subject goes on to develop Alzheimer's disease (this type of patient belongs to the MCI-coverting (MCIc) group). Symptoms appear, on average, within 2 to 3 years (Lopez et al., 2012). Prospective population-based studies in older adults have shown that the conversion rate of MCIc patients to AD or other forms of dementia is approximately 10–15% per year (Lopez et al., 2012; Mitchell and Shiri-Feshki, 2009; Wei et al., 2009). al., 2016). Despite considerable knowledge about MCIc, between 47% and 67% of subjects diagnosed with MCI do not convert to normal or dementia (Clem et al., 2017; Ganguli et al., 2004; Lopez et al., 2012). There is little understanding. In a study of people identified with MCI over a 10-year period in a large community sample, 21% of those suspected to be at greater risk of converting to dementia (Dubois and Albert, 2004; Jicha et al., 2006) were diagnosed with MCI (Clem et al., 2017; Ganguli et al., 2004). The above studies suggest that certain subjects may remain diagnostically stable for a period of time without converting to AD (this type of patient belongs to the MCI-stable group or MCI group). Recent results from neuroimaging studies have shown that AD, even in the early stages of MCI or prior to the transition to AD, is a disconnection syndrome (FIG. This supports the hypothesis that changes involve functional changes throughout the network (Bishop et al., 2010; Clem et al., 2017; Daianu et al., 2013; de Vos et al., 2018; Dubois). and Albert, 2004).

Previous studies have shown the potential of invasive and non-invasive biomarkers to predict conversion from MCI to AD dementia. For invasive markers, APOE-ε4 genotype (for the APOE-ε4 allele, brain changes associated with AD may occur at an early age) (Dean et al., 2014; Liu et al., 2013) and amyloid-beta (amyloid-β). beta, Aβ) accumulation and neurofibrillary lesions are considered the most important biomarkers for AD (Murphy and LeVine, 2010). Apolipoprotein-E (APOE) genotype polymorphism is considered the most common polymorphism in neurodegenerative diseases and is consistently associated with normal cognitive decline in AD and MCI patients. APOE-ε4 is the strongest genetic risk factor and increases AD risk by 2- to 3-fold. It also lowers the age of onset of AD (Michaelson, 2014). Recent developments in non-invasive neuroimaging techniques, including functional and structural imaging, have resulted in a variety of commonly used neuroimaging biomarkers for AD. Among several neuroimaging modalities, structural magnetic resonance imaging (sMRI) has attracted considerable attention due to its ready availability and high spatial resolution for patients with mild symptoms (Cuingnet et al., 2011; Gupta et al., 2011). 2019b, 2019c; Long et al., 2017; Salvatore et al., 2015; Sun et al., 2019; Wei et al., 2016). sMRI can also reveal abnormalities in a wide range of brain regions, including gray matter (GM) atrophy in the medial temporal lobe and the hippocampus/entorhinal cortex, which has been identified as an important AD-specific biomarker for differentiation or classification of AD patients. (Cuingnet et al., 2011). In diffusion tensor imaging (DTI or diffusion MRI), water diffusion in the brain is interpreted as MR signal loss. Because neurodegenerative processes are accompanied by the loss of barriers that restrict the movement of water molecules (Acosta-Cabronero and Nestor, 2014), DTI may represent a promising marker of ultrastructural white matter (WM) damage in AD and MCI patients. You can. Connectivity-based analysis applying graph theory to DTI data disrupts the topological properties of structural brain networks in AD, supporting disconnection theory. In particular, regional diffusion metrics over the limbic WM of the fornix, posterior cingulum, and parahippocampal gyrus showed better performance than volumetric measures of GM in predicting MCI transformation (Sun et al., 2019). In clinical studies, fluorodeoxyglucose-positron emission tomography (FDG-PET), florbetapir-PET AV45 (amyloid protein imaging), and resting-state fMRI (rs-fMRI) are the most commonly used functional neuroimaging methods for AD diagnosis (Gupta et al., 2019a; Hojjati et al., 2018, 2017; Pan et al., 2019b, 2019a). FDG-PET measures cerebral glucose metabolism via 18F-FDG in the brain and helps detect the characteristic regional hypometabolism in AD patients, which reflects neuronal dysfunction in the brain of AD patients (Mosconi et al., 2008 ). In contrast, florbetapir-PET AV45 measures accumulation of amyloid protein in AD brain homogenates and has faster in vivo kinetics. The use of florbetapir in amyloid imaging was recently validated in an autopsy study, and its safety profile allows for clinical application in brain imaging (Camus et al., 2012). Additionally, rs-fMRI imaging has been developed as a tool to map brain intrinsic activity and depict synchronization of FC between regions (Bi et al., 2018; de Vos et al., 2018). Recent rs-fMRI studies have shown that FC patterns may be altered in some specific functional networks (default mode networks) in AD and MCI patients. The authors found decreased FC between the hippocampus and several regions throughout the cerebral cortex, namely decreased FC within the default mode network and increased FC within the frontal network (de Vos et al., 2018).

In the above studies, there was no clear evidence supporting the superiority of other biomarkers (CSF vs. APOE-ε4 vs. imaging) for diagnostic assessment of Alzheimer's disease. The choice of biomarker largely depends on price and availability. Nevertheless, some authors argue for the perfection of imaging over fluid biomarkers, considering that the imaging modality can distinguish between different stages of the disease anatomically and temporally (Khoury and Ghossoub, 2019; Marquez and Yassa, 2019 ). The studies mentioned above used only a single modality to detect biomarkers for detecting conversion from MCI to AD. The performance of the proposed algorithm is around 80-90%, which is lower compared to the performance of recently published multimodal studies (Fayao Liu et al., 2014; Gupta et al., 2019a; Ritter et al., 2015; Xu et al., 2016; Zhang et al., 2011). It is true that no single imaging modality for biomarkers to date meets all the diagnostic requirements established by previous studies (Hyman et al., 2012; Jack et al., 2018, 2016) and that each biomarker Although no single biomarker (whether genotypic, fluid, or imaging) can by itself accurately distinguish the heterogeneous disorders of AD with high accuracy, multiple methods can provide complementary information. This may lead to the need to develop panels containing CSF data or a combination of imaging and APOE that merge information about disease mode to improve diagnostic accuracy as neuroimaging biomarkers (Marquez and Yassa, 2019) . Combining information (multimodal) from different types of neuroimaging (structural and functional) with genotypic (APOE) or biochemical (CSF) information, such as sMRI, AV45-PET, FDG-PET, DTI, and rs-fMRI, results in a single modality. It may help improve the diagnostic performance of AD or MCI compared to methods (Gupta et al., 2019a; Schouten et al., 2016; Wei et al., 2016; Young et al., 2013; Zhang et al. , 2011). Moreover, recent combinations of biomarkers yield powerful diagnostic techniques to classify AD groups into cognitively healthy subjects, with specificity and sensitivity scores reaching over 90% (Bloudek et al., 2011; Khoury and Ghossoub, 2019; Rathore et al., 2017).

Several studies have reported combinations of various neuroimaging techniques to investigate AD or MCI. Dai and collaborators (Dai et al., 2012) used regional GM volumetric measurements and functional measures (amplitude of low-frequency fluctuations, regional homogeneity, and regional FC intensity) as features. They trained distinct maximum uncertainty LDA classifiers on functional and structural features and merged the outputs of the classifiers through weighted voting. Zhang and collaborators (Zhang et al., 2011) used a multimodal method to distinguish healthy controls (HC) from AD patients. They used a kernel-based support vector machine (SVM) classifier for classification and combined volumetric features with regional FDG-PET and CSF biomarkers. Young and collaborators (Young et al., 2013) proposed a method to combine sMRI, FDG-PET, and APOE genotyping data to distinguish HC from AD. These authors used Gaussian process as a multi-mode kernel method and applied SVM classifier for classification of MCI vs. MCIc groups. However, diagnostic accuracy was low. In another study, multiple kernel learning (MKL) using the Fourier transform of the Gaussian kernel was applied to AD classification using both sMRI and rs-fMRI (Fayao Liu et al., 2014) neural images. suggested. Moradi and co-workers (Moradi et al., 2015) used GM density maps, age and cognitive tests as features and used classification algorithms such as low-density separation and random forest for AD transformation discrimination. Another study proposed a new method to classify AD using multi-feature techniques (regional thickness, regional correlation calculated from thickness measurements, and APOE genotype) using an SVM classifier (Zheng et al., 2015). Schouten and collaborators (Schouten et al., 2016) combined regional volumetric measurements, diffusion measurements, and correlation measurements between all brain regions calculated from functional MRI. They used a logistic elastic net for classification. Additionally, Liu and collaborators (Liu et al., 2017) used independent component analysis and COX model to distinguish between MCI and MCIc. In the study, they used sMRI and FDG-PET scans along with APOE data and some cognitive measures. Hojjati and co-workers (Hojjati et al., 2018) used SVM for classification by combining features extracted from sMRI (cortical thickness) and rs-fMRI (graph measurements) for AD detection. It is worth noting that most of the multimodal methods presented above used brain atrophy in several regions manually extracted as features of sMRI and PET images for AD detection among different groups. However, the use of only a small number of brain regions as a feature of any imaging modality may not accurately reflect the overall spatiotemporal pattern of structural and physiological abnormalities (Fan et al., 2008). Additionally, combining modalities simply by increasing the number of modalities does not increase predictive power.

Therefore, the main goal of the present invention was to combine five different imaging modalities (sMRI, AV45, FDG-PET, DTI and rs-fMRI) with APOE genotyping to build a multimodal system for AD detection. Additionally, three methods (completely different from each other) were used to distinguish AD from other groups. Additionally, no single modality of neuroimaging can achieve high performance or classify six binary classification groups (AD vs. HC, MCI vs. MCIc, AD vs. MCIc, AD vs. MCI, HC vs. MCIc, HC vs. MCI). Based on this, we wanted to know which combined methods performed well in the classification step (whole brain or voxel analysis). We also aimed to use voxel of interest (VOI) and graph methods to discover areas where the above six binary groups differ significantly from each other. Whole brain segmentation and voxel approach were used to study regional and voxel differences in all six binary classification groups. NiftyReg (Gupta et al., 2019a; Young et al., 2013), pyClusterROI script (Craddock et al., 2012), PANDA (Cui et al., 2013), DPRASF (Yan, 2010) and SPM12 (Ashburner and Friston , 2001), the integrated CAT12 toolbox extracted features from structural and functional neuroimaging data. Additionally, graph-based analyzes (John et al., 2017; Peraza et al., 2019) were performed to study the organization of network connectivity using anatomical features (including GM volume, cortical thickness, and WM pathways between GM regions). , regional time series of 200 brain regions included in the Craddock atlas were used. For this graph-based analysis, we used the BRAPH toolbox (Mijalkov et al., 2017). Later, the MKL algorithm based on the EasyMKL (Aiolli and Donini, 2015; Donini et al., 2019) classifier was applied for classification and data fusion. This classifier works by simultaneously learning predictor variables and kernel combination weights. We also applied a one-time cross-validation technique to help find the optimal hypermeter for the MKL classifier. In this study, a radial basis function (RBF)-SVM classifier was applied and the results were compared with those obtained from EasyMKL. Our results show that grouping different measurements (or complementary information) from six different modalities results in better results for all six binary classification groups (combined ROI or combined VOI or (using a combination of all) showed much better performance.

Korean Patent No. 1929965 Korean Patent No. 2241357 Korean Patent Publication No. 2143940

The purpose of the present invention is to combine neuroimaging methods (sMRI, AV45, FDG, rs-fMRI, and DTI) with genetic biomarkers (APOE) to detect Alzheimer's disease or mild cognitive impairment in normal subjects using whole-brain, voxel-wise, and graphical analysis methods. It provides a method of providing diagnostic information for classification of disorders.

The present invention includes 1) calculating a kernel matrix; 2) Kernel matrix combination step; 3) Kernel matrix normalization step; 4) Kernel matrix selection and classification step; 5) Cross-validation step; and 6) classifying the stage of Alzheimer's disease, wherein the 1) kernel matrix calculation step includes 1-1) calculating the kernel matrix through whole-brain parcellation analysis. ; 1-2) Calculating a kernel matrix using voxel-based morphometry; and 1-3) calculating a kernel matrix using a graph theory method.

According to one embodiment of the present invention, the kernel matrix includes sMRI, FDG-PET, AV45-PET, DTI, and rs-fMRI, and the kernel matrix includes apolipoprotein E genotype ( apolipoprotein E genotype).

