Research ArticleClassification of Alzheimerrsquos Disease and Mild CognitiveImpairment Based onCortical and Subcortical Features fromMRIT1 Brain Images Utilizing Four Different Types of Datasets

SaidjalolToshkhujaev1KunHoLee23KyuYeongChoi2 JangJaeLee2Goo-RakKwon 12

Yubraj Gupta12 and Ramesh Kumar Lama12

1School of Information Communication Engineering Chosun University 309 Pilmun-Daero Dong-Gu Gwangju 61452Republic of Korea2National Research Center for Dementia Chosun University 309 Pilmun-Daero Dong-Gu Gwangju 61452Republic of Korea3Department of Biomedical Science College of Natural Sciences Chosun University 309 Pilmun-Daero Dong-GuGwangju 61452 Republic of Korea

Correspondence should be addressed to Goo-Rak Kwon grkwonchosunackr

Received 9 November 2019 Revised 9 July 2020 Accepted 14 July 2020 Published 1 September 2020

Academic Editor Belayat Hossain

Copyright copy 2020 Saidjalol Toshkhujaev et al is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Alzheimerrsquos disease (AD) is one of the most common neurodegenerative illnesses (dementia) among the elderly Recentlyresearchers have developed a new method for the instinctive analysis of AD based on machine learning and its subfield deeplearning Recent state-of-the-art techniques consider multimodal diagnosis which has been shown to achieve high accuracycompared to a unimodal prognosis Furthermore many studies have used structural magnetic resonance imaging (MRI) tomeasure brain volumes and the volume of subregions as well as to search for diffuse changes in whitegray matter in thebrain In this study T1-weighted structural MRI was used for the early classification of AD MRI results in high-intensityvisible features making preprocessing and segmentation easy To use this image modality we acquired four types of datasetsfrom each datasetrsquos server In this work we downloaded 326 subjects from the National Research Center for Dementiahomepage 123 subjects from the Alzheimerrsquos Disease Neuroimaging Initiative (ADNI) homepage 121 subjects from theAlzheimerrsquos Disease Repository Without Borders homepage and 131 subjects from the National Alzheimerrsquos CoordinatingCenter homepage In our experiment we used the multiatlas label propagation with expectationndashmaximization-basedrefinement segmentation method We segmented the images into 138 anatomical morphometry images (in which 40 featuresbelonged to subcortical volumes and the remaining 98 features belonged to cortical thickness) e entire dataset was splitinto a 70 30 (training and testing) ratio before classifying the data A principal component analysis was used for di-mensionality reduction en the support vector machine radial basis function classifier was used for classification betweentwo groupsmdashAD versus health control (HC) and early mild cognitive impairment (MCI) (EMCI) versus late MCI (LMCI)e proposed method performed very well for all four types of dataset For instance for the AD versus HC group theclassifier achieved an area under curve (AUC) of more than 89 for each dataset For the EMCI versus LMCI group theclassifier achieved an AUC of more than 80 for every dataset Moreover we also calculated Cohen kappa and Jaccard indexstatistical values for all datasets to evaluate the classification reliability Finally we compared our results with those ofrecently published state-of-the-art methods

HindawiJournal of Healthcare EngineeringVolume 2020 Article ID 3743171 14 pageshttpsdoiorg10115520203743171

1 Introduction

e occurrence of the most serious and common neuro-degenerative disease Alzheimerrsquos disease (AD) is dramat-ically increasing among the elderly Among people of agesranging from 60 to 84 243 million are suffering from AD[1] e early diagnosis of AD and the best prognosis of mildimpairment are possible because of an increasing list ofpossible biomarkers (from genetics cognition proteomicsand neuroimaging) [2] Mild cognitive impairment (MCI) isthe level between the predictable cognitive deterioration ofregular aging and the more serious decline of dementia Atthe beginning of the MCI stage there might be difficultieswith thinking language and memory that are more thanordinary age-related changes A reason for differentiatingthe above patients from those with prodromal AD at theMCI level is that intervention early in the course of theillness may help postpone the onset and reduce the risk ofAD [3] Such intervention later in the progression of theillness might limit the disease but it might not be possible toshift the pathology-induced neurological damage after it hasalready occurred erefore analysis and identification ofpresymptomatic AD at the MCI point are highly importantSuch treatment will be much more central and compelling asimproved treatment becomes available

In neuroimaging the main task is labeling anatomicalstructures in magnetic resonance imaging (MRI) brainscans with accuracy For clinical decision-making re-gional volume measurement is important as well asaccurate segmentation [4] Currently there is no treat-ment method for AD but many drugs are under devel-opment and it is predicted that a cure will be found soonerefore neuroimaging makes an optimistic prognosismore likely and assessments by structural MRI (sMRI)can be used to check medial temporal lobe (MTL) andpositron emission tomography (PET) fluorodeoxyglucose(FDG) or amyloid results Medial temporal lobe frame-works are essential for producing new memories and forthe improvement of AD [5] Medial temporal lobe at-rophy (MTA) degeneration and the associated episodicmemory impairment are label features of AD and bothimpair over the method of illness [5 6] Moreover MTAis determined by utilizing region of interest- (ROI-) based[7] voxel-based [8] and vertex-based [5] approaches Inthis research the focus is on binary arrangement amongAD health control (HC) early MCI (EMCI) and lateMCI (LMCI) utilizing sMRI According to the atrophyassessment from MRI scans the level of neuro-degeneration and intensity can be determined Studieshave used morphometric approaches such as the volumeof interest (VOI) and ROI voxels for automatic seg-mentation of sMRI images e sMRI volume involvesmeasurement of the medial progressive lobe and hip-pocampus [9] Numerous machine-learning methodshave been implemented to differentiate the binary clas-sifications of AD HC and MCI due to AD (mAD) andasymptomatic AD (aAD) Only using unique modalitiessuch as the hippocampus or amyloid imaging biomarkerscould be less sensitive in analyzing AD progression

mostly at the symptomatic level Currently the relevanceof biomarkers for neurodegeneration which is an ana-lytical component of AD pathophysiology in prodromaland early-stage dementia is widely acknowledged [10]e most-important biomarkers for early detection of ADare the volumetric measurement of cortical thickness andsubcortical volume Studying the cortical thickness is anextensively recognized method for investigating the sizeof gray matter atrophy and is at the cutting edge of ADresearch Cortical thinning has been found in MCI andAD [11] However for subcortical neurofibrillary tangleand amyloid construction in AD MRI investigation hasrecently drawn attention to AD-correlated subcorticalcomplex changes New segmentation methods can assesssubcortical volumes and provide a basis for subcorticalshape analysis [11] Many classification algorithms andapproaches are based on machine learning such assupport vector machine (SVM) k-nearest neighbor(KNN) random forest (RF) and other ensemble classi-fiers Among these the SVM algorithm is commonlyutilized because of its good accuracy and sensitivity thatcan deal with high-dimensional data e SVM classifi-cation method provides a first step for recognizing datafrom the training dataset included in well-characterizedsubjects with known states for which labels are given forthe subjects [12] e margin of the training data ismaximized by composing the optimal splitting hyper-plane or regular hyperplanes in a solitary or higher-di-mensional plane with proposed classifier At the testingstage a test dataset is based on the hyperplane learned inthe classification [13] It is common for T1-weighted MRIimages of each subject to be separated automatically intoROIs according to three anatomical views (sagittalcoronal and axial) as shown in Figure 1 ey are used asfeatures for classification