According to another embodiment of the present invention, step 1-1) calculates the kernel matrix using NiftyReg, pyClusterROI, and PANDA toolboxes, and the 1-1) Step 2) is to calculate the kernel matrix using SPM12, DPARSF and FSL toolbox.

According to another embodiment of the present invention, the rs-fMRI features include REHO, ALFF, and FALFF.

In addition, according to one implementation of the present invention, step 4) is classified as multiple kernel learning, and multiple kernel learning uses the following [Equation 12].

[Equation 12]

In [Equation 12] above, sMRI, FDG, AV45, DTI, rs-fMRI, and APOE are normalized by K(x,y), the variable is the normalized value of the kernel, and w is the corresponding weight. A w value is assigned to each normalized value so that the sum of the six values is 1.

According to another implementation of the present invention, step 5) uses leave-one-out cross-validation.

According to another embodiment of the present invention, the progression stage of Alzheimer's disease is one selected from the group consisting of healthy control group, mild cognitive impairment, conversion cognitive disorder, and Alzheimer's disease.

The method of providing diagnostic information for classifying the progression stage of Alzheimer's disease according to the present invention is the optimal method for classifying the progression stage of Alzheimer's disease according to the characteristics of various clinical indicators and changes in the extraction process compared to tests performed only with specific indicators. It can have the effect of providing a diagnostic tool.

Figure 1 shows an overview of a multimodal framework.
Figure 2 shows a pipeline showing the feature extraction process for rs-fMRI images.
Figure 3 shows the MKL classification process.
Figure 4 shows (A) AD vs. whole-brain parcellation analysis. HC, (B)MCIs vs. MCIc, (C) AD vs. MCIs, (D) AD vs. MCIc, (E)HC vs. MCIs, and (F) HC vs. This shows the ROC curve for MCIc.
Figure 5 shows Cohen's kappa plot using whole-brain parcellation analysis.
Figure 6 shows the classification results using region of interest (ROI) features (EasyMKL).
Figure 7 shows the classification results using voxel-wise (VOI) features (EasyMKL).
Figure 8 shows Cohen's kappa plot using voxel-wise analysis.
Figure 9 shows (A) AD vs. AD using AV45-PET images. HC, and (B) MCIs vs. This shows the area affected by MCIc.
Figure 10 shows (A) AD vs. AD using DTI images. HC, and (B) MCIs vs. This shows the WM voxels selected for MCIc classification.
Figure 11 shows (A) AD vs. voxel-wise analysis. HC, (B)MCIs vs. MCIc, (C) AD vs. MCIs, (D) AD vs. MCIc, (E)HC vs. MCIs, and (F) HC vs. This shows the ROC curve for MCIc.
Figure 12 shows classification results using combined-(VOI+ROI) features with whole-brain and voxel-wise features (EasyMKL).
Figure 13 shows (A) AD vs. AD using combined-(VOI+ROI). HC, (B) MCIs vs. MCIc, (C) AD vs. MCIs, (D) AD vs. MCIc, (E)HC vs. MCIs, and (F) HC vs. This shows the ROC curve for MCIc.
Figure 14 shows a weighted correlation matrix graph of 384 regions of AD, HC, MCIs, and MCIc for sMRI biomarkers.
15 shows AD vs. AD in global structural topology. This shows the difference between the HC group.
Figure 16 shows MCIs vs. MCIs in global structural topology. This shows the difference between the MCIc group.
17 shows AD vs. HCs and MCIs vs. This is a brain map showing the most expected area to distinguish between the MCIc group.
Figure 18 shows a comparison of classification results between EasyMKL and RBF-SVM classifiers.
Figure 19 shows the most significant brain regions found in all six binary classification groups using voxel-wise and graph analysis.

Hereinafter, the present invention will be described in more detail through examples. Since these examples are merely for illustrating the present invention, the scope of the present invention is not to be construed as limited to these examples.

<Example 1> Participants

For this study, we downloaded the data of all subjects for whom all five imaging modalities (sMRI, rs-fMRI, FDG-PET, AV45-PET, and DTI) along with APOE genotype were available from the ADNI website. A total of 129 subjects were classified as healthy control (HC; n = 35), MCIc (n = 31), MCI (n = 30), or AD (n = 33) and matched in gender and age ratio. In the HC group, participants had a global clinical dementia rating (CDR) score of 0, a mini-mental state examination (MMSE) score of 27 to 30, and a functional activities questionnaire (FAQ). ) score is 0 to 4, and geriatric depression scale (GDS) score is 0 to 4. In the MCI group, the MMSE score is 25 to 30 points, the FAQ score is 0 to 16 points, and the GDS score is 0 to 13 points. In the MCIc group, the MMSE score was 19 to 30 points, the FAQ score was 0 to 18 points, and the GDS score was 0 to 10 points. In the AD group, patients had an overall CDR score of 1, MMSE score of 14 to 24, FAQ score of 3 to 28, and GDS score of 0 to 7.

MCI subjects who were followed for less than 18 months and did not convert within this period were not considered. Table 1 shows participant demographic information, including mean age and gender ratio by group.

Group AD (n = 33) MCIc (n = 31) MCIs (n = 30) HC (n = 35) Sex (M/F) 21/12 16/15 17/13 14/21 Age 75.65± 8.61 72.27± 7.40 72.90± 7.86 77.83±6.17 Body Weight (kg) 75.81±13.60 80.71±17.43 82.14±15.03 75.76±18.62 FAQ score 19.34±6.53 6.16± 7.38 1.86± 2.99 0.13± 0.48 NPI-Q score 4.46± 4.01 2.64±3.32 1.66± 1.61 0.33± 0.78 GDS score 2.37± 2.59 2.03± 2.02 1.26± 0.96 1.13± 1.79 MMSE score 19.59± 4.56 26.32± 3.85 28.03± 1.25 29.13± 1.20

To assess statistically significant changes in demographic and clinical characteristics between these groups, Student's t-test was used with the significance level set at 0.05. No significant differences (p-value > 0.05) were found between groups in age or gender ratio.

To obtain unbiased performance estimates, the classification group was randomly split into two clusters in a 70:30 ratio for the training and test sets. The model was trained on the training set, and performance measurements of diagnostic specificity and sensitivity were performed on a separate test set. Age and gender distribution were maintained during the segmentation stage.

<Example 2> Three Features

2-1. sMRI acquisition

1.5T T1-weighted MR images were obtained from the ADNI website. MRI images were obtained in a data center using Philips, GE, or Siemens Medical system scanners. Because each scanner has different acquisition protocols, the image normalization step was performed by ADNI. Image correction included correction, distortion of image geometry due to gradient nonlinearity (grad-warp), reduction of intensity non-uniformity due to undulation, or residual intensity non-uniformity of the 1.5T scan used for each image by ADNI. Detailed information about sMRI imaging can be found on the ADNI website (http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/). All scans had resolution with intervals between each scan. In the present invention, the obtained sMRI images were preprocessed again using the toolbox of the FMRIB software library (FSL, v.6.0) ( Smith, 2002 ). For anatomical sMRI images, this involves extracting non-brain tissue from each image using the BET function. We then passed the skull-stripped images into the ANT (Tustison et al., 2010) toolbox for N4 bias field correction to correct for non-uniform artifacts in each image. We also used the FSL toolbox (Jenkinson et al., 2012) for coregistration to the standard Montreal Neurological Institute (MNI) 152 template (Grabner et al., 2006).

2-2. FDG-PET image acquisition

ADNI offers four types of FDG-PET samples: 1) Co-registered Dynamic; 2) co-registration, average; 3) Co-registration, average, normalized image and voxel size and 4) Co-reg, Avg, Std Img and Vox Siz, named Uniform Resolution. Type (3) standard FDG-PET images were downloaded from the ADNI website.

The downloaded baseline FDG-PET sample was in DICOM format. In the first step, these DICOM format images were converted to Nifty format using the dcm2nii (Li et al., 2016) toolbox. Later, these images were spatially normalized to the MNI 152 template using the SPM12 toolbox (integrated in MATLAB 2019b) with a standard 91 x 109 x 91 tensor dimension image grid with voxel size of 2 x 2 x 2 ^mm3 . This image grid was derived from the patient's posterior (AC-PC) axis being parallel to the AC-PC line. The above regularization step is performed at two levels: a global affine transformation followed by a non-rigid spatial transformation. A typical affine transform requires a 12-parameter design, whereas a non-rigid space transform uses a sequence of the lowest frequency elements of a three-dimensional cosine transform. Additionally, an intensity normalization step was performed by segmenting each voxel by depth with the average score of the global GM obtained with the help of the AAL template image. Detailed information about FDG-PET imaging can be found on the ADNI website (http://adni.loni.usc.edu/methods/pet-analysis-method/pet-analysis/). Additionally, after completing these preprocessing steps, the FDG-PET images were coregistered to the corresponding sMRI T1-weighted images using the SPM12 toolbox.

2-3. AV45-PET image acquisition

ADNI offers four types of AV45-PET samples: 1) AV45 co-registered, dynamic; 2) AV45 co-registration, average; 3) AV45 Co-reg, Avg, standardized image and voxel size and 4) AV45 Co-reg, Avg, Std Img and Vox Siz, uniform resolution. Type (3) standard AV45-PET image was downloaded from the ADNI website. For each scan, 5-min frames (4 for flobetapyr, acquired 50–70 min post-injection) were captured using the NeuroStat “mcoreg” routine (followed by ADNI histology) (rigid-body translation/rotation, 6 free). (Jagust et al., 2015). The downloaded baseline AV45-PET samples were in DICOM and ECAT format. In the first step, these DICOM and ECAT format images were converted to Nifty format using the dcm2nii (Li et al., 2016) toolbox. Later, these images were scanned into the SPM12 toolbox (MATLAB ^2019b ) with a standard 91 was spatially normalized to the MNI 152 template using (incorporated in). More detailed information about AV45-PET imaging can be found on the ADNI website (http://adni.loni.usc.edu/methods/pet-analysis-method/pet-analysis/). Additionally, after completing the preprocessing steps mentioned above, AV45-PET images were coregistered to the corresponding sMRI T1-weighted images using the SPM12 toolbox.

2-4. Resting-state functional MR image (rs-fMRI) acquisition

fMRI images were acquired using a 3.0-T Philips Medical sMRI scanner. All rs-fMRI images were retrieved from the ADNI website. The sample for each subject consisted of 6720 DICOM images. Patients were asked to lie in the scanner, relaxed and without thinking during the scanning process. Sequence parameters are as follows: pulse sequence = GR, TR = 3000 ms, TE = 30 ms, flip angle = 80°, data matrix = 64 x 64, pixel spacing X, Y = 3.31 mm and 3.31 mm, slice thickness. = 3.33 mm, axial slice = 48, no slice gap, viewpoint = 140. Because the signal-to-noise ratio (SNR) of rs-fMRI images is limited, the collected data does not reflect the effect of noise on fMRI images. preprocessed to reduce For preprocessing of rs-fMRI images, we used the Data Processing Assistant for Resting-state fMRI (DPARSF) (Yan, 2010) software, available at (http://d.rnet.co/DPABI/DPABI_V2.3_170105.zip) ) can be downloaded from. For each subject, the entire preprocessing is divided into 9 steps: converting DICOM format to NIFTY format, removing the first 10 viewpoints, slicing timing, head movement correction to adjust the time series of images, and ensuring that the brain is positioned in the same orientation in all images. , normalization, smoothing at full width at half maximum (FWHM), removes residual noise that systematically increases or decreases over time by removing linear trends (Smith et al., 1999), and temporal noise to maintain 0.01-0.08 Hz variation. Filtering, and covariate removal to remove physiological artifacts (Fransson, 2005), non-neuronal blood oxygen level-dependent (BOLD) fluctuations, and head movements.

2-5. Diffusion tensor imaging (DTI) imaging acquisition

DTI images were also downloaded from the ADNI website. The DTI protocol uses spin- ^echo diffusion with TR/TE of 12000/1046 ms, voxel size of ^0.9375 Weighted echo-planar imaging was used. More information about DTI imaging can be found on the ADNI webpage (http://adni.loni.usc.edu/data-samples/data-types/). This ADNI protocol was selected after a detailed comparison of several other DTI protocols to optimize SNR at a fixed scan time (Chen et al., 2018; Jahanshad et al., 2013). Preprocessing of DTI data was performed using the diffusion toolbox in FSL (Version 6.0, FMRIB, Oxford, UK). FSL preprocessing includes (i) correction for eddy currents and head motion, (ii) skull stripping, and (iii) fitting the data to a diffusion tensor model to determine fractional anisotropy (FA) and mean diffusivity (MD). It involved calculating a map of . A single diffusion tensor was fitted to each voxel of the vortex- and EPI-corrected DWI image using the FSL toolbox. Scalar anisotropy maps were obtained from the resulting diffusion tensor eigenvalues λ ₁ , λ ₂ , and λ ₃ . FA, which measures the degree of diffusion anisotropy, was defined in a standard way as follows.

where <λ> is equal to MD or the average diffusion rate in all directions. The resulting images were smoothed with a Gaussian kernel of 5 mm FWHM to improve SNR and ensure a Gaussian distribution of the maps.

2-6. APOE genotype

The APOE genotype for each subject was also obtained from the ADNI website. APOE genotype is known to affect the risk of developing sporadic Alzheimer's disease in carriers. Each subject's APOE genotype was recorded as a pair of numbers indicating which two alleles were present in the blood. This genetic trait was a single categorical variable for each participant and could take on one of five possible values: (ε2, ε3), (ε2, ε4), (ε3, ε3), (ε3, ε4), and ( ε4, ε4). The most common allele is APOE ε3, but carriers of the APOE-ε4 variant have an increased risk of developing Alzheimer's disease, while the APOE ε2 variant protects carriers (Michaelson, 2014). Among the three genetic polymorphisms, APOE shows the highest correlation with MCI status and stability (Brainerd et al., 2013). Several recent studies have linked APOE-ε4 status to a relatively late risk of preclinical progression to Alzheimer's disease. Until recently, ε4 variants (including ε4/ε4 and ε4/ε3 combinations) were inconsistently linked to MCI status, likely reflecting both clinical and methodological differences in classification of the condition. In the present invention, genotype data was obtained from a 10 ml blood sample taken at the time of scanning and immediately sent to the AD Biomarker Fluid Bank Laboratory at the University of Pennsylvania for analysis.

<Example 3> Three Feature Extraction Processes

Figure 1 shows a block diagram of the proposed framework. In the present invention, three types of analyzes were performed to extract features from each imaging modality.

(1) Whole brain segmentation (atlas-based segmentation) is a quantitative method that provides a non-invasive way to measure brain regions through neuroimaging. It works by assigning tissue labels to unlabeled images using sMRI scans as well as corresponding manual segmentation.

- For sMRI, FDG and AV45-PET scans, a 2mm Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA) template image (Joliot et al., 2015) was used to extract 384 regions of interest (ROIs) from each.

- For rs-fMRI and DTI scans, we used the 2mm Craddock atlas template (Craddock et al., 2012) to extract 200 ROIs from each rs-fMRI image and 2mm Johns Hopkins University.

- Extract 200 ROIs from each rs-fMRI image and the 2mm Johns Hopkins University (JHU) WMlabels atlas to extract 50 ROIs from each DTI image.

(2) Voxel-wise analysis or morhpometry (voxel-based morphometry, VBM) is a computerized study of neuroanatomy that measures local density differences in brain tissue through voxel-wise comparison of multiple brain images. This is the approach used. Statistics can be used to recognize differences in brain structure between patient groups, which in turn can be used to infer the presence of atrophy or normal tissue expansion in patients with the disease.

- sMRI, FDG and AV45-PET scans apply VBM using SPM12 toolbox

- rs-fMRI scans apply VBM using DPRASF toolbox integrated with SPM12

- DTI scans use the TBSS (tract-based spatial statistics) feature of FSL (v.6.0)

(3) Graph theory methods are a robust method for quantifying the organization of networks through brain anatomical features, including cortical thickness, gray matter volume, and white matter tracks between gray matter regions. When applied to different neuroimaging modalities, graph theory suggests that the result, represented by brain voxels or regions defined by a predetermined compartmental structure, is a network with vertices or nodes, and edges are individual connections between regions, estimated by the strength of the correlation between the estimated regions. It is expressed as a data relationship. The edges of a structural network are not always represented by correlations between regional capacities, but rather by the density or number of white matter track-connected areas. In the analysis of sMRI, FDG and AV45-PET networks, nodes are usually defined by anatomical subdivisions or dividing the brain into multiple regions.

- Used a 2mm AICHA atlas template image (already segmented into 384 individual regions) to extract 384 ROIs from each sMRI, FDG and AV45-PET image

- For functional networks, a 2mm Craddock atlas template image was used to extract 200 ROIs, and for rs-fMRI and DTI images, a 2mm JHU-WM (ICBM-DTI-81) label atlas was used.

- Extract 50 ROIs from each DTI image

- After the nodes of the network are defined, edges representing relationships between different regions are calculated

- Uses the BRAPH toolbox (Mijalkov et al., 2017) integrated into MATLAB 2019a

<Example 4> Feature Extraction Using Atlas-Based Segmentation

(1) After completing a series of image preprocessing steps for each imaging method, as shown in Figure 1, features are extracted from each method. The pink dotted line in Figure 1 shows feature extraction for whole brain analysis.

- For sMRI, FDG and AV45-PET images, a 2 mm AICHA atlas template image was used to extract 384 ROIs from each image (Gupta et al., 2019a).

- Then process the images using the open source NiftyReg toolbox (Modat et al., 2010), a registration tool that performs fast diffeomorphic non-rigid registration on images.

After the registration procedure, target labeled images were obtained based on the template with 384 segments.

(2) For each of the 384 ROIs in labeled MR and PET images, calculate gray matter volume and relative cerebral metabolic rate of glucose from baseline MRI, FDG, and AV45-PET data, respectively, and later use these as features.

- Acquire 384 features for each sMRI and PET image

- For rs-fMRI images, run the pyClusterROI Python script downloaded from https://ccraddock.github.io/cluster_roi/ to extract 200 ROIs from each rs-fMRI image.

This method uses a spatially constrained and normalized truncated spectral clustering algorithm to divide individual and group levels to create compartments. Spatial constraints are hatched to ensure that voxels in the resulting ROI are connected so that the ROI is spatially consistent.

- For DTI images, we run a pipeline to analyze brain diffusion images (PANDA toolbox), available for download at https://www.nitrc.org/projects/panda/, integrated with MATLAB R2019a on the Ubuntu 18.04 operating system.

- The PANDA pipeline uses the FMRIB Software Library (FSL), the Pipeline System for Octave and MATLAB (PSOM), the Diffusion Toolkit, and the MRIcron package to extract diffusion metrics (e.g. FA and MD) amenable to statistical analysis at the voxel level or atlas level.

- For segmentation of DTI images, we used the 2 mm JHUWM (ICBM-DTI-81) label atlas already segmented into 50 individual ROIs.

After completing this process, we obtain a target labeled image based on a template with 50 distinct regions.

<Example 5> Feature Extraction Using Voxel-Wise Morphometry

The red dotted line in Figure 1 shows the pipeline for voxel-wise analysis (or feature extraction) from sMRI, FDG, AV45-PET, rs-fMRI, and DTI images.

- For voxel-wise analysis of sMRI, FDG and AV45 images we used the Computational Anatomy Toolbox (CAT version 12) and the Integrated Statistical Mapping Methods (SPM12) toolbox.

- https://www.fil.ion.ucl.ac.uk/spm/software/spm12/

- http://www.neuro.uni-jena.de/cat/

(1) First, the MRI data were anatomically normalized using a 12-parameter affine transformation provided by the SPM template to compensate for differences in brain size. I selected the East Asian brain image template and left all other elements at the default settings.

(2) Second, the sMRI images were segmented into gray matter, white matter and CSF images using an integrated tissue segmentation method after image intensity inhomogeneity correction was completed. The obtained linearly transformed and segmented images were nonlinearly distorted using the diffeomorphic anatomical registration via exponentiated lie algebra (DARTEL) method, modulated to generate an improved template for the DARTEL-based MNI152 template image, and then smoothed using an 8 mm FWHM kernel. do. In the case of the DARTEL-based MNI152 template image, it is modulated and smoothed using an 8mm FWHM kernel to create an improved template.

(3) Finally, it consists of voxelwise statistical evaluation. To construct a statistical parameter map, contrast values are calculated based on the general linear model estimated regression parameters. This technique runs a two-sample t-test to determine whether there are major regional density differences between two sets of gray matter images.

After running false discovery rate (FDR) and family wise error rate (FWER) corrections, local information values representing the main density differences between the two GM image sets were obtained. Based on this data, cluster values were obtained to create an ROI binary mask, which was then used to obtain gray matter volume from gray matter brain images for use as morphometric features. Additionally, before submitting the FDG and AV45 images to VBM analysis, the first step was to register the FDG and AV45 images with the corresponding sMRI images. The SPM12 toolbox was used for this registration. After enrollment, the same method used for sMRI VBM analysis was used. SPM12 was used to generate statistical maps of between-group changes at the regional-whole-brain level in cerebral metabolic rate for glucose consumption (CMRgl) in AD, MCI, MCIc and HC groups.

For rs-fMRI images, the selected data must be preprocessed to reduce the impact of SNR. In this study, we used the DPARSF toolbox integrated with MATLAB R2019a to calculate whole brain ALFF, fractional (fALFF) and REHO feature maps, as shown in Fig. 2. Spatial smoothing was performed with a 4X4X4mm FWHM Gaussian kernel before ALFF, FALFF and REHO calculations. To minimize low-frequency drift, linear trending is eliminated from the process.

To investigate the differences in ALFF, fALFF, and REHO between the AD patient group and other groups, Random-effects two samplet-test was performed on individual ALFF, fALFF, and REHO maps, considering patient's age as a confounding covariate, voxel-wise as follows: Implement. ALFF is calculated by filtering the time course of each individual voxel of the subject with a fast Fourier transform in the frequency domain, and then obtains the power range or spectrum. Since the power at a given frequency is relative to the square of the amplitude of this frequency, the square root is calculated in each frequency domain of the power range and then the root mean square is obtained over 0.01-0.08 Hz in each individual voxel. Later, this mean square root is used as the ALFF index. fALFF is calculated as the ratio of the amplitude within the low-frequency spectrum (0.01-0.08 Hz) to the total amplitude over the entire frequency spectrum (0-0.25 Hz). It is usually calculated as the fraction of the sum of the amplitudes over the entire detectable frequency range of a given signal. REHO is a voxel-based brain activity measure that assesses the similarity or synchronization between the time series of a given voxel and its nearest neighbors (Zang et al., 2004). This value is calculated using the time series and Kendall's coefficient of agreement for every 27 adjacent voxels. A Kendall's coefficient of agreement value (ranging from 0 to 1) is then assigned to each voxel centroid. In REHO maps, voxels with higher intensity show greater similarity to the time series of neighboring voxels.

For DTI images, FA maps were created using the DTIfit approach and input into the TBSS environment to examine changes in diffusion measurements along the WMtract. First, we non-linearly aligned all FA data to a common space (FMRIB58_FA), then averaged the normalized FA images to generate an average FA image, and set a threshold (FA>0.2) (excluding voxels that were mainly gray matter or CSF) to determine the average FA scaffold was generated. Next, we projected each participant's FA data onto the average FA skeleton and then performed voxel-wise statistical analysis. We perform voxel-based statistical analysis of FAs in the white matter skeleton using Randomize, FSL's non-parametric permutation interference tool. Multiple comparisons were corrected using threshold-free cluster enhancement (p<0.05). White matter regions are identified with the JHU-WM (ICBM-DTI-81) label atlas included in the FSL.

<Example 6> Graph Generation and Construction of sMRI, FDG-PET, AV45-PET, fMRI, and DTI Brain Networks

To evaluate sMRI, FDG-PET, and AV45-PET network topologies in AD, MCI, MCIc, and HC subjects, T1 images of these subjects are integrated and preprocessed into a 2mm-AICHA-atlas template image using the NiftyReg toolbox. A total of 384 regions are extracted from each image and included as nodes in the network study. Edges between these brain regions were computed with Pearson correlations, with negative correlations set to 0. Connectivity analysis of networks is performed on binary undirected graphs, controlling the number of networks over a density range of 5–25% with a step size of 0.5%.

- Use the DPARSF toolbox integrated in MATLAB 2019a to evaluate the functional network topology of rs-fMRI images (we also follow the same method for feature extraction as we followed for rs-fMRI images (“Feature Extraction Using Voxel-Wise Morphometry "See section))

- Regional time series of 200 brain regions included in the Craddock template atlas are extracted for each patient (using Pearson coefficients to calculate relationships between regions and performing network analysis on binary undirected graphs)

- To evaluate the DTI network topology in AD, MCI, MCIc and HC targets, DWI images of the targets are preprocessed using the PANDA toolbox integrated in MATLAB 2019b.

- The PANDA toolbox uses a 2mm JHU-WM (ICBMDTI-81) labeled atlas template image to extract 50 WM regions from each DTI image (the extracted regions are later included as nodes in the network analysis).