Recently Liu et al [14] proposed deep learning based onmulticlass classification among normal control (NC) MCInot converted (ncMCI) MCI transformers converted(cMCI) and AD patients grounded on 83 ROI of MRIimages and the conforming disclosed PET images Stackedautoencoders were utilized as unsupervised learning to gainhigh-level features and softmax logistic regression wasadopted as a classifier Nozadi et al [15] researched anapproach for a pipeline utilizing learned features from se-mantically labeled PET images to show group classificationIn that research the ADNI dataset was used and the resultswere validated In classification they used SVM classifierwith radical basis function (RBF) kernel and random forest(RF) with FDG and AV-45 biomarkers of PET image mo-dality for AD NC EMCI and LMCI groups e FDG-PETshows good accuracy such as 917 for AD versus NC withRBF-SVM classifier compared to AV-45-PET MoreoverFDG-PET demonstrates better results for EMCI versusLMCI with RBF-SVM than AV-45-PET Gupta et al [16]proposed a machine-learning-based framework to distin-guish subjects with AD from those with MCI by using fourdifferent biomarkers sMRI the apolipoprotein E (APOE)genotype cerebrospinal fluid (CSF) protein level and FDG-PET from the ADNI dataset According to binary

2 Journal of Healthcare Engineering

classification the combined method showed area under thereceiver operating characteristic curves of 9833 93599683 9464 9643 and 9524 for AD versus HC MCIstable (MCIs) versus cMCI AD versus MCIs AD versuscMCI HC versus cMCI and HC versus MCIs respectivelyGorji and Naima [17] have employed a convolutional neuralnetwork (CNN) based deep learning approach for dis-criminating healthy people from patients with EMCI andLMCI In their research the ADNI dataset was used andtheir proposed method has gained 9454 accuracy 9170sensitivity and 9796 specificity with the sagittal part ofMRI for CN versus LMCI Chyzhyk et al [18] utilized lattice-independent component analysis to combine the kerneltransformation of data with the feature section stage egeneralization of the dendritic computing classifiers wasdeveloped by that approach For classification of NC versusAD patients they also used the OASIS dataset and methodwith an accuracy of 7475 sensitivity of 96 and speci-ficity of 525 Zang et al [19] utilized operative featureconsequent from functional brain network of three fre-quency bands during resting states for the efficiency of theclassification context to classify subjects with EMCI versusLMCI eir approached method demonstrates that thefunctional network features chosen by the minimal re-dundancy maximal relevance (mRMR) algorithm improvethe distinguishing between EMCI versus LMCI comparedwith others chosen by stationary selection (SS-LR) andFisher score (FS) algorithms e chosen slow-5 band showsbetter accuracy compared with other bands such as 8387accuracy 8621 sensitivity and 8121 specification forEMCI versus LMCI Cuingnet et al [20] employed 10 ap-proaches for clinically abnormal subject versus healthygroups by using an sMRI-based feature extraction techniqueis approach included three methods based on corticalthickness and five voxel-based and two other methods forthe hippocampus When the technique was used AD versusHC achieved 81 sensitivity and 95 specificity stable MCIand progressive MCI (P-MCI) had a sensitivity of 70 andspecificity of 61 and HC versus P-MCI had 73 and 85sensitivity and specificity respectively Farhan et al [21]utilized the right and left areas of the hippocampus as well as

the volume of gray matter white matter and CSF extractedfrom sMRI brain images en four types of classificationwere evaluated to achieve good accuracy SVM multilayerperceptron j48 and an ensemble classifier e ensembleclassifier had a high accuracy of 9375 Cho et al [22]studied the incremental learning process based on thelongitudinal frequency which is represented by corticalthickness implemented on 131 ncMCI and 72 cMCI subjectsWhen the method was used for cMCI versus ncMCI thesensitivity and specificity were 63 and 76 respectivelywhich is a better result than that reported previously [21]Wolz et al [23] employed four kinds of automatic featureextraction methods (manifold-based learning corticalthickness tensor-based morphometry and hippocampalvolume) based on sMRI for 834 subjects from the ADNIdataset AD versus MCI and AD versus HC groups elinear discriminant analysis (LDA) and SVM classificationtechniques were compared by manipulating MCI predictionand AD classification In AD versus HC classification LDAachieved 89 accuracy and the sensitivity and specificitywere 93 and 85 respectively Specifically fusion featuresand the LDA classifier showed the best result for the clas-sification of MCI-converted and MCI-stable subjects (68accuracy 67 sensitivity and 69 specification) RecentlyGupta et al [5] proposed four classifier methodsmdashSVM k-nearest neighbors softmax and naıve Bayes (NB)mdashfor bi-nary classification of AD versus HC HC versus mAD andmAD versus aAD and for tertiary classification of AD versusHC versus mAD and AD versus HC versus aAD utilizingsubcortical and cortical features based on 326 subjectsdownloaded from the Gwangju Alzheimerrsquos disease andRelated Dementia (GARD) dataset website e segmenteddataset was parceled into a 70 30 ratio and 70 was used asa training set e remainder was used to obtain unbiasedestimation performance as a test set PCA was manipulatedfor dimensionality reduction purposes and obtained a9906 F1 score by the softmax classifier for AD versus HCbinary classification e SVM classifier achieved for HCversus mAD AD versus aAD binary and AD versus HCversus mAD tertiary classification F1 scores of 9951975 and 9999 respectively NB performed well for AD

(a) (b) (c)

Figure 1 Cross-sectional segmentation results for T1-weighted MRI images (a) axial (b) coronal and sagittal and (c) view planes

Journal of Healthcare Engineering 3

versus HC versus aAD tertiary classification with an F1 scoreof 9588 Moreover to confirm the efficiency of the modelthe OASIS dataset was employed

Compared to related early works this research ad-dresses improving the accuracy and constancy of binaryclassification by comparing it with three kinds of geo-graphical sMRI dataset Moreover all related works useddifferent types of automated feature extraction methodand segmentation toolbox is work focused on the bestaccurate and clear segmentation method which is mul-tiatlas label propagation (MALP) with expect-ationndashmaximization- (EM-) based refinement(MALPEM) [24] e best classifier RBF-SVM wasutilized with the proposed classifier method e GARDdataset was employed to classify the AD versus NC andEMCI versus LMCI binary classifications based on sub-cortical volume and cortical thickness from the sMRIbrain images Finally the segmented features were passedthrough the RBF-SVM classifier and compared to theother three sMRI datasets

2 Materials and Methods

21 Subjects e data used in this study were collected fromthe NRCD National Alzheimerrsquos Coordinating Center(NACC) Alzheimerrsquos Disease Repository Without Borders(ARWIBO) and ADNI All preprocessed brain images wereselected e GARD consists of 81 AD subjects (39 males 42females ageplusmn SD= 7186plusmn 709 years educationlevel = 734plusmn 488 range = 0ndash18) 171 cognitively normal HCsubjects (83 males 88 females ageplusmn SD= 7166plusmn 543 yearseducation level = 916plusmn 554 range = 0ndash22) 39 patients withmAD (25 males 14 females ageplusmn SD= 7323plusmn 709 yearseducation level = 820plusmn 519 range = 0ndash18) and 35 patientswith aAD (15 males 20 females ageplusmn SD= 7274plusmn 482years education level = 788plusmn 630 range = 0ndash18) eNACC includes 26 AD subjects (11 males 15 femalesageplusmn SD= 7333plusmn 943 years education level = 1444plusmn 358range = 0ndash18) 42 cognitively normal HC subjects (22 males20 females ageplusmn SD= 6598plusmn 1191 years educationlevel = 1589plusmn 296 range = 0ndash22) 30 patients with mAD (10males 20 females ageplusmn SD= 7552plusmn 862 years educationlevel = 1491plusmn 345 range = 0ndash18) and 33 patients with aAD(16 males 17 females ageplusmn SD= 7312plusmn 892 years educa-tion level = 1439plusmn 408 range = 0ndash18) e ARWIBO con-sists of 29 AD subjects (10 males 19 femalesageplusmn SD= 712plusmn 414 years education level = 837plusmn 931range = 0ndash18) 33 cognitively normal HC subjects (16 males17 females ageplusmn SD= 655plusmn 909 years educationlevel = 100plusmn 682 range = 0ndash22) 34 patients with mAD (14males 20 females ageplusmn SD= 697plusmn 711 years educationlevel = 767plusmn 421 range = 0ndash18) and 25 patients with aAD(10 males 15 females ageplusmn SD= 6945plusmn 322 years educa-tion level = 797plusmn 521 range = 0ndash18) e ADNI consists of32 AD subjects (17 males 15 females ageplusmn SD= 7214plusmn 421years education level = 941plusmn 378 range = 0ndash18) 28 cog-nitively normal HC subjects (18 males 10 femalesageplusmn SD= 6402plusmn 645 years education level = 1141plusmn 656range = 0ndash22) 25 patients with mAD (12 males 13 females