- Perform the same procedure using the same parameters as described above to construct a network for sMRI brain images.

- Additionally, several graph metrics are computed to quantify the structure and structure of nodes or global topological organization of feature networks, including local efficiency, characteristic path length, transitivity, and modularity.

- Local efficiency is a measure of the average efficiency of data transmission within a local subgraph or region, defined as the inverse of the shortest average path length between a given node and all other nodes.

The local efficiency of node i is defined as follows.

Here, d _i represents the number of nodes in the subgraph G _i , and l _i,j is the shortest path length between nodes i and j. The distance between two vertices in a graph is the length of the shortest path between them. Otherwise, the distance is infinite and the average of the shortest path between one node and all remaining nodes is called the characteristic path length.

The characteristic path length l _G is calculated as follows.

Here, n is the number of vertices (v) in the graph network and d(v _i ,v _j ) represents the shortest distance between vertices v _i and v _j . The transition rate (or clustering coefficient) is defined as the ratio of paths that intersect two edges to the number of triangles. Additionally, if a node is connected to a second node and then to a third node, transitivity reproduces the probability that the initial node is connected to the third node. It can be calculated as follows:

Here a _ij is the (i,j) entry of the binary connectivity matrix. If there is a link between nodes i and j, a _ij =a _ji =1, otherwise a _ij =a _ji =0. Since the network has no loops of its own, a _ij =0. Modularity is the proportion of network edges belonging to a given group minus the proportion expected if the edges were randomly distributed. It also calculates the extent to which the network can be divided into communities. It can be calculated as follows:

Here, the network is completely divided into M non-overlapping modules (or clusters), and q _ij represents the ratio of all links connecting nodes in module i and nodes in module j.

<Example 7> Classification Techniques

In supervised learning, classification is similar to the task of determining the category to which a new sample belongs based on a data training set containing already identified instances with associated connections. In neuroimaging, different sources of information include different imaging modalities (sMRI, FDG, AV45, DTI, and rs-fMRI), different ways of extracting features from the same modality or from different feature subsets (e.g. ROI-based for all images). or voxel-based) may be included.

In the present invention, we applied three different methods to extract features from the same neuroimaging modality: whole brain sections, voxel-wise analysis, and graphical representation. We were primarily interested in the first two approaches, which use feature subsets as kernels for each of the two methods. You combine (or chain) the approaches later. We were particularly interested in examining models based on subsets of features extracted by voxel or anatomical criteria to achieve predictions that can estimate anatomical location.

7-1. Multiple Kernel Learning

Kernel methods such as SVM (Cortes and Vapnik, 1995; Samper-Gonzalez et al., 2018), which are based on similarity measures between data points, have achieved great success in dimensionality reduction and classification. Kernelization projects the basic data space into a higher-dimensional feature space. Nonlinear relationships between variables in the original space become linear in the transformed space. [Equation 5] is the training sample, where [Equation 6] is the data sample, M is the number of features of all forms, and [Equation 7] is the corresponding class label. The goal is to simultaneously obtain optimal feature description and maximum margin classifier in kernel space due to a systematic and elegant way to form complex patterns. Therefore, we use kernel tricks to project the data.

As is known, for the kernel (K) of the input space (Χ) there exists a Hilbert space (f), called feature space, whose projection Ø is given by the mapping ϕ:X → f. K=(x,y)=<ϕ(x),ϕ(y)> with 2-object(x,y) of Examples of kernel functions include linear, RBF, etc. A recent paper shows that using multiple kernels instead of a single kernel improves the interpretability of the decision function and in some cases improves the final performance. In MKL, data is represented as a combination of base kernels. Each basic kernel represents a different form or characteristic of an entity. MKL attempts to find the optimal combination of underlying kernels so that the next analysis task can benefit the most. Classification tasks are particularly well represented by MKL. The optimal combination is the one that provides maximum classification accuracy. The dual form of MKL optimization solved by existing solvers such as LIBSVM is given by [Equation 8].

Here, α _i , α _j are the Lagrange multipliers, which are variables obtained when converting the raw support vectors into a dual problem, and k ^m (x ^m _i , x ^m _j ) is the mth applied to each pair of samples. is the kernel function, and C is a regularization parameter that controls the distance between the hyperplane and the support vector. In a set of n training samples, the feature of the ith sample of the mth modality is in the vector x ^m _i and its class label is +1 or -1. The weights of the mth modality kernel, denoted by β _m , are separated into separate optimization problems using grid search or with fixed α. For each new test sample s, a kernel function is computed over the training samples. The MKL overview is shown in Figure 3. Recent studies have shown that embedding the underlying dataset in two or more kernels with different kernel parameter choices improves performance. Recent studies have shown that including the underlying data set in more than one kernel, each with different kernel parameter choices, improves performance. Then, all kernels are normalized to unit traces through [Equation 11].

Here, weights are set for sMRI, FDG, AV45, DTI, rs-fMRI, and APOE genotype features. The combined kernel can be described as follows.

Then, the EasyMKL (Aiolli and Donini, 2015; Donini et al., 2019) solver is used to calculate optimal weights to search for the underlying kernel combination that maximizes classifier performance by optimizing the simple quadratic problem solved by SVM. Besides its proven empirical success, a clear advantage of EasyMKL over other MKL approaches is its high scalability with respect to the number of kernels to be combined. Find the coefficient η that maximizes the edges in the training data set. Here, the margin is calculated as the distance between the convex hull of the positive and negative samples. In particular, the general problems that EasyMKL seeks to optimize are as follows.

where y is the diagonal matrix where the training samples are on the diagonal, λ is the regularization hyperparameter, while the domain Γ represents the two probability distributions over the negative and positive sample sets of the training set, i.e. Equation 11 (Aiolli and Donini, 2015; Donini et al., 2019).

Every element is similar to a pair of samples in the convex hull of the positive and negative training samples. The first expression of the objective function in the solution represents the obtained (squared) edges. That is, it is the (squared) distance between the points in the convex hull of positive samples and the points in the convex hull of negative samples in feature space. Enforces sparsity across modalities, allowing more than one distinct kernel to be selected from the same modality. In other words, it becomes a convex optimization problem because there is sparsity and non-sparsity for forms within forms. Additionally, in addition to the EasyMKL classifier, the RBF kernel (or Gaussian kernel) was applied together with the SVM classifier to compare the obtained results. The RBF kernel is a function whose score varies depending on the distance from the origin (or some point). The above is expressed as [Equation 12], where ∥x ₁ - x ₂ ∥ ² is the squared Euclidean distance between the two data points x ₁ and x ₂ .

The RBF kernel has two parameters, gamma (γ) and C, and the performance depends on them. When the C value is small, the classifier is ok with misclassified input points (high bias, low variance), but when the C value is high, the classifier is heavily penalized for misclassified data and therefore has poor quality control over misclassified input points (low bias, high variance). ) leans backwards to avoid Additionally, when the value is low, the curve of the decision margin is quite low, resulting in a very large decision area, but when the value is high, the curve of the decision margin is high, creating a decision boundary bar around the input point. In the case of the present invention, the GridSearch method of the scikit-learn (v0.20) (Pedregosa et al., 2011) library was applied to find the optimal hyperparameter (C and gamma) values for the RBF-SVM classifier. GridSearch was performed in the range C = 1 to 9 and γ=le ^-4 to 7.

7-2. Cross validation (CV)

Cross-validation (CV) is one of the most widely used data resampling methods to estimate generalization knowledge in predictive designs and prevent under- or overfitting. CV is mainly applied in settings where the purpose is prediction and there is a need to evaluate the accuracy of the prediction model. For classification problems, a model is usually fitted with a set of known samples, called training samples, and a set of unknown samples against which the model is tested, called test samples. The goal is to have samples to test the proposed model during the training period and to provide insight into how a particular model adapts to independent samples. The CV round involves splitting the sample into complementary subsets and then performing analysis on the individual subsets. Afterwards, the study is validated on another subset (called the test sample). To reduce variability, we perform several rounds of CV using several different partitions and later take the average of the results. CV is a powerful procedure for evaluating model performance. In the present invention, LOOCV (leave-one-out CV) of the scikit-learn (v0.20) (Pedregosa et al., 2011) library was used. LOOCV is a CV process where the majority of the folds are "1" and "k" is fixed to the number of attributes in the data set. This means that the number of folds is equal to the number of instances in the sample. Therefore, the learning algorithm is used once for each instance, using all other instances as training samples and the selected instance as a single-item test sample. This type of CV is useful when the size of the training sample is limited and the number of attributes to be tested is not large.

7-3. Implementation

Our classification framework and validation experiments use an interface to the scikit-learn v0.20 (Pedregosa et al., 2011) library to measure performance and MKLpy (v0.5, https://github) for the MKL framework. It was conducted in Python 3.5 using com/IvanoLauriola/MKLpy). The main source code is available on the GitHub website (https://github.com/Alzheimer1/Classification-of-Alzheimer-s-disease). A list of data sets is available in the supplementary section. However, the original image features must be prepared separately.

<Result>

This section presents each classification (AD vs. HC, MCI vs. MCIc, AD vs. MCI, AD vs. MCIc, HC vs. MCI, AD vs. HC, MCI vs. MCIc, AD vs. We present performance results for HC vs. MCIc).

We implemented a multimodal fusion approach for integration of sMRI, FDG, AV45, rs-fMRI, DTI and APOE genotype data. Combined expressions were used to distinguish AD patients from healthy subjects.

The combined approach is considered successful if the classification task performs with higher accuracy, higher AUC score, higher precision, better sensitivity, and specificity for unimodal classification. Along with the unimodal approach, we evaluated the classification of a concatenated data vector containing data from five neuroimaging modalities and two APOE genotyping modalities and used this as a baseline. Additionally, after completing feature extraction from each modality using whole-brain segmentation and voxel-wise analysis, the extracted features are passed through a polynomial kernel function to map the original nonlinear low-dimensional features into a separable higher-dimensional space.

Afterwards, data fusion technology was used to combine multiple kernel features into a single format and then passed through the EasyMKL classifier for binary classification of six groups. For unbiased performance evaluation, the classification group was randomly split into two sets with a 70:30 ratio as training and testing sets, respectively. It is very difficult to find a suitable value for the lambda (λ) parameter in the training set, and its value affects the classification results. Therefore, we used the leaveone-out cross-validation technique on the training set to find the optimal hyperparameter values for lambda from 0 to 1 of the EasyMKL algorithm.

For each method, an EasyMKL classifier was trained using the training group using the optimized values obtained for the hyperparameters. The performance of the resulting classifier was estimated on the remaining 30% data of the test dataset that was not used in the training phase.

<Test Example 1> Classification Performance Across Single and Combined Modalities Using Whole-Brain Parcelation Analysis

- For whole brain analysis, 384 ROIs were extracted from each neuroimaging modality using 2 mm AICHA atlas template images for sMRI, FDG and AV45-PET images with the NiftyReg toolbox (see Figure 1B).

- For each rs-fMRI and DTI image, a 2mm Craddock atlas template and a 2mm JHU-WM (ICBM-DTI-81) label atlas were used to extract 200 and 50 ROIs from each rs-fMRI and DTI image. Use the pyClusterROI Python script and the PANDA toolbox, respectively (see Figure 1B).

- In total, 1404 features were obtained for a single image, 384 features from each sMRI, FDG, and AV45-PET image, 200 features from each rsfMRI image, 50 features from each DTI image, and 2 features from APOE genotype data. Afterwards, the obtained features were passed through a normalization technique to minimize redundancy within the data set. Additionally, low-dimensional normalized features are transferred from the polynomial kernel matrix and mapped to the high-dimensional feature space. All these high-dimensional features were then integrated into one form before passing through the EasyMKL algorithm for classification. The obtained AUC-ROC graph and Cohen's kappa score are shown in Figures 4 and 5.

- For a single modality, whole-brain MKL analysis for AD vs. HC using only the APOE genotype (Figure 6) achieved an accuracy of 85.71%. Similar accuracies were obtained using sMRI (90.48%), FDG-PET (91.5%), AV45-PET (89.39%), and rs-fMRI (92.42%). Using DTI-FA, accuracy increased to 93.17% compared to genotypic and functional imaging. When the combined ROI features passed the classifier, the accuracy increased to 96.05%. Additionally, the obtained Cohen's kappa value was 0.9066, which is closer to 1 than the kappa value of the individual forms. Figure 5 shows Cohen's kappa plot for the combined ROIs.

- Here, for AD vs HC classification, the combined ROI features performed very well compared to single modalities. Similarly, for single modality, whole-brain MKL analysis for MCI vs. MCIc (Figure 6) using APOE genotype alone achieved an accuracy of 85.24%. Similar accuracies were obtained using FDG-PET (86.88%), AV45-PET (89.47%), and rs-fMRI (88.52%). Using sMRI-extracted ROI features, the achieved accuracy was lower (84.71%) compared to other single modalities. Using DTI-FA, the accuracy increased to 91.80% compared to genotype and feature images. Additionally, when the combined ROI features passed the classifier, the accuracy increased to 94.74% and the obtained Cohen's kappa value was 0.8950 (Figure 4), which is closer to 1 compared to the individual modalities. Here, for MCI vs. MCIc classification, the combined ROI features performed very well compared to the single modality. Similarly, for the AD vs. MCIc classification problem, the best performance was obtained using a combination of six feature modalities, achieving an accuracy of 94.89% with a Cohen's kappa of 0.8502. In this case, rs-fMRI and DTI-FA unimodal features performed better in classifying groups (AD vs. MCIc) than other unimodal features, and the obtained accuracies were 92.06 and 93.65%. AD vs. For the MCI group, our proposed technique achieved an accuracy of 93.59%, which is 1.66% higher than the best accuracy obtained from DTI-FA (unimodal) features to classify this group. The Cohen's kappa score obtained for the AD versus MCI group was 0.8562 (Figure 5), which is close to the maximum agreement of 1. For the HC vs. MCIc classification problem, the proposed method combining all six modalities of biomarkers to distinguish between HC and MCIc achieved good results compared to single-modality biomarkers. For this classification problem, the proposed method achieved an accuracy of 94.24% with Cohen's kappa of 0.8814 (Figure 5). In this case, it can be seen from Figure 6 that the three (FDG-PET, AV45-PET, rs-fMRI) functional imaging features performed better than other unimodal features, and the obtained Cohen's kappa scores were 0.7610, 0.7981, and 0.8129, all of which were 1. close to Similarly, for the HC vs. MCI group, the accuracy of the proposed technique is 95.55%, which is 1.62% higher than the highest accuracy obtained from DTIFA (unimodal) features for classifying this group. The Cohen's kappa score obtained for the HC versus MCI groups was 0.8697 (Figure 5), which is close to the maximum agreement value of 1. Therefore, from Figures 4 to 6, it can be said that for all classification combinations, the proposed method achieved a high level of performance compared to individual modalities of different biomarkers of 1 to 3%, and the proposed scheme also achieved it. It shows a higher level of agreement between the six classification combinations than individual modality-based methods.

- The number of ROIs extracted was slightly higher for sMRI, FDG-PET, and AV45-PET images compared to other modalities, but the actual number of features used as input for each single and combined model varied across models. HC vs. ROIs extracted for sMRI, FDG-PET, AV45-PET, and rsfMRI in the classification set (AD vs. HC, MCIs vs. MCIc, AD vs. MCIc, AD vs. MCIs, HC vs. MCIs) excluding the MCIc group. The number of features was higher than the number of DTI-FA ROI features, and the obtained accuracy was lower than that of DTI-FA ROI features for the specified classification group. Among 1404 ROIs, 384 ROIs were selected from sMRI, 384 ROIs were selected from FDG-PET, 384 ROIs were selected from AV45-PET, 200 ROIs were from rs-fMRI, 50 ROIs were from DTI-FA, the remaining 2 were from APOE genotype. Among the two ROIs, they correspond to 27.3% (sMRI, FDG-PET, and AV45-PET respectively), 14.5% (rs-fMRI), 3.5% (DTI-FA), and 0.1% (APOE) of the total number of features. Figure 4 shows the ROC curves (plots of the TPR vs. the FPR for dissimilar possible cutpoints) for each study presented in Figure 6. The obtained AUC is displayed in each plot. Figure 4 shows that the proposed method achieved higher AUC values for all classification sets than for individual modalities. For the AD vs. HC and HC vs. MCI classification groups, the proposed method achieved an AUC of over 95%, while for the MCI vs. MCIc, AD vs. MCI, AD vs. MCIc, and HC vs. MCIc groups, the proposed method achieved an AUC of less than 95%. AUCs were achieved (MCI vs. MCIc, AD vs. MCI, AD vs. MCIc, and HC vs. MCIc) <95% <(AD vs. HC and HC vs. MCI).

<Test Example 2> Classification Performance Across Single and Combined Modalities Using Voxel-Wise Analysis

- The SPM12 toolbox integrated with the CAT12 toolbox in MATLAB R2019a was used for voxel-wise analysis of sMRI, FDG-PET, and AV45-PET images. For DTI images, the DTIfit and TBSS functions from the FSL toolbox were used. For voxel-wise analysis of rs-fMRI images, we used the DPARSF toolbox in MATLAB R2019a. Afterwards, these obtained features were propagated with the two features in the APOE genotype data through a normalization technique to minimize redundancy within the data set. Additionally, these low-dimensional normalized features were transferred from the polynomial kernel matrix and mapped to the high-dimensional feature space. All high-dimensional features were then fused into one form before passing through the EasyMKL algorithm for classification. The results obtained are shown in Figure 7, and Figure 8 shows Cohen's kappa plots for all six classification groups using voxel-wise analysis.

- After comparing the results of statistical two-sample t-tests for AD vs. HC groups, statistical values representing the significance level of the groups in the activation map were calculated. This table specifies the main affected regions observed in the AD vs. HC set and the voxel clusters of each group obtained along with detailed information including peak area in MNI space shape, cluster-level p-score, and peak intensity in T-score. To achieve bias change for multiple comparisons, we used an uncorrelation threshold of P _uncorrected ≤0.001, an FDR value of P _FDR =0.05 at the voxel level, and a FWER value of P _FWER =0.05 at the cluster level. ROI binary masks were created from selected clusters of each modality and later gray and white matter volumes were removed from both image sets (AD vs. HC). Figure 9A shows the most important regions in which these groups differ from each other using AV45-PET neuroimages. Likewise, Figure 10A shows the most important regions in which these groups differed using DTI-FA neuroimaging.

- For MCI vs. MCIc groups, change bias for multiple comparisons using a no-correlation threshold of P _uncorrected ≤0.001, an FDR value of P _FDR =0.05 at the voxel level, and a FWER value of P _FWER =0.05 at the cluster level. Perform. ROI binary masks were generated from selected clusters of each modality and subsequently gray matter and white matter volumes were removed from both image sets (MCI vs. MCIc). Figure 9B shows the most important regions in which these groups differ from each other using AV45-PET neuroimages. Similarly, Figure 10B uses DTI-FA neuroimaging to show the most important regions where these groups differ. We also follow the same procedure for calculating voxel clusters for the AD vs. MCIc, AD vs. MCI, HC vs. MCIc, and HC vs. MCI groups as we followed for AD vs. HC and MCI vs. MCIc. After completing a series of feature extraction from each individual modality, the obtained features are passed through a polynomial kernel matrix to map these low-dimensional features into a high-dimensional feature space. All these high-dimensional features were then fused into one form before passing through the EasyMKL algorithm for classification. Figure 7 shows the classification results for AD versus HC groups using voxel-wise features. Compared to the single modality results, we show that the combined features performed very well for this group. The combined features achieved 95.55% AUC in classifying AD versus HC groups as shown in Figure 11, with a Cohen's kappa value of 0.9014, which is close to 1 as shown in Figures 7 and 8. The two groups have a good level of agreement between them. Figure 7 shows the classification results of MCI versus MCIc groups using voxel-wise features. It also shows that the combined features performed very well in classifying MCI versus MCIc groups compared to single-mode performance. The combined features achieved an AUC of 94.90% when classifying MCI versus MCIc groups, as shown in Figure 11. Cohen's kappa value was 0.8825, close to 1, as shown in Figures 7 and 8. Also for the AD vs. MCIc and AD vs. MCI classification groups, our proposed system achieved a high level of performance and agreement (0.9145 and 0.8869) compared to individual modality biomarkers.

- AD vs. For the MCIc group, AV45-PET and DTI-FA achieved higher classification accuracy compared to other unimodal biomarkers, but the obtained accuracy was 0.9145 (Figure 8), with Cohen's kappa score of 96.20%, which is the accuracy obtained from the combined VOI process. It was 3% lower than Likewise AD vs. For the MCI group, the proposed method achieved an accuracy of 95.16% with a Cohen's kappa score of 0.8869 (Figure 8). For the HC vs. MCIc classification group, the AV45-PET individual modality biomarker performed very well compared to other single modality biomarkers. The accuracy and Cohen's kappa scores obtained using the AV45-PET biomarker were 94.02% and 0.8481 (Figure 8). Additionally, the combined VOI features were passed through the EasyMKL classifier for classification of HC vs. MCIc groups, and the accuracy increased by 1.5% after applying the classifier. This suggests that the combined VOI features were helpful in classifying this group. Similarly, for the HC versus MCI classification groups, the rs-fMRI features of ALFF and fALFF achieved good levels of performance and agreement (0.8386 and 0.8248) compared to other individual modality biomarkers. In this case, individual characterization of the APOE genotype also performed well compared to the FDG-PET imaging modality biomarkers, but the performance of the individual modality biomarkers was not very good compared to the combined VOI results (Figure 7). The combined VOI features achieved an accuracy of 94.43% and an AUC of 94.67% with a Cohen's kappa score of 0.8864 in classifying HC versus MCI groups. Figure 11 shows the ROC curves for all six classification groups.

The solid red lines across all classification tasks represent the combined VOI features for that particular group. As a result, it can be seen from Figures 7 and 8, 11 that for all classification combinations, the proposed method achieved a higher level of performance of 1-3% compared to single-modality biomarkers, and the proposed scheme also achieved 6% better performance compared to single-modality-based methods. There is a high level of agreement between each classification combination. Figure 7 clearly shows the advantage of using combined features over single features.

- To compare different most important regions in AD vs HC subjects, (left/right) temporal-mid, (left/right) frontal-sup, (left/right) occipital-mid, (left/right) occipital- The most different are inf, (left/right) temporalsup, (left/right) fusiform, (left/right) hippocampus, (left/right) temporal-inf, (left/right) precentral, and (left/right) sagittal stratum. It was confirmed that it was an area.

- Similarly, for the MCI vs. MCIc group, (left/right) precentral, (left/right) precuneus, (left/right) frontal-mid, (left/right) cingulum-mid, (left/right) temporal-inf , (left/right) temporal-sup, (left/right) frontal-dup-medial, (left/right) cerebellum-9, (left/right) thalamus, and (left/right) fusiform are the most different regions. I was able to confirm.

- For AD vs. MCIc group, (left/right) frontal-inf-tri, (left/right) frontal-inf-oper, (left/right) frontal-inf-orb, (right) hippocampus, (left/right) ) It was confirmed that the precentral, (left/right) thalamus, (left) pallidum, (left/right) lingual, and (left/right) inferior longitudinal fasciculus were the most different regions.

- For AD vs. MCIs group, (left/right) precentral, (left/right) frontal-mid, (left/right) hippocampus, (left/right) temporal-inf, (left/right) frontal-inf-orb , (left/right) occipitalmid, (right) posterior corona radiate, and (left/right) precuneus were confirmed to be the most different regions.

- For HC vs. MCIc and HC vs. MCI groups ((left/right) precentral, (left/right) cerebellum-6, (left/right) precuneus, (left/right) frontal-mid, (left/right) corticospinal It was confirmed that the most different regions were the tract, (left/right) lingual, (right) amygdala, and the (left/right) occipital-sup. When calculating voxel clusters, the (left/right) precentral region was used in all classification problems. It is interesting that this was discovered.

- Figure 12 shows the combined (VOI + ROI) classification results for all six classification groups after linking features extracted from whole brain and voxel-wise with APOE genotype. This revealed the importance of all features for classification by mapping low-dimensional features to a high-dimensional feature space by applying a polynomial kernel matrix before passing the features to the EasyMKL classifier. These high-level features were fused into one form. The features were then passed to the MKL algorithm for classification.