ageplusmn SD= 6914plusmn 835 years education level = 799plusmn 420range = 0ndash18) and 38 patients with aAD (22 males 16 fe-males ageplusmnSD=6711plusmn581 years education level=802plusmn710range=0ndash18)

Tables 1ndash4 demonstrate the demographics of the 326subjects from the GARD 121 subjects from the ARWIBO131 subjects from the NACC and 123 subjects from theADNI for this study e clinical variables and dis-tinguishing statistics in demographics among the researchgroups were determined using the Welch independentsamples t-test In this research the significance level was005 which is a normal alpha value In the GARD data casethe rate for females was greater than that of other groupsexcept for the EMCI group e levels of education werecompletely disparate in the pairwise comparisons amongstudy groups According to the comparison among thegroups the AD groups had the lowest level of education Inthis study to gain unbiased estimated performance everydataset was randomly split into two parts with a 70 30 ratiofor training and testinge approach involved training witha training algorithm code We tested the remaining datasetusing the trained algorithm Moreover the analysis of thegroup performance was classified in terms of accuracyspecificity sensitivity precision and F1 score utilizing aunique test set

22MRIAcquisition A brain image occupies space in three-dimensional (3D) images so we could use volume data to fillthis space e volume data are measured voxels which looklike the pixels utilized to display images only in 3D

Standard 3T T1-weighted images were obtained utilizingthe volumetric 3D MPRAGE protocol with a resolution1times 1times 1mm (voxel size) All images were N4 bias corrected

23 Feature Selection Segmented MRI brain images havebeen used mostly for classification with machine learningand the subfield of machine-learning techniques Accordingto many types of studies the subcortical part of the brain iseasily affected by dementia and AD compared to the corticalpart but cortical thickness is an outstanding candidate forthe treatment of AD In this research subcorticalcorticalfeatures were extracted by using MALPEM and 138 featureswere achieved from 3D sMRI T1-weighted images MAL-PEM is a collection of tools for distinguishing and visual-ization of corticalsubcortical parts of the brain based on thesagittal axial and coronal views in Figure 1 MALPEM wasconstructed using an automatic workflow consisting ofseveral standard image-processing techniques dividing 138ROIs In recent years multiatlas segmentation has developedinto one of the most accurate methods for the segmentationof T1-weighted images mostly focusing on graph-cut or EMoptimization MALPEM [24] was evaluated as a top 3method in a Grand Challenge on whole-brain segmentationat MICCAI 2012 (httpwwwchristianledigcom) e sMRimages of all 701 subjects were segmented individuallyutilizing MALPEM as designated in previous research [3] Itconsumed between 8 and 10 h for each subject For thissegmentation the automatically explained

4 Journal of Healthcare Engineering

classification the combined method showed area under thereceiver operating characteristic curves of 9833 93599683 9464 9643 and 9524 for AD versus HC MCIstable (MCIs) versus cMCI AD versus MCIs AD versuscMCI HC versus cMCI and HC versus MCIs respectivelyGorji and Naima [17] have employed a convolutional neuralnetwork (CNN) based deep learning approach for dis-criminating healthy people from patients with EMCI andLMCI In their research the ADNI dataset was used andtheir proposed method has gained 9454 accuracy 9170sensitivity and 9796 specificity with the sagittal part ofMRI for CN versus LMCI Chyzhyk et al [18] utilized lattice-independent component analysis to combine the kerneltransformation of data with the feature section stage egeneralization of the dendritic computing classifiers wasdeveloped by that approach For classification of NC versusAD patients they also used the OASIS dataset and methodwith an accuracy of 7475 sensitivity of 96 and speci-ficity of 525 Zang et al [19] utilized operative featureconsequent from functional brain network of three fre-quency bands during resting states for the efficiency of theclassification context to classify subjects with EMCI versusLMCI eir approached method demonstrates that thefunctional network features chosen by the minimal re-dundancy maximal relevance (mRMR) algorithm improvethe distinguishing between EMCI versus LMCI comparedwith others chosen by stationary selection (SS-LR) andFisher score (FS) algorithms e chosen slow-5 band showsbetter accuracy compared with other bands such as 8387accuracy 8621 sensitivity and 8121 specification forEMCI versus LMCI Cuingnet et al [20] employed 10 ap-proaches for clinically abnormal subject versus healthygroups by using an sMRI-based feature extraction techniqueis approach included three methods based on corticalthickness and five voxel-based and two other methods forthe hippocampus When the technique was used AD versusHC achieved 81 sensitivity and 95 specificity stable MCIand progressive MCI (P-MCI) had a sensitivity of 70 andspecificity of 61 and HC versus P-MCI had 73 and 85sensitivity and specificity respectively Farhan et al [21]utilized the right and left areas of the hippocampus as well as

the volume of gray matter white matter and CSF extractedfrom sMRI brain images en four types of classificationwere evaluated to achieve good accuracy SVM multilayerperceptron j48 and an ensemble classifier e ensembleclassifier had a high accuracy of 9375 Cho et al [22]studied the incremental learning process based on thelongitudinal frequency which is represented by corticalthickness implemented on 131 ncMCI and 72 cMCI subjectsWhen the method was used for cMCI versus ncMCI thesensitivity and specificity were 63 and 76 respectivelywhich is a better result than that reported previously [21]Wolz et al [23] employed four kinds of automatic featureextraction methods (manifold-based learning corticalthickness tensor-based morphometry and hippocampalvolume) based on sMRI for 834 subjects from the ADNIdataset AD versus MCI and AD versus HC groups elinear discriminant analysis (LDA) and SVM classificationtechniques were compared by manipulating MCI predictionand AD classification In AD versus HC classification LDAachieved 89 accuracy and the sensitivity and specificitywere 93 and 85 respectively Specifically fusion featuresand the LDA classifier showed the best result for the clas-sification of MCI-converted and MCI-stable subjects (68accuracy 67 sensitivity and 69 specification) RecentlyGupta et al [5] proposed four classifier methodsmdashSVM k-nearest neighbors softmax and naıve Bayes (NB)mdashfor bi-nary classification of AD versus HC HC versus mAD andmAD versus aAD and for tertiary classification of AD versusHC versus mAD and AD versus HC versus aAD utilizingsubcortical and cortical features based on 326 subjectsdownloaded from the Gwangju Alzheimerrsquos disease andRelated Dementia (GARD) dataset website e segmenteddataset was parceled into a 70 30 ratio and 70 was used asa training set e remainder was used to obtain unbiasedestimation performance as a test set PCA was manipulatedfor dimensionality reduction purposes and obtained a9906 F1 score by the softmax classifier for AD versus HCbinary classification e SVM classifier achieved for HCversus mAD AD versus aAD binary and AD versus HCversus mAD tertiary classification F1 scores of 9951975 and 9999 respectively NB performed well for AD

(a) (b) (c)

Figure 1 Cross-sectional segmentation results for T1-weighted MRI images (a) axial (b) coronal and sagittal and (c) view planes

Journal of Healthcare Engineering 3

versus HC versus aAD tertiary classification with an F1 scoreof 9588 Moreover to confirm the efficiency of the modelthe OASIS dataset was employed

Compared to related early works this research ad-dresses improving the accuracy and constancy of binaryclassification by comparing it with three kinds of geo-graphical sMRI dataset Moreover all related works useddifferent types of automated feature extraction methodand segmentation toolbox is work focused on the bestaccurate and clear segmentation method which is mul-tiatlas label propagation (MALP) with expect-ationndashmaximization- (EM-) based refinement(MALPEM) [24] e best classifier RBF-SVM wasutilized with the proposed classifier method e GARDdataset was employed to classify the AD versus NC andEMCI versus LMCI binary classifications based on sub-cortical volume and cortical thickness from the sMRIbrain images Finally the segmented features were passedthrough the RBF-SVM classifier and compared to theother three sMRI datasets

2 Materials and Methods

21 Subjects e data used in this study were collected fromthe NRCD National Alzheimerrsquos Coordinating Center(NACC) Alzheimerrsquos Disease Repository Without Borders(ARWIBO) and ADNI All preprocessed brain images wereselected e GARD consists of 81 AD subjects (39 males 42females ageplusmn SD= 7186plusmn 709 years educationlevel = 734plusmn 488 range = 0ndash18) 171 cognitively normal HCsubjects (83 males 88 females ageplusmn SD= 7166plusmn 543 yearseducation level = 916plusmn 554 range = 0ndash22) 39 patients withmAD (25 males 14 females ageplusmn SD= 7323plusmn 709 yearseducation level = 820plusmn 519 range = 0ndash18) and 35 patientswith aAD (15 males 20 females ageplusmn SD= 7274plusmn 482years education level = 788plusmn 630 range = 0ndash18) eNACC includes 26 AD subjects (11 males 15 femalesageplusmn SD= 7333plusmn 943 years education level = 1444plusmn 358range = 0ndash18) 42 cognitively normal HC subjects (22 males20 females ageplusmn SD= 6598plusmn 1191 years educationlevel = 1589plusmn 296 range = 0ndash22) 30 patients with mAD (10males 20 females ageplusmn SD= 7552plusmn 862 years educationlevel = 1491plusmn 345 range = 0ndash18) and 33 patients with aAD(16 males 17 females ageplusmn SD= 7312plusmn 892 years educa-tion level = 1439plusmn 408 range = 0ndash18) e ARWIBO con-sists of 29 AD subjects (10 males 19 femalesageplusmn SD= 712plusmn 414 years education level = 837plusmn 931range = 0ndash18) 33 cognitively normal HC subjects (16 males17 females ageplusmn SD= 655plusmn 909 years educationlevel = 100plusmn 682 range = 0ndash22) 34 patients with mAD (14males 20 females ageplusmn SD= 697plusmn 711 years educationlevel = 767plusmn 421 range = 0ndash18) and 25 patients with aAD(10 males 15 females ageplusmn SD= 6945plusmn 322 years educa-tion level = 797plusmn 521 range = 0ndash18) e ADNI consists of32 AD subjects (17 males 15 females ageplusmn SD= 7214plusmn 421years education level = 941plusmn 378 range = 0ndash18) 28 cog-nitively normal HC subjects (18 males 10 femalesageplusmn SD= 6402plusmn 645 years education level = 1141plusmn 656range = 0ndash22) 25 patients with mAD (12 males 13 females

ageplusmn SD= 6914plusmn 835 years education level = 799plusmn 420range = 0ndash18) and 38 patients with aAD (22 males 16 fe-males ageplusmnSD=6711plusmn581 years education level=802plusmn710range=0ndash18)

Tables 1ndash4 demonstrate the demographics of the 326subjects from the GARD 121 subjects from the ARWIBO131 subjects from the NACC and 123 subjects from theADNI for this study e clinical variables and dis-tinguishing statistics in demographics among the researchgroups were determined using the Welch independentsamples t-test In this research the significance level was005 which is a normal alpha value In the GARD data casethe rate for females was greater than that of other groupsexcept for the EMCI group e levels of education werecompletely disparate in the pairwise comparisons amongstudy groups According to the comparison among thegroups the AD groups had the lowest level of education Inthis study to gain unbiased estimated performance everydataset was randomly split into two parts with a 70 30 ratiofor training and testinge approach involved training witha training algorithm code We tested the remaining datasetusing the trained algorithm Moreover the analysis of thegroup performance was classified in terms of accuracyspecificity sensitivity precision and F1 score utilizing aunique test set

22MRIAcquisition A brain image occupies space in three-dimensional (3D) images so we could use volume data to fillthis space e volume data are measured voxels which looklike the pixels utilized to display images only in 3D

Standard 3T T1-weighted images were obtained utilizingthe volumetric 3D MPRAGE protocol with a resolution1times 1times 1mm (voxel size) All images were N4 bias corrected

23 Feature Selection Segmented MRI brain images havebeen used mostly for classification with machine learningand the subfield of machine-learning techniques Accordingto many types of studies the subcortical part of the brain iseasily affected by dementia and AD compared to the corticalpart but cortical thickness is an outstanding candidate forthe treatment of AD In this research subcorticalcorticalfeatures were extracted by using MALPEM and 138 featureswere achieved from 3D sMRI T1-weighted images MAL-PEM is a collection of tools for distinguishing and visual-ization of corticalsubcortical parts of the brain based on thesagittal axial and coronal views in Figure 1 MALPEM wasconstructed using an automatic workflow consisting ofseveral standard image-processing techniques dividing 138ROIs In recent years multiatlas segmentation has developedinto one of the most accurate methods for the segmentationof T1-weighted images mostly focusing on graph-cut or EMoptimization MALPEM [24] was evaluated as a top 3method in a Grand Challenge on whole-brain segmentationat MICCAI 2012 (httpwwwchristianledigcom) e sMRimages of all 701 subjects were segmented individuallyutilizing MALPEM as designated in previous research [3] Itconsumed between 8 and 10 h for each subject For thissegmentation the automatically explained

4 Journal of Healthcare Engineering

versus HC versus aAD tertiary classification with an F1 scoreof 9588 Moreover to confirm the efficiency of the modelthe OASIS dataset was employed

Compared to related early works this research ad-dresses improving the accuracy and constancy of binaryclassification by comparing it with three kinds of geo-graphical sMRI dataset Moreover all related works useddifferent types of automated feature extraction methodand segmentation toolbox is work focused on the bestaccurate and clear segmentation method which is mul-tiatlas label propagation (MALP) with expect-ationndashmaximization- (EM-) based refinement(MALPEM) [24] e best classifier RBF-SVM wasutilized with the proposed classifier method e GARDdataset was employed to classify the AD versus NC andEMCI versus LMCI binary classifications based on sub-cortical volume and cortical thickness from the sMRIbrain images Finally the segmented features were passedthrough the RBF-SVM classifier and compared to theother three sMRI datasets