- For the AD vs HC classification group, the Combined-(VOI + ROI) feature performed very well compared to the Combined-VOI and Combined-ROI features. AUC and Cohen's kappa scores increased by 2-2.5% (97.78%, 0.9456) compared to combined ROI and combined VOI (Figures 12, 13). Additionally, for the MCI vs. MCIc classification group, Figures 12 and 13 show that the feature of Combined-(VOI + ROI) achieved a very high AUC of 96.94% (2% increase) compared to the other Combined-VOI and Combined-ROI features, respectively. . The obtained Cohen's kappa value (0.9247) was also higher than that of Combined-VOI and Combined-ROI. Also AD vs. MCI, HC vs. MCI, HC vs. For the MCIc classification group, our proposed system performed very well when compared to the results obtained with Combined-VOI and Combined-ROI for each group. In Figure 12, it can be seen that all measured outcomes increased by 2-3% for the three groups (AD vs. MCI, HC vs. MCI, and HC vs. MCIc). The AUC scores obtained for these three groups are 96.25%, 96.59%, and 96.67%. Likewise AD vs. For the MCIc group, the Combined-ROI method performed very well compared to the Combined-(VOI + ROI) and Combined-VOI methods. The measured outcome difference was not very high between the Combined-ROI and Combined-(VOI C ROI) methods (only 1%) for the AD vs. MCIc groups. Figure 12 clearly shows the advantage of using the Combined-(VOI + ROI) feature over the features of Combined-ROI and Combined-VOI.

<Test Example 3> Graph Network Construction and Analysis for All Six Classification Groups

- We used the BRAPH toolbox integrated in MATLAB 2019a for graph analysis or graph network construction. We also performed structural and functional graph theory for all six binary classification groups (AD vs. HC, MCI vs. MCIc, AD vs. MCIc, AD vs. MCI, HC vs. MCIc, and HC vs. MCI). Nodal measurements were taken and comparisons were made using binary undirected graphs, and the measurements were made for a set of network densities (ranging from 5% to 25%), representing the ratio between the number of connections in the network and the number of possible connections, with a step size of 0.5%. evaluated. Several graph metrics have been computed to quantify the node or global topological organization of structural and functional networks, including local efficiency, characteristic path length, transitivity, and modularity. For all six binary classification groups, a non-parametric permutation test sample with 1000 permutations each was performed to assess differences between groups, which were significant against a two-tailed test of the null hypothesis at p < 0.05. A structural correlation matrix graph of AD, HC, MCI, and MCIc of sMRI subjects is shown in Figure 14. All groups showed strong correlations between bilateral homologous regions. The plots in Figures 15 and 16 show the lower and upper bounds (dark red spheres) of the 95% confidence interval (CI) (dark gray shading) as a function of density. Blue, green, pink and purple spheres show differences between sets, beyond the CI indicating that the change was statistically significant at p < 0.05.

- The small dark red dot in the middle (value near 0) specifies the average value of the global network measure between the random sets after the permutation test.

- We also compared the levels of nodes for all six binary classification groups using all modalities. The FDR correction value was kept constant at 0.05 for all six binary classification groups. Regarding the global network topology shown for AD vs. HC in Figure 15, we found longer characteristic path lengths using only FDG (starting at 0.15) than other neuroimaging modalities, with path lengths exceeding the average of the difference values. did. For local efficiency, we found that the AV45 form was the only mode that started above the mean value (at 0.05). We also compared the modularity graphs of the AD and HC groups and found that sMRI showed the largest difference among all modalities (modularity starting at 0.14), while network density was nearly constant up to 25% (network topology was widespread). The rs-fMRI modality performed very well on transition graphs compared to other modalities, despite a decrease in some network densities, starting at 0.06 and increasing up to 25%. We also calculated regional or node network topologies for the AD versus HC groups shown in Figure 17A,B. Figures 17A,B show that AV45 and FDG-PET modalities are the only neuroimages showing the number of significant regional changes in node network topology for AD vs. HC groups. Nodule severity was right g-frontal-sup-1, left g-cuneus-2, left g-frontal-sup-3, left g-frontal-sup-1, left g-frontal-med-orb-1, and right A significant increase was seen in the s-precentral-3 region. Similarly, for the MCI vs. MCIc group, we found an increase in characteristic path length and local efficiency using FDG (starting at 0.01 in both cases, as shown in Figure 15). For modularization, sMRI achieved a network density of 5% but then decreased to 25%. At the same time, the modularity of DTI neural images increased from 10% to 25% in network density. Likewise, for transitivity, compared to the other modalities that started below the mean value, the AV45 modality was the only modality that started at 0 (difference) and showed network increases at every single network density.

- This transitivity plot shows the most prevalent topological changes for the MCI versus MCIc groups. Figure 17C,D shows that AV45-FDG is the most important modality showing the number of significant area changes in node topology for MCI vs. MCIc groups. Left s-postcentral-2, right g-parietal-inf-1, left g-frontal-inf-tri, right g-lingual-2, right g-parahippocampus-5, left n-thalamus-6, left s-parietooccipital -4, left-lingual-3, and left g-cingulum-post-2 are the most important regions represented by the nodal degrees of MCI vs. MCIc classification groups. rs-fMRI is the only modality that shows increases in characteristic path length measurements. It starts at 0.38 (difference) and is inside the CI and above the mean value, but at 15-17% network density, some networks are outside the CI upper limit, and again at 18% density there are networks that are inside the CI until 25% density is reached. . We can also see that at a network density of around 13%, some networks are close to the average value (indicated by small dark red dots). The most prevalent topological changes are shown for the AD versus MCIc groups. The FDG-PET form shows an increase only for local efficiency measures. Modularity and transitivity increase at almost all network densities for sMRI and AV45-PET modalities compared to other single modalities. The FDG-PET format is the only way to show the characteristic path lengths within the CI and above the average value. Comparing local efficiency measures from all modalities, we found that although all modalities were within the CI, the rs-fMRI modality was the only neural image above the mean (indicated in small dark red). Modularity increased and at the same time transitivity decreased in AD vs. MCI groups, as seen in the sMRI modality, but both plots are within the 95% CI. right n-caudate-2, left g-insula-anterior-2, left g-frontal-mid-orb-1, left g-cuneus-2, right g-angular-3, right g-occipital-pole-1, Right n-thalamus-6 was the most significant region indicated by nodal extent for AD vs. MCI classification groups using FDG and sMRI images. In this plot we can see that the characteristic path length starts at 0.8 (difference) at 5% density in the rs-fMRI modality (very high compared to other modalities). However, the density suddenly starts to decrease at 8-13%, crosses the lower limit (indicated by the dark red sphere), and later increases again at a network density of 14%. The rs-fMRI modality is the only neuroimaging where local efficiency, modularity, and transitivity are above the mean values (small dark red spheres). This image is also in the middle of the 95% confidence interval.

<Discussion>

- Our results provided insight into multimodal behavior to classify all six binary classification groups and show which domains or single modalities are most important for future analysis of AD differentials at the clinical level. The idea we propose here is to combine multiple neuroimaging modalities (sMRI, AV45, FDG, rs-fMRI, and DTI) with genetic biomarkers (APOE) for classification of AD patients and other groups in a whole-brain, voxel-wise, and graphical manner. Use analytical methods. The present invention used three types of Atlas (AICHA, pyClusterROI, and JHU-WM) to segment sMRI/PET, rs-fMRI, and DTI neural images. Additionally, to segment sMRI and PET images, we used the AICHA atlas, which was already segmented into 384 ROIs. Similarly, for rs-fMRI neural images, we apply the pyClusterROI Python script with the Craddock atlas to segment the rs-fMRI images into 200 brain regions (this is a well-known example of segmenting fMRI images using a spatially constrained normalized cut spectral clustering process). Because it is a known technology). Additionally, because fMRI images are composed of time series (ADNI rs-fMRI data consists of 140 time series, or time points), segmenting the brain using fMRI data required grouping voxels from similar time points to form regions. This is typically done using data-driven clustering methods where each cluster constitutes one region. For the above reasons, we chose the pyClusterROI script for rs-fMRI images. Additionally, in the case of DTI images, the JHU-WM (ICBM-DTI-81) label atlas, which was already segmented into 50 brain regions, was applied. In the present invention, different atlases were selected for different modalities of neuroimaging because of their respective advantages and to extract higher brain ROIs for specific neuroimaging modalities. After feature extraction was completed in each method, the low-dimensional extracted features were transmitted through a polynomial kernel function and mapped to high-dimensional features. Afterwards, all six high-dimensional features were fused into one form before further analysis. We later passed these unimodal and multimodal features through EasyMKL classifier for classification and reported the average accuracy of each method. This procedure is widely used to compare the performance of machine learning approaches. Although previous studies have already applied multimodal methods to AD classification (Zhang et al., 2011; Young et al., 2013; Liu et al., 2014; Ritter et al., 2015; Schouten et al., 2016; Hojjati et al., 2018; Gupta et al., 2019a) along with other groups, we present six binary classification groups (AD vs. HC, MCI vs. MCIc, AD vs. MCIc, AD vs. MCI, HC vs. MCIc, HC vs. MCI). This is the first study to combine five types of neuroimaging modalities and two APOE genotype scores for classification. Compared with recently published results, the method proposed in this invention clearly showed an improvement using multimodal features in terms of performance compared to single-mode features for classifying six binary groups. Additionally, we study the plots of six binary groups (characteristic path length, local efficiency, modularity and transitivity) and adopt a graph-theoretic strategy (global and node network topology) to find the most abundant regions in the groups.

- Influence of the different types of neuroimaging modality

Cohen's kappa scores obtained for each modality using both ROI and VOI features were compared. Scores were calculated for each of the six classification groups. Cohen's kappa results shown in Figure 5 were calculated using ROI-based features. Similarly, Cohen's kappa results shown in Figure 8 were calculated using VOI-based features. The results shown in Figures 5 and 8 clearly show that the DTI-FA modality biomarker achieved a high level of agreement between groups, classifying the six binary classification groups compared to the other five biomarkers. It can also be said that the VOI-based DTI-FA (above 0.8 to 0.87) features performed slightly better than the ROI-based DTI-FA (above 0.75 to 0.84) features.

- Influence of the type of features (ROI and VOI)

Cohen's kappa scores obtained for region (ROI) features with reference atlases (AICHA, Craddock and JHU-WM) were used to compare voxel (VOI) features (SPM12, DPARSF, TBSS). Scores were evaluated for the same six binary classification groups. Cohen's kappa results shown in Figure 5 are for ROI-based analysis, and similarly, Cohen's kappa results shown in Figure 8 are for VOI-based analysis. Results shown in Figures 5 and 8 showed no noticeable differences in Cohen's kappa scores obtained using regional or voxel features. Both features performed very well and achieved a high level of agreement with each other for all six binary classification groups.

- Influence of the classification method

The results obtained from the EasyMKL classifier were compared with the results obtained from the RBF-SVM classifier. Figures 6, 7, and 12 show the classification results obtained for all six binary groups using EasyMKL and RBF-SVM classifiers. Similarly, Figure 18 shows a plot comparing the accuracy obtained for all six binary classification tasks using EasyMKL and RBF-SVM classifiers. The results shown in Figure 18 show that the EasyMKL classifier achieved high classification accuracy for all six binary groups compared to RBF-SVM. It also suggests that EasyMKL optimizes the simple quadratic problem solved by SVM in a more efficient manner.