2 Materials and Methods

21 Subjects e data used in this study were collected fromthe NRCD National Alzheimerrsquos Coordinating Center(NACC) Alzheimerrsquos Disease Repository Without Borders(ARWIBO) and ADNI All preprocessed brain images wereselected e GARD consists of 81 AD subjects (39 males 42females ageplusmn SD= 7186plusmn 709 years educationlevel = 734plusmn 488 range = 0ndash18) 171 cognitively normal HCsubjects (83 males 88 females ageplusmn SD= 7166plusmn 543 yearseducation level = 916plusmn 554 range = 0ndash22) 39 patients withmAD (25 males 14 females ageplusmn SD= 7323plusmn 709 yearseducation level = 820plusmn 519 range = 0ndash18) and 35 patientswith aAD (15 males 20 females ageplusmn SD= 7274plusmn 482years education level = 788plusmn 630 range = 0ndash18) eNACC includes 26 AD subjects (11 males 15 femalesageplusmn SD= 7333plusmn 943 years education level = 1444plusmn 358range = 0ndash18) 42 cognitively normal HC subjects (22 males20 females ageplusmn SD= 6598plusmn 1191 years educationlevel = 1589plusmn 296 range = 0ndash22) 30 patients with mAD (10males 20 females ageplusmn SD= 7552plusmn 862 years educationlevel = 1491plusmn 345 range = 0ndash18) and 33 patients with aAD(16 males 17 females ageplusmn SD= 7312plusmn 892 years educa-tion level = 1439plusmn 408 range = 0ndash18) e ARWIBO con-sists of 29 AD subjects (10 males 19 femalesageplusmn SD= 712plusmn 414 years education level = 837plusmn 931range = 0ndash18) 33 cognitively normal HC subjects (16 males17 females ageplusmn SD= 655plusmn 909 years educationlevel = 100plusmn 682 range = 0ndash22) 34 patients with mAD (14males 20 females ageplusmn SD= 697plusmn 711 years educationlevel = 767plusmn 421 range = 0ndash18) and 25 patients with aAD(10 males 15 females ageplusmn SD= 6945plusmn 322 years educa-tion level = 797plusmn 521 range = 0ndash18) e ADNI consists of32 AD subjects (17 males 15 females ageplusmn SD= 7214plusmn 421years education level = 941plusmn 378 range = 0ndash18) 28 cog-nitively normal HC subjects (18 males 10 femalesageplusmn SD= 6402plusmn 645 years education level = 1141plusmn 656range = 0ndash22) 25 patients with mAD (12 males 13 females

ageplusmn SD= 6914plusmn 835 years education level = 799plusmn 420range = 0ndash18) and 38 patients with aAD (22 males 16 fe-males ageplusmnSD=6711plusmn581 years education level=802plusmn710range=0ndash18)

Tables 1ndash4 demonstrate the demographics of the 326subjects from the GARD 121 subjects from the ARWIBO131 subjects from the NACC and 123 subjects from theADNI for this study e clinical variables and dis-tinguishing statistics in demographics among the researchgroups were determined using the Welch independentsamples t-test In this research the significance level was005 which is a normal alpha value In the GARD data casethe rate for females was greater than that of other groupsexcept for the EMCI group e levels of education werecompletely disparate in the pairwise comparisons amongstudy groups According to the comparison among thegroups the AD groups had the lowest level of education Inthis study to gain unbiased estimated performance everydataset was randomly split into two parts with a 70 30 ratiofor training and testinge approach involved training witha training algorithm code We tested the remaining datasetusing the trained algorithm Moreover the analysis of thegroup performance was classified in terms of accuracyspecificity sensitivity precision and F1 score utilizing aunique test set

22MRIAcquisition A brain image occupies space in three-dimensional (3D) images so we could use volume data to fillthis space e volume data are measured voxels which looklike the pixels utilized to display images only in 3D

Standard 3T T1-weighted images were obtained utilizingthe volumetric 3D MPRAGE protocol with a resolution1times 1times 1mm (voxel size) All images were N4 bias corrected

23 Feature Selection Segmented MRI brain images havebeen used mostly for classification with machine learningand the subfield of machine-learning techniques Accordingto many types of studies the subcortical part of the brain iseasily affected by dementia and AD compared to the corticalpart but cortical thickness is an outstanding candidate forthe treatment of AD In this research subcorticalcorticalfeatures were extracted by using MALPEM and 138 featureswere achieved from 3D sMRI T1-weighted images MAL-PEM is a collection of tools for distinguishing and visual-ization of corticalsubcortical parts of the brain based on thesagittal axial and coronal views in Figure 1 MALPEM wasconstructed using an automatic workflow consisting ofseveral standard image-processing techniques dividing 138ROIs In recent years multiatlas segmentation has developedinto one of the most accurate methods for the segmentationof T1-weighted images mostly focusing on graph-cut or EMoptimization MALPEM [24] was evaluated as a top 3method in a Grand Challenge on whole-brain segmentationat MICCAI 2012 (httpwwwchristianledigcom) e sMRimages of all 701 subjects were segmented individuallyutilizing MALPEM as designated in previous research [3] Itconsumed between 8 and 10 h for each subject For thissegmentation the automatically explained

4 Journal of Healthcare Engineering

neuromorphometrics brain atlas (n 30 provided byNeuromorphometrics Inc under academic subscriptionhttpNeuromorphometric last accessed 15 March 2018)was used is atlas automatically distinguishes the whole-brain images into 40 noncortical and 98 cortical partsMALP was utilized to acquire the specific probability of abrain atlas for sMRI brain images as K which should besegmented [3] is probability is integrated into the EMframework as a spatial anatomical task n is indicated as avoxel of K by j 1 n therefore the intensities of the voxelZi isin D image should be described as K Z1 Z2 Zn1113864 1113865e probabilistic priors are produced through the trans-formation of manually generated L atlases into the coor-dinate space of the unseen image For the propagation of thetag the L transformations were measured using a nonrigidregistering technique based on free-form deformationwhich follows a previous rigid technique and some align-ment By using a locally weighted multiatlas fusion strategythe probabilistic atlas was designed for image intensitynormalization and rescaling by the Gaussian weighted sumof squared differentiation e estimated hidden segmen-tation employing the observed intensities j followed theapproach of Van Leemput et al [25] It was considered that

the observed log-transformed intensities of the voxels referto a spatial class N and are dispensed with mean φN andstandard deviation ρN

ψ φ1 ρ1( 1113857 φ2 ρ2( 1113857 φN ρN( 11138571113864 1113865 (1)

e global Markov random fields approach was used forapplying the regularization of the resulting segmentatione EM algorithm makes segmentations with very low-in-tensity variance within the region (intraclass variance)compared to the ldquogold-standardrdquo segmentations ereforenormalized intraclass variance was determined for eachregion (ρNGoldk) by averaging the normalized standarddeviations (ρN φN) of every group over the trainingsubjects Moreover the averaged (averaged over all trainingsubjects segmented with a leave-one-out strategy) distrib-uted standard deviation was measured within every regionassembled by the EM algorithm (ρEMk) By determiningΔN (pNGoldk minus pEMk)2 it was evaluated by which valuethe intraclass variance of the spatial group might be en-hanced on average to match the gold-standard innates better[26] e final segmentation was created by fusing the re-fined labels for this subset with the labels from the MALP

Table 1 Demographic characteristics of the studied population (from the GARD database)

Group Subject number AgeGender

EducationM F

AD 81 7186plusmn 709 [56ndash83] 39 42 734plusmn 488 [0ndash18]EMCI 39 7323plusmn 734 [49ndash87] 25 14 820plusmn 519 [0ndash18]LMCI 35 7274plusmn 482 [61ndash83] 15 20 788plusmn 630 [0ndash18]HC 171 7166plusmn 543 [60ndash85] 83 88 916plusmn 554 [0ndash22]

Table 3 Demographic characteristics of studied population (from the NACC dataset)

Group Subject number AgeGender

EducationM F

AD 26 7333plusmn 943 [50ndash78] 11 15 1444plusmn 358 [0ndash18]EMCI 30 7552plusmn 862 [47ndash85] 10 20 1491plusmn 345 [0ndash18]LMCI 33 7312plusmn 892 [58ndash80] 16 17 1439plusmn 408 [0ndash18]HC 42 6598plusmn 1191 [59ndash83] 22 20 1589plusmn 296 [0ndash22]

Table 4 Demographic characteristics of studied population (from the ADNI dataset)

Group Subjects number AgeGender

EducationM F

AD 32 7214plusmn 421 [54ndash79] 17 15 941plusmn 378 [0ndash18]EMCI 25 6914plusmn 835 [48ndash83] 12 13 799plusmn 420 [0ndash18]LMCI 38 6711plusmn 581 [60ndash81] 22 16 802plusmn 710 [0ndash18]HC 28 6402plusmn 645 [63ndash84] 18 10 1141plusmn 656 [0ndash22]

Table 2 Demographic characteristics of studied population (from the ARWIBO dataset)

Group Subject number AgeGender

EducationM F

AD 29 7124plusmn 1409 [59ndash80] 10 19 837plusmn 378 [0ndash18]EMCI 34 697plusmn 711 [51ndash82] 14 20 767plusmn 421 [0ndash18]LMCI 25 6945plusmn 322 [59ndash79] 10 15 797plusmn 521 [0ndash18]HC 33 6559plusmn 912 [58ndash83] 16 17 1006plusmn 343 [0ndash22]

Journal of Healthcare Engineering 5

approach for enduring parts After completion of segmen-tation all the segmented data were normalized to zero meanand component variance for every feature as demonstratedin Figure 2 utilizing the ordinary scalar function of thescikit-learn library According to normalization ξ is a givendata matrix where the subjects are in rows and the featuresof subjects are in columns e elements of normalizedmatrix illustrate ξ (m n) and the equation is given by

ξ norm(mn) ξ (mn) minus mean ξ n( 1113857

std ξ n( 1113857 (2)

en for reduction of dimensionality PCA was exe-cuted [27] with all features designed into a lower-dimen-sional space PCA map features in a new N-dimensionalsubspace should be less than those in the initial L-dimen-sional spacee newN variance is the principal componentand every principal component eliminates the maximumvariance that is accounted for in all accomplishing com-ponents e principal components can be illustrated by thefollowing equation

PCj m1k1 + m2k2 + middot middot middot middot middot middot + mzkz (3)

where PCj is a principal component in j yz is an originalfeature in z and mz is a numerical coefficient of yz eobserved original features are greater than or equal to thenumber of principal components e achieved principalcomponents for AD versus HC are illustrated in Figure 3e component number was regulated by controlling thefeatures greater than 96 e first principal componentgained 96 among all the other 83 features received afterpassing all 138 features Hence the first component wasutilized for the classification of EMCI versus LMCI as well

24 Classification Method Previously the cortical andsubcortical region features were extracted by using an au-tomatic toolbox (MALP-EM) based on sMRI T1-weightedimages e feature vectors covering mean-centered voxelintensities were constructed by fusing all features eclassification method used here is designed to fuse tworesources of features subcortical volume and corticalthickness e above features are utilized for a frameworkdecision to distinguish AD from other subjects Moreoverwhen comparing the classification for two-part brainimagesmdashcortical and subcorticalmdashby RBF-SVM machine-learning classifiers in some dataset cases the corticalthickness was determined with good accuracy in anothercase the subcortical provided better results

25 SVM is machine-learning classifier is being usedwidely to investigate sMRI data [5 12 20] RBF-SVM issuitable for binary classification for separable and non-separable data In the past decade it has been employed asthe most popular machine-learning tool in neuroimagingand neuroscience is approach depends on selecting acritical point for the classification work Support vectors are

elements that are distinguished into two groups RBF-SVMis a supervised learning classification algorithm and deter-mines the optimal hyperplane that discriminates bothmodules with an extreme margin from support vectorsduring the training phase e estimated hyperplane de-termines the classifier for the testing of a new data point Insome cases such as linear ones SVM might not guarantee agood result so linear SVM is expanded by utilizing a kerneltrick e central idea of kernel methods is that the inputdata are designed into a higher-dimensional planeemploying linear and nonlinear functions for known ker-nels SVM kernels exploit nonlinear and linear RBFMoreover the linear kernel is a superior case of RBF elinear kernel with a penalty parameter C has the samepresentation as the RBF kernel with the parameters (c y)e RBF has two parameters c and y that are good for anassumed problem e main goal of analyzing good (c y) isto enable the classification method to predict testing dataaccurately

3 Results and Discussion

31 Background In this experiment the proposed methodutilized the RBF-SVM classification algorithm en allextracted features were split into the cortical thickness andsubcortical volume in the MALPEM toolbox to distinguishbetween the AD versus HC and EMCI versus LMCI groupsis classification was performed to recognize how well theapproach performs on the sMRI 3D images In the beginningstage normalization was designed for every subject Fur-thermore the PCA reduction of dimensionality method wasemployed to select the optimal number of principal com-ponents for every binary classification In a different binaryclassification group different numbers of principal com-ponents were gained Moreover to confirm the robustness ofthe classification result 10-fold stratified K-fold (SKF) cross-validation (CV) was utilized

32 Evaluation e set of subjects was randomly split intotwo groups in a 70 30 ratio for training and testing to obtainunbiased estimates of the performance e training set wasused for analyzing the optimal hyperparameters of everytechnique and classifier Furthermore the testing set wasutilized for classification achievement For optimal hyper-parameter estimation SKF-CV was used based on thetraining set Following the LIBSVM library the linear SVMC kernel value and c y parameters of the kernel wereregulated While choosing the default parameter featuresSVM performed poorly on the preliminary data Hence toregulate the optimal parameters for c and y the grid searchmethod was used before c and y were utilized for trainingFor c and y CV accuracy was selected as the best parameterIn this work two parameters were setmdashc 1 to 10 andy (1eminus4 1eminus2 00001)mdashfor every classification group ento analyze the classifier for the training groups the obtainedoptimized values were regulated e assessment of binaryclassifiers was determined by using confusionmetrics whichis a precise test covering binary classification tasks e

6 Journal of Healthcare Engineering

diagonal elements of the metric demonstrate the correctedpredictions created by the classifier en elements could besplit into two groups that express the controls of correctlyidentified true positive (TP) and true negative (TN)However the subjects classified incorrectly can be defined asfalse positive (FP) and false negative (FN) e determi-nation of accuracy in equation (4) is the number of samplesin which the classifiers regulated correctly

Acc TP + TN

TP + TN + FP + FN (4)

Moreover considering only accurate measurement wasnot sufficient for the unstable class dataset and resulted inmisleading estimation Hence the additional four assess-ment metrics should be adjusted sensitivity specificityprecision and F1 score ey are designated as follows

Sen TP

TP + FN (5)

Spec TN

TN + FP (6)

Ppv TP

TP + FP (7)

F1 score 2TP

2TP + FP + FN (8)