<Reference>

Acosta-Cabronero, J., Nestor, P.J., 2014. Diffusion tensor imaging in Alzheimer's disease: insights into the limbic-diencephalic network and methodological considerations. Frontiers in Aging Neuroscience 6. https://doi.org/10.3389/fnagi.2014.00266

Aiolli, F., Donini, M., 2015. EasyMKL: a scalable multiple kernel learning algorithm. Neurocomputing 169, 215-224. https://doi.org/10.1016/j.neucom.2014.11.078

Albert, M.S., 2011. Changes in cognition. Neurobiology of Aging 32, S58-S63. https://doi.org/10.1016/j.neurobiolaging.2011.09.010

Angelucci, F., Spalletta, G., di Iulio, F., Ciaramella, A., Salani, F., Varsi, A.E., Gianni, W., Sancesario, G., Caltagirone, C., Bossu, P., 2010. Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI) Patients are Characterized by Increased BDNF Serum Levels. Current Alzheimer's Research 7, 15-20. https://doi.org/10.2174/156720510790274473

Ashburner, J., Friston, K.J., 2001. Why Voxel-Based Morphometry Should Be Used. NeuroImage 14, 1238-1243. https://doi.org/10.1006/nimg.2001.0961

Bi, Frontiers in Neuroinformatics 12. https://doi.org/10.3389/fninf.2018.00060

Bishop, N.A., Lu, T., Yankner, B.A., 2010. Neural mechanisms of aging and cognitive decline. Nature 464, 529-535. https://doi.org/10.1038/nature08983

Bloudek, L.M., Spackman, D.E., Blankenburg, M., Sullivan, S.D., 2011. Review and Meta-Analysis of Biomarkers and Diagnostic Imaging in Alzheimer's Disease. JAD 26, 627-645. https://doi.org/10.3233/JAD-2011-110458

Brainerd, C.J., Reyna, V.F., Petersen, R.C., Smith, G.E., Kenney, A.E., Gross, C.J., Taub, E.S., Plassman, B.L., Fisher, G.G., 2013. The apolipoprotein E genotype predicts longitudinal transitions to mild cognitive impairment but not to Alzheimer's dementia: Findings from a nationally representative study. Neuropsychology 27, 86-94. https://doi.org/10.1037/a0030855

Camus, V., Payoux, P., Barre, L., Desgranges, B., Voisin, T., Tauber, C., La Joie, R., Tafani, M., Hommet, C., Chetelat, G. , Mondon, K., de La Sayette, V., Cottier, J.P., Beaufils, E., Ribeiro, M.J., Gissot, V., Vierron, E., Vercouillie, J., Vellas, B., Eustache, F. , Guilloteau, D., 2012. Using PET with 18F-AV-45 (florbetapir) to quantify brain amyloid load in a clinical environment. European Journal of Nuclear Medicine and Molecular Imaging 39, 621-631. https://doi.org/10.1007/s00259-011-2021-8

Chen, H.-J., Gao, Y.-Q., Che, C.-H., Lin, H., Ruan, X.-L., 2018. Diffusion Tensor Imaging With Tract-Based Spatial Statistics Reveals White Matter Abnormalities in Patients With Vascular Cognitive Impairment. Frontiers in Neuroanatomy 12. https://doi.org/10.3389/fnana.2018.00053

Cheng, B., Liu, M., Zhang, D., Munsell, B.C., Shen, D., 2015. Domain Transfer Learning for MCI Conversion Prediction. IEEE Transactions on Biomedical Engineering 62, 1805-1817. https://doi.org/10.1109/TBME.2015.2404809

Clem, M.A., Holliday, R.P., Pandya, S., Hynan, L.S., Lacritz, L.H., Woon, F.L., 2017. Predictors That a Diagnosis of Mild Cognitive Impairment Will Remain Stable 3 Years Later: Cognitive And Behavioral Neurology 30, 8- 15. https://doi.org/10.1097/WNN.0000000000000119

Cohen, J., 1960. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement 20, 37-46. https://doi.org/10.1177/001316446002000104

Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach Learn 20, 273-297. https://doi.org/10.1007/BF00994018

Craddock, R.C., James, G.A., Holtzheimer, P.E., Hu, X.P., Mayberg, H.S., 2012. A whole brain fMRI atlas generated via spatially constrained spectral clustering. Human Brain Mapping 33, 1914-1928. https://doi.org/10.1002/hbm.21333

Cui, Y., Liu, B., Luo, S., Zhen, X., Fan, M., Liu, T., Zhu, W., Park, M., Jiang, T., Jin, J.S., the Alzheimer's Disease Neuroimaging Initiative, 2011. Identification of Conversion from Mild Cognitive Impairment to Alzheimer's Disease Using Multivariate Predictors. PLoS ONE 6, e21896. https://doi.org/10.1371/journal.pone.0021896

Cui, Z., Zhong, S., Xu, P., He, Y., Gong, G., 2013. PANDA: a pipeline toolbox for analyzing brain diffusion images. Frontiers in Human Neuroscience 7. https://doi.org/10.3389/fnhum.2013.00042

Cuingnet, R., Gerardin, E., Tessieras, J., Auzias, G., Lehericy, S., Habert, M.-O., Chupin, M., Benali, H., Colliot, O., 2011. Automatic classification of patients with Alzheimer's disease from structural MRI: A comparison of ten methods using the ADNI database. NeuroImage 56, 766-781. https://doi.org/10.1016/j.neuroimage.2010.06.013

Dai, Z., Yan, C., Wang, Z., Wang, J., Xia, M., Li, K., He, Y., 2012. Discriminative analysis of early Alzheimer's disease using multi-modal imaging and multi -level characterization with multi-classifier (M3). NeuroImage 59, 2187-2195. https://doi.org/10.1016/j.neuroimage.2011.10.003

Daianu, M., Jahanshad, N., Nir, TM, Toga, AW, Jack, CR, Weiner, MW, Thompson, for the Alzheimer's Disea, PM, 2013. Breakdown of Brain Connectivity Between Normal Aging and Alzheimer's Disease: A Structural k -Core Network Analysis. Brain Connectivity 3, 407-422. https://doi.org/10.1089/brain.2012.0137

de Vos, F., Koini, M., Schouten, T.M., Seiler, S., van der Grond, J., Lechner, A., Schmidt, R., de Rooij, M., Rombouts, S.A.R.B., 2018. A. Comprehensive analysis of resting state fMRI measures to classify individual patients with Alzheimer's disease. NeuroImage 167, 62-72. https://doi.org/10.1016/j.neuroimage.2017.11.025

Dean, D.C., Jerskey, B.A., Chen, K., Protas, H., Thiyyagura, P., Roontiva, A., O'Muircheartaigh, J., Dirks, H., Waskiewicz, N., Lehman, K.; Siniard, A.L., Turk, M.N., Hua, Onset Alzheimer Disease: A Cross-sectional Imaging Study. JAMA Neurol 71, 11. https://doi.org/10.1001/jamaneurol.2013.4544

Donini, M., Monteiro, J.M., Pontil, M., Hahn, T., Fallgatter, A.J., Shawe-Taylor, J., Mourao-Miranda, J., 2019. Combining heterogeneous data sources for neuroimaging based diagnosis: re- weighting and selecting what is important. NeuroImage 195, 215-231. https://doi.org/10.1016/j.neuroimage.2019.01.053

Dubois, B., Albert, M.L., 2004. Amnestic MCI or prodromal Alzheimer's disease? The Lancet Neurology 3, 246-248. https://doi.org/10.1016/S1474-4422(04)00710-0

Fan, Y., Resnick, S.M., Wu, X., Davatzikos, C., 2008. Structural and functional biomarkers of prodromal Alzheimer's disease: A high-dimensional pattern classification study. NeuroImage 41, 277-285. https://doi.org/10.1016/j.neuroimage.2008.02.043

Fayao Liu, Luping Zhou, Chunhua Shen, Jianping Yin, 2014. Multiple Kernel Learning in the Primal for Multimodal Alzheimer's Disease Classification. IEEE Journal of Biomedical and Health Informatics 18, 984-990. https://doi.org/10.1109/JBHI.2013.2285378

Fransson, P., 2005. Spontaneous low-frequency BOLD signal fluctuations: An fMRI investigation of the resting-state default mode of brain function hypothesis. Human Brain Mapping 26, 15-29. https://doi.org/10.1002/hbm.20113

Ganguli, M., Dodge, H.H., Shen, C., DeKosky, S.T., 2004. Mild cognitive impairment, amnestic type: An epidemiologic study. Neurology 63, 115-121. https://doi.org/10.1212/01.WNL.0000132523.27540.81

Grabner, G., Janke, A.L., Budge, M.M., Smith, D., Pruessner, J., Collins, D.L., 2006. Symmetric Atlasing and Model Based Segmentation: An Application to the Hippocampus in Older Adults, in: Larsen, R ., Nielsen, M., Sporring, J. (Eds.), Medical Image Computing and Computer-Assisted Intervention - MICCAI 2006. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 58-66. https://doi.org/10.1007/11866763_8

Gupta, Y., Lama, R.K., Kwon, G.-R., Alzheimer's Disease Neuroimaging Initiative, 2019a. Prediction and Classification of Alzheimer's Disease Based on Combined Features From Apolipoprotein-E Genotype, Cerebrospinal Fluid, MR, and FDG-PET Imaging Biomarkers. Frontiers in Computational Neuroscience 13. https://doi.org/10.3389/fncom.2019.00072

Gupta, Y., Lee, K.H., Choi, K.Y., Lee, J.J., Kim, B.C., Kwon, G.-R., 2019b. Alzheimer's Disease Diagnosis Based on Cortical and Subcortical Features. Journal of Healthcare Engineering 2019, 1-13. https://doi.org/10.1155/2019/2492719

Gupta, Y., Lee, K.H., Choi, K.Y., Lee, J.J., Kim, B.C., Kwon, G.R., the National Research Center for Dementia, Alzheimer's Disease Neuroimaging Initiative, 2019c. Early diagnosis of Alzheimer's disease using combined features from voxel-based morphometry and cortical, subcortical, and hippocampus regions of MRI T1 brain images. PLoS ONE 14, e0222446. https://doi.org/10.1371/journal.pone.0222446

Hojjati, S.H., Ebrahimzadeh, A., Khazaee, A., Babajani-Feremi, A., 2018. Predicting conversion from MCI to AD by integrating rs-fMRI and structural MRI. Computers in Biology and Medicine 102, 30-39. https://doi.org/10.1016/j.compbiomed.2018.09.004

Hojjati, S.H., Ebrahimzadeh, A., Khazaee, A., Babajani-Feremi, A., 2017. Predicting conversion from MCI to AD using resting-state fMRI, graph theoretical approach and SVM. Journal of Neuroscience Methods 282, 69-80. https://doi.org/10.1016/j.jneumeth.2017.03.006

Hua, K., Zhang, J., Wakana, S., Jiang, H., Li, probability maps in stereotaxic spaces: Analyzes of white matter anatomy and tract-specific quantification. NeuroImage 39, 336-347. https://doi.org/10.1016/j.neuroimage.2007.07.053

Hyman, B.T., Phelps, C.H., Beach, T.G., Bigio, E.H., Cairns, N.J., Carrillo, M.C., Dickson, D.W., Duyckaerts, C., Frosch, M.P., Masliah, E., Mirra, S.S., Nelson, P.T.; Schneider, J.A., Thal, D.R., Thies, B., Trojanowski, J.Q., Vinters, H.V., Montine, T.J., 2012. National Institute on Aging-Alzheimer's Association guidelines for the neuropathologic assessment of Alzheimer's disease. Alzheimer's & Dementia 8, 1-13. https://doi.org/10.1016/j.jalz.2011.10.007

Jack, C.R., Bennett, D.A., Blennow, K., Carrillo, M.C., Dunn, B., Haeberlein, S.B., Holtzman, D.M., Jagust, W., Jessen, F., Karlawish, J., Liu, E.; Molinuevo, J.L., Montine, T., Phelps, C., Rankin, K.P., Rowe, C.C., Scheltens, P., Siemers, E., Snyder, H.M., Sperling, R., Elliott, C., Masliah, E. , Ryan, L., Silverberg, N., 2018. NIA-AA Research Framework: Toward a biological definition of Alzheimer's disease. Alzheimer's & Dementia 14, 535-562. https://doi.org/10.1016/j.jalz.2018.02.018

Jack, C.R., Bennett, D.A., Blennow, K., Carrillo, M.C., Feldman, H.H., Frisoni, G.B., Hampel, H., Jagust, W.J., Johnson, K.A., Knopman, D.S., Petersen, R.C., Scheltens, P. , Sperling, R.A., Dubois, B., 2016. A/T/N: An unbiased descriptive classification scheme for Alzheimer disease biomarkers. Neurology 87, 539-547. https://doi.org/10.1212/WNL.0000000000002923

Jagust, W.J., Landau, S.M., Koeppe, R.A., Reiman, E.M., Chen, K., Mathis, C.A., Price, J.C., Foster, N.L., Wang, A.Y., 2015. The Alzheimer's Disease Neuroimaging Initiative 2 PET Core: 2015. Alzheimer's & Dementia 11, 757-771. https://doi.org/10.1016/j.jalz.2015.05.001

Jahanshad, N., Kochunov, P.V., Sprooten, E., Mandl, R.C., Nichols, T.E., Almasy, L., Blangero, J., Brouwer, R.M., Curran, J.E., de Zubicaray, G.I., Duggirala, R.; Fox, P.T., Hong, L.E., Landman, B.A., Martin, N.G., McMahon, K.L., Medland, S.E., Mitchell, B.D., Olvera, R.L., Peterson, C.P., Starr, J.M., Sussmann, J.E., Toga, A.W., Wardlaw, J.M., Wright, M.J., Hulshoff Pol, H.E., Bastin, M.E., McIntosh, A.M., Deary, I.J., Thompson, P.M., Glahn, D.C., 2013. Multi-site genetic analysis of diffusion images and voxelwise heritability analysis: A pilot project of the ENIGMA-DTI working group. NeuroImage 81, 455-469. https://doi.org/10.1016/j.neuroimage.2013.04.061

Jenkinson, M., Beckmann, C.F., Behrens, T.E.J., Woolrich, M.W., Smith, S.M., 2012. FSL. NeuroImage 62, 782-790. https://doi.org/10.1016/j.neuroimage.2011.09.015

Jicha, G.A., Parisi, J.E., Dickson, D.W., Johnson, K., Cha, R., Ivnik, R.J., Tangalos, E.G., Boeve, B.F., Knopman, D.S., Braak, H., Petersen, R.C., 2006. Neuropathologic. Outcome of Mild Cognitive Impairment Following Progression to Clinical Dementia. ARCH NEUROL 63, 8.