Sensitivity given in equation (5) illustrates the accuracyof the predicted group Specificity given in equation (6)illustrates the accuracy of the predicted absence groupSensitivity which is also called ldquorecallrdquo or ldquoprobability ofdetectionrdquo is the proportion of actual positives that arecorrectly determined Similarly specificity investigates theproportion of actual negatives that is not included in the

class Precision given in equation (7) (positive predictivevalue) is the element of appropriate incidences between therepossessed incidences F1 score given in equation (8)determines the accuracy of a test To assess that each methodperforms significantly better than a random classifier theCohen kappa and Jaccard distance were utilized Cohenkappa determination is usually utilized to analyze interraterreliability Rater reliability demonstrates the degree to whichthe data represented in the research are appropriate illus-trations of the variables evaluated e Cohen [28] is astatistic helpful for an interrater accuracy testing e de-termination of the Cohen kappa can be executed based onthe following equation

k Pr(a) minus Pr(e)

1 minus Pr(e) (9)

Here Pr (a) demonstrates the observed agreement andPr (e) illustrates chance agreement e kappa can rangebetween minus1 and + 1 and the kappa result is explained asfollows if the values are le0 there is no agreement between001 and 020 there is minor arrangement between 021 and040 there is known fair agreement between 041 and 060there is moderate agreement between 061 and 080 there issubstantial agreement and from 081 to 100 there is nearlyperfect agreement [29] e Jaccard index determines howclose the commonality of the two datasets can be a measured[30] e Jaccard coefficient is given in the followingequation

J(M N) Mcap N| |

Mcup N| | (10)

e Jaccard index is a statistic utilized for measuring thediversity and similarity of sample sets

Image preprocessing MALPEM toolbox

Normalization

Principle component analysis

Radial basic function-SVM

Testing RF modelTraining RF model

Label

3D image

10-fold stratified K-foldcross-validation (SK-CV)

Diagnosis ofADEMCILMCI

HC

138 segmented ROIimages

Figure 2 Proposed technique workflow

Journal of Healthcare Engineering 7

e range of Jaccard coefficient measures from 0 to100 is as follows

Jaccard index (the amount in equal sets)(the amountin either set)times 100

e process of calculating the Jaccard index is as follows(1) calculate the number of both members that are pro-portional for both sets (2) calculate the entire number ofmembers (proportional and nonproportional) and split thenumber of common members (1) by the entire number ofmembers (2) and (3) multiply by 100 [31]

33 Classification Results e binary classification wasutilized to analyze subcortical volume and cortical thicknessvolume of AD versus HC and EMCI versus LMCI subjectgroups e results of classification are demonstrated inTable 5 and all results are illustrated in Figures 4(a) and 4(b)Kappa and Jaccard are shown in Figures 4(c) and 4(d) Allprocesses were performed in a 64-bit Python 36 environ-ment on Intel Core i7-8700 at 320Hz and 16GB of RAMrunning Ubuntu 1604

331 Binary Classification AD versus HC and EMCI versusLMCI According to the four datasets two binary group-smdashAD versus HC and EMCI versus LMCImdashwere classifiedwith subcortical volume and cortical thickness features byutilizing the RBF-SVM classification algorithm and theresult is illustrated in Table 5

(1) AD versus HC e result of classification for ADversus HC is given in Table 5 and Figures 4(a) and 4(c) Inevery classification situation the database was divided intotwo separations in a 70 30 ratio Each dataset shows betterresults for kappa and Jaccard statistics in the RBF-SVMclassification technique In the kappa statistics case theGARD dataset cortical thickness feature gives the highestinterrater reliability of 09342 and the ARWIBO dataset

subcortical volume feature has 08939 In the Jaccard sta-tistics case the GARDdataset cortical thickness is 09091 thehighest and the ARWIBO dataset subcortical volume fea-ture is 08889 Moreover the GARD cortical thickness has agood F1 score precision specificity and sensitivity (98189687 9524 and 9818 respectively)

(2) EMCI versus LMCI e classification performancefor EMCI versus LMCI is shown in Table 5 and Figures 4(b)and 4(d) Similarly the kappa statistics for the GARD datasetsubcortical volume feature shows the highestresultmdash09043mdashand the ARWIBO dataset cortical feature is08979 e Jaccard statistic for the GARD dataset subcor-tical volume feature had the highest resultmdash09186 eNACC dataset subcortical volume feature is 09167 Fur-thermore the GARD dataset subcortical volume feature hasbetter results with an F1 score precision specificity andsensitivity of 9645 100 100 and 9275 respectivelycompared to the cortical thickness

332 Comparison with Related Recently Published MethodsAs mentioned above in the proposed method four datasetswere used GARD NACC ARWIBO and ADNI Except forthe GARD dataset all are available for the public andanyone can download them e websites available areNACC (httpswwwalzwashingtonedu) ARWIBO(httpswwwgaaindataorgpartnerARWIBO) and ADNI(httpadniloniuscedu)

ere are 701 subjects 326 (GARD) 131 (NACC) 121(ARWIBO) and 123 (ADNI) e ages of all subjects aremore than 47 and converting of mild cognitive stage to ADbetween 0 and 18 months For segmentation the T1-weighted sMRI imaging modality was utilized from eachdataset Moreover the results of the proposed method werecompared to recently published results as shown in Tables 6and 7

100

90

80

70

60

50

40

30

20

10

0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 39 41 43 45 49 51 53 55 57 59 61 63 65 67 69 71 73 75 79 81 83

PCA for AD vs HC NRCD dataset

Figure 3 Number of principal components for AD versus HC

8 Journal of Healthcare Engineering

diagonal elements of the metric demonstrate the correctedpredictions created by the classifier en elements could besplit into two groups that express the controls of correctlyidentified true positive (TP) and true negative (TN)However the subjects classified incorrectly can be defined asfalse positive (FP) and false negative (FN) e determi-nation of accuracy in equation (4) is the number of samplesin which the classifiers regulated correctly

Acc TP + TN

TP + TN + FP + FN (4)

Moreover considering only accurate measurement wasnot sufficient for the unstable class dataset and resulted inmisleading estimation Hence the additional four assess-ment metrics should be adjusted sensitivity specificityprecision and F1 score ey are designated as follows

Sen TP

TP + FN (5)

Spec TN

TN + FP (6)

Ppv TP

TP + FP (7)

F1 score 2TP

2TP + FP + FN (8)

Sensitivity given in equation (5) illustrates the accuracyof the predicted group Specificity given in equation (6)illustrates the accuracy of the predicted absence groupSensitivity which is also called ldquorecallrdquo or ldquoprobability ofdetectionrdquo is the proportion of actual positives that arecorrectly determined Similarly specificity investigates theproportion of actual negatives that is not included in the

class Precision given in equation (7) (positive predictivevalue) is the element of appropriate incidences between therepossessed incidences F1 score given in equation (8)determines the accuracy of a test To assess that each methodperforms significantly better than a random classifier theCohen kappa and Jaccard distance were utilized Cohenkappa determination is usually utilized to analyze interraterreliability Rater reliability demonstrates the degree to whichthe data represented in the research are appropriate illus-trations of the variables evaluated e Cohen [28] is astatistic helpful for an interrater accuracy testing e de-termination of the Cohen kappa can be executed based onthe following equation

k Pr(a) minus Pr(e)

1 minus Pr(e) (9)

Here Pr (a) demonstrates the observed agreement andPr (e) illustrates chance agreement e kappa can rangebetween minus1 and + 1 and the kappa result is explained asfollows if the values are le0 there is no agreement between001 and 020 there is minor arrangement between 021 and040 there is known fair agreement between 041 and 060there is moderate agreement between 061 and 080 there issubstantial agreement and from 081 to 100 there is nearlyperfect agreement [29] e Jaccard index determines howclose the commonality of the two datasets can be a measured[30] e Jaccard coefficient is given in the followingequation

J(M N) Mcap N| |

Mcup N| | (10)

e Jaccard index is a statistic utilized for measuring thediversity and similarity of sample sets