John, M., Ikuta, T., Ferbinteanu, J., 2017. Graph analysis of structural brain networks in Alzheimer's disease: beyond small world properties. Brain Structure and Function 222, 923-942. https://doi.org/10.1007/s00429-016-1255-4

Joliot, M., Jobard, G., Naveau, M., Delcroix, N., Petit, L., Zago, L., Crivello, F., Mellet, E., Mazoyer, B., Tzourio-Mazoyer, N. ., 2015. AICHA: An atlas of intrinsic connectivity of homotopic areas. Journal of Neuroscience Methods 254, 46-59. https://doi.org/10.1016/j.jneumeth.2015.07.013

Khoury, R., Ghossoub, E., 2019. Diagnostic biomarkers of Alzheimer's disease: A state-of-the-art review. Biomarkers in Neuropsychiatry 1, 100005. https://doi.org/10.1016/j.bionps.2019.100005

Liu, C.-C., Kanekiyo, T., Xu, H., Bu, G., 2013. Apolipoprotein E and Alzheimer disease: risk, mechanisms and therapy. Nature Reviews Neurology 9, 106-118. https://doi.org/10.1038/nrneurol.2012.263

Liu, K., Chen, K., Yao, L., Guo, X., 2017. Prediction of Mild Cognitive Impairment Conversion Using a Combination of Independent Component Analysis and the Cox Model. Frontiers in Human Neuroscience 11. https://doi.org/10.3389/fnhum.2017.00033

Long, PLOS ONE 12, e0173372. https://doi.org/10.1371/journal.pone.0173372

Lopez, O.L., Becker, J.T., Chang, Y.-F., Sweet, R.A., DeKosky, S.T., Gach, M.H., Carmichael, O.T., McDade, E., Kuller, L.H., 2012. Incidence of mild cognitive impairment in the Pittsburgh Cardiovascular Health Study-Cognition Study. Neurology 79, 1599-1606. https://doi.org/10.1212/WNL.0b013e31826e25f0

Marquez, F., Yassa, M.A., 2019. Neuroimaging Biomarkers for Alzheimer's Disease. Mol Neurodegeneration 14, 21. https://doi.org/10.1186/s13024-019-0325-5

Michaelson, D.M., 2014. APOE ε4: The most prevalent yet understudied risk factor for Alzheimer's disease. Alzheimer's & Dementia 10, 861-868. https://doi.org/10.1016/j.jalz.2014.06.015

Mijalkov, M., Kakaei, E., Pereira, J.B., Westman, E., Volpe, G., for the Alzheimer's Disease Neuroimaging Initiative, 2017. BRAPH: A graph theory software for the analysis of brain connectivity. PLOS ONE 12, e0178798. https://doi.org/10.1371/journal.pone.0178798

Mitchell, A.J., Shiri-Feshki, M., 2009. Rate of progression of mild cognitive impairment to dementia - meta-analysis of 41 robust inception cohort studies. Acta Psychiatrica Scandinavica 119, 252-265. https://doi.org/10.1111/j.1600-0447.2008.01326.x

Moradi, E., Pepe, A., Gaser, C., Huttunen, H., Tohka, J., 2015. Machine learning framework for early MRI-based Alzheimer's conversion prediction in MCI subjects. NeuroImage 104, 398-412. https://doi.org/10.1016/j.neuroimage.2014.10.002

Morris, J.C., 1993. The Clinical Dementia Rating (CDR): Current version and scoring rules. Neurology 43, 2412-2412. https://doi.org/10.1212/WNL.43.11.2412-a

Mosconi, L., Pupi, A., De Leon, M.J., 2008. Brain Glucose Hypometabolism and Oxidative Stress in Preclinical Alzheimer's Disease. Annals of the New York Academy of Sciences 1147, 180-195. https://doi.org/10.1196/annals.1427.007

Murphy, M.P., LeVine, H., 2010. Alzheimer's Disease and the Amyloid- Peptide. Journal of Alzheimer's Disease 19, 311-323. https://doi.org/10.3233/JAD-2010-1221

Pan, X., Adel, M., Fossati, C., Gaidon, T., Guedj, E., 2019a. Multilevel Feature Representation of FDG-PET Brain Images for Diagnosing Alzheimer's Disease. IEEE Journal of Biomedical and Health Informatics 23, 1499-1506. https://doi.org/10.1109/JBHI.2018.2857217

Pan, Multiscale spatial gradient features for 18F-FDG PET image-guided diagnosis of Alzheimer's disease. Computer Methods and Programs in Biomedicine 180, 105027. https://doi.org/10.1016/j.cmpb.2019.105027

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V.; Vanderplas, J., Passos, A., Cournapeau, D., 2011. Scikit-learn: Machine Learning in Python. MACHINE LEARNING IN PYTHON 6.

az-Parra, A., Kennion, O., Moratal, D., Taylor, J.-P., Kaiser, M., Bauer, R., 2019. Structural connectivity centrality changes mark the path toward Alzheimer's disease. Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring 11, 98-107. https://doi.org/10.1016/j.dadm.2018.12.004Peraza, L.R., D.

az-Parra, A., Kennion, O., Moratal, D., Taylor, J.-P., Kaiser, M., Bauer, R., 2019. Structural connectivity centrality changes mark the path toward Alzheimer's disease. Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring 11, 98-107. https://doi.org/10.1016/j.dadm.2018.12.004

Petersen, R.C., 2004. Mild cognitive impairment as a diagnostic entity. J Intern Med 256, 183-194. https://doi.org/10.1111/j.1365-2796.2004.01388.x

Qiu, C., Kivipelto, M., 2009. Epidemiology of Alzheimer's disease: occurrence, determinants, and strategies toward intervention. Dialogues in Clinical Neuroscience 11, 18.

Rathore, S., Habes, M., Iftikhar, M.A., Shacklett, A., Davatzikos, C., 2017. A review on neuroimaging-based classification studies and associated feature extraction methods for Alzheimer's disease and its prodromal stages. NeuroImage 155, 530-548. https://doi.org/10.1016/j.neuroimage.2017.03.057

Ritter, K., Schumacher, J., Weygandt, M., Buchert, R., Allefeld, C., Haynes, J.-D., 2015. Multimodal prediction of conversion to Alzheimer's disease based on incomplete biomarkers. Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring 1, 206-215. https://doi.org/10.1016/j.dadm.2015.01.006

Salvatore, C., Cerasa, A., Battista, P., Gilardi, M.C., Quattrone, A., Castiglioni, I., 2015. Magnetic resonance imaging biomarkers for the early diagnosis of Alzheimer's disease: a machine learning approach. Frontiers in Neuroscience 9. https://doi.org/10.3389/fnins.2015.00307

Samper-Gonzalez, J., Burgos, N., Bottani, S., Fontanella, S., Lu, P., Marcoux, A., Routier, A., Guillon, J., Bacci, M., Wen, J. ., Bertrand, A., Bertin, H., Habert, M.-O., Durrleman, S., Evgeniou, T., Colliot, O., 2018. Reproducible evaluation of classification methods in Alzheimer's disease: Framework and application to MRI and PET data. NeuroImage 183, 504-521. https://doi.org/10.1016/j.neuroimage.2018.08.042

Schouten, T.M., Koini, M., de Vos, F., Seiler, S., van der Grond, J., Lechner, A., Hafkemeijer, A., Mϕller, C., Schmidt, R., de Rooij, M., Rombouts, S.A.R.B., 2016. Combining anatomical, diffusion, and resting state functional magnetic resonance imaging for individual classification of mild and moderate Alzheimer's disease. NeuroImage: Clinical 11, 46-51. https://doi.org/10.1016/j.nicl.2016.01.002

Smith, A.M., Lewis, B.K., Ruttimann, U.E., Ye, F.Q., Sinnwell, T.M., Yang, Y., Duyn, J.H., Frank, J.A., 1999. Investigation of Low Frequency Drift in fMRI Signal. NeuroImage 9, 526-533. https://doi.org/10.1006/nimg.1999.0435

Smith, S.M., 2002. Fast robust automated brain extraction. Human Brain Mapping 17, 143-155. https://doi.org/10.1002/hbm.10062

Sun, Y., Bi, Q., Wang, X., Hu, X., Li, H., Li, Han, Y., 2019. Prediction of Conversion From Amnestic Mild Cognitive Impairment to Alzheimer's Disease Based on the Brain Structural Connectome. Frontiers in Neurology 9. https://doi.org/10.3389/fneur.2018.01178

Tustison, N.J., Avants, B.B., Cook, P.A., Yuanjie Zheng, Egan, A., Yushkevich, P.A., Gee, J.C., 2010. N4ITK: Improved N3 Bias Correction. IEEE Transactions on Medical Imaging 29, 1310-1320. https://doi.org/10.1109/TMI.2010.2046908

Wei, R., Li, C., Fogelson, N., Li, L., 2016. Prediction of Conversion from Mild Cognitive Impairment to Alzheimer's Disease Using MRI and Structural Network Features. Frontiers in Aging Neuroscience 8. https://doi.org/10.3389/fnagi.2016.00076

Xu, L., Wu, X., Li, R., Chen, K., Long, Z., Zhang, J., Guo, -Modal Neuroimaging Biomarkers. Journal of Alzheimer's Disease 51, 1045-1056. https://doi.org/10.3233/JAD-151010

Yan, 2010. DPARSF: a MATLAB toolbox for “pipeline” data analysis of resting-state fMRI. Frontiers in System Neuroscience. https://doi.org/10.3389/fnsys.2010.00013

Young, J., Modat, M., Cardoso, M.J., Mendelson, A., Cash, D., Ourselin, S., 2013. Accurate multimodal probabilistic prediction of conversion to Alzheimer's disease in patients with mild cognitive impairment. NeuroImage: Clinical 2, 735-745. https://doi.org/10.1016/j.nicl.2013.05.004

Zang, Y., Jiang, T., Lu, Y., He, Y., Tian, L., 2004. Regional homogeneity approach to fMRI data analysis. NeuroImage 22, 394-400. https://doi.org/10.1016/j.neuroimage.2003.12.030

Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., 2011. Multimodal classification of Alzheimer's disease and mild cognitive impairment. NeuroImage 55, 856-867. https://doi.org/10.1016/j.neuroimage.2011.01.008

Zheng, W., Yao, Z., Hu, B., Gao, X., Cai, H., Moore, P., 2015. Novel Cortical Thickness Pattern for Accurate Detection of Alzheimer's Disease. Journal of Alzheimer's Disease 48, 995-1008. https://doi.org/10.3233/JAD-150311

Claims (10)

1) Kernel matrix calculation step;
2) Kernel matrix combination step;
3) Kernel matrix normalization step;
4) Kernel matrix selection and classification step;
5) Cross-validation step; and
6) Includes the step of classifying the stages of Alzheimer's disease,
The 1) kernel matrix calculation step is,
1-1) Calculating a kernel matrix by whole-brain parcellation analysis;
1-2) Calculating a kernel matrix using voxel-based morphometry; and
1-3) calculating the kernel matrix using a graph theory method; comprising one or more steps selected from the group consisting of,
The kernel matrix includes sMRI, FDG-PET, AV45-PET, DTI, rs-fMRI, and apolipoprotein E genotype,
Step 1-1) is to calculate the kernel matrix using NiftyReg, pyClusterROI and PANDA toolbox,
Steps 1-2) above calculate the kernel matrix using SPM12, DPARSF and FSL toolbox,
The rs-fMRI features include REHO, ALFF and FALFF,
Step 4) is classified as multiple kernel learning, and the multiple kernel learning uses the following [Equation 12],
Step 5) above uses leave-one-out cross-validation,
The progression stage of Alzheimer's disease is one selected from the group consisting of healthy control group, mild cognitive impairment, transitional cognitive disorder, and Alzheimer's disease. Method for providing diagnostic information using a machine learning system for classifying the progression stage of Alzheimer's disease.
[Equation 12]

In [Equation 12] above, sMRI, FDG, AV45, DTI, rs-fMRI, and APOE are normalized by K(x,y), the variable is the normalized value of the kernel, and w is the corresponding weight. A w value is assigned to each normalized value so that the sum of the six values is 1. delete delete delete delete delete delete delete delete delete