Image preprocessing MALPEM toolbox

Normalization

Principle component analysis

Radial basic function-SVM

Testing RF modelTraining RF model

Label

3D image

10-fold stratified K-foldcross-validation (SK-CV)

Diagnosis ofADEMCILMCI

HC

138 segmented ROIimages

Figure 2 Proposed technique workflow

Journal of Healthcare Engineering 7

e range of Jaccard coefficient measures from 0 to100 is as follows

Jaccard index (the amount in equal sets)(the amountin either set)times 100

e process of calculating the Jaccard index is as follows(1) calculate the number of both members that are pro-portional for both sets (2) calculate the entire number ofmembers (proportional and nonproportional) and split thenumber of common members (1) by the entire number ofmembers (2) and (3) multiply by 100 [31]

33 Classification Results e binary classification wasutilized to analyze subcortical volume and cortical thicknessvolume of AD versus HC and EMCI versus LMCI subjectgroups e results of classification are demonstrated inTable 5 and all results are illustrated in Figures 4(a) and 4(b)Kappa and Jaccard are shown in Figures 4(c) and 4(d) Allprocesses were performed in a 64-bit Python 36 environ-ment on Intel Core i7-8700 at 320Hz and 16GB of RAMrunning Ubuntu 1604

331 Binary Classification AD versus HC and EMCI versusLMCI According to the four datasets two binary group-smdashAD versus HC and EMCI versus LMCImdashwere classifiedwith subcortical volume and cortical thickness features byutilizing the RBF-SVM classification algorithm and theresult is illustrated in Table 5

(1) AD versus HC e result of classification for ADversus HC is given in Table 5 and Figures 4(a) and 4(c) Inevery classification situation the database was divided intotwo separations in a 70 30 ratio Each dataset shows betterresults for kappa and Jaccard statistics in the RBF-SVMclassification technique In the kappa statistics case theGARD dataset cortical thickness feature gives the highestinterrater reliability of 09342 and the ARWIBO dataset

subcortical volume feature has 08939 In the Jaccard sta-tistics case the GARDdataset cortical thickness is 09091 thehighest and the ARWIBO dataset subcortical volume fea-ture is 08889 Moreover the GARD cortical thickness has agood F1 score precision specificity and sensitivity (98189687 9524 and 9818 respectively)

(2) EMCI versus LMCI e classification performancefor EMCI versus LMCI is shown in Table 5 and Figures 4(b)and 4(d) Similarly the kappa statistics for the GARD datasetsubcortical volume feature shows the highestresultmdash09043mdashand the ARWIBO dataset cortical feature is08979 e Jaccard statistic for the GARD dataset subcor-tical volume feature had the highest resultmdash09186 eNACC dataset subcortical volume feature is 09167 Fur-thermore the GARD dataset subcortical volume feature hasbetter results with an F1 score precision specificity andsensitivity of 9645 100 100 and 9275 respectivelycompared to the cortical thickness

332 Comparison with Related Recently Published MethodsAs mentioned above in the proposed method four datasetswere used GARD NACC ARWIBO and ADNI Except forthe GARD dataset all are available for the public andanyone can download them e websites available areNACC (httpswwwalzwashingtonedu) ARWIBO(httpswwwgaaindataorgpartnerARWIBO) and ADNI(httpadniloniuscedu)

ere are 701 subjects 326 (GARD) 131 (NACC) 121(ARWIBO) and 123 (ADNI) e ages of all subjects aremore than 47 and converting of mild cognitive stage to ADbetween 0 and 18 months For segmentation the T1-weighted sMRI imaging modality was utilized from eachdataset Moreover the results of the proposed method werecompared to recently published results as shown in Tables 6and 7

100

90

80

70

60

50

40

30

20

10

0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 39 41 43 45 49 51 53 55 57 59 61 63 65 67 69 71 73 75 79 81 83

PCA for AD vs HC NRCD dataset

Figure 3 Number of principal components for AD versus HC

8 Journal of Healthcare Engineering

e range of Jaccard coefficient measures from 0 to100 is as follows

Jaccard index (the amount in equal sets)(the amountin either set)times 100

e process of calculating the Jaccard index is as follows(1) calculate the number of both members that are pro-portional for both sets (2) calculate the entire number ofmembers (proportional and nonproportional) and split thenumber of common members (1) by the entire number ofmembers (2) and (3) multiply by 100 [31]

33 Classification Results e binary classification wasutilized to analyze subcortical volume and cortical thicknessvolume of AD versus HC and EMCI versus LMCI subjectgroups e results of classification are demonstrated inTable 5 and all results are illustrated in Figures 4(a) and 4(b)Kappa and Jaccard are shown in Figures 4(c) and 4(d) Allprocesses were performed in a 64-bit Python 36 environ-ment on Intel Core i7-8700 at 320Hz and 16GB of RAMrunning Ubuntu 1604

331 Binary Classification AD versus HC and EMCI versusLMCI According to the four datasets two binary group-smdashAD versus HC and EMCI versus LMCImdashwere classifiedwith subcortical volume and cortical thickness features byutilizing the RBF-SVM classification algorithm and theresult is illustrated in Table 5

(1) AD versus HC e result of classification for ADversus HC is given in Table 5 and Figures 4(a) and 4(c) Inevery classification situation the database was divided intotwo separations in a 70 30 ratio Each dataset shows betterresults for kappa and Jaccard statistics in the RBF-SVMclassification technique In the kappa statistics case theGARD dataset cortical thickness feature gives the highestinterrater reliability of 09342 and the ARWIBO dataset

subcortical volume feature has 08939 In the Jaccard sta-tistics case the GARDdataset cortical thickness is 09091 thehighest and the ARWIBO dataset subcortical volume fea-ture is 08889 Moreover the GARD cortical thickness has agood F1 score precision specificity and sensitivity (98189687 9524 and 9818 respectively)

(2) EMCI versus LMCI e classification performancefor EMCI versus LMCI is shown in Table 5 and Figures 4(b)and 4(d) Similarly the kappa statistics for the GARD datasetsubcortical volume feature shows the highestresultmdash09043mdashand the ARWIBO dataset cortical feature is08979 e Jaccard statistic for the GARD dataset subcor-tical volume feature had the highest resultmdash09186 eNACC dataset subcortical volume feature is 09167 Fur-thermore the GARD dataset subcortical volume feature hasbetter results with an F1 score precision specificity andsensitivity of 9645 100 100 and 9275 respectivelycompared to the cortical thickness

332 Comparison with Related Recently Published MethodsAs mentioned above in the proposed method four datasetswere used GARD NACC ARWIBO and ADNI Except forthe GARD dataset all are available for the public andanyone can download them e websites available areNACC (httpswwwalzwashingtonedu) ARWIBO(httpswwwgaaindataorgpartnerARWIBO) and ADNI(httpadniloniuscedu)

ere are 701 subjects 326 (GARD) 131 (NACC) 121(ARWIBO) and 123 (ADNI) e ages of all subjects aremore than 47 and converting of mild cognitive stage to ADbetween 0 and 18 months For segmentation the T1-weighted sMRI imaging modality was utilized from eachdataset Moreover the results of the proposed method werecompared to recently published results as shown in Tables 6and 7

100

90

80

70

60

50

40

30

20

10

0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 39 41 43 45 49 51 53 55 57 59 61 63 65 67 69 71 73 75 79 81 83

PCA for AD vs HC NRCD dataset

Figure 3 Number of principal components for AD versus HC

8 Journal of Healthcare Engineering

70

75

80

85

90

95

100

AD

NI-

C

AD

NI-

S

ARW

IBO

-C

ARW

IBO

-S

NRC

D-C

NRC

D-S

NAC

C-C

NAC

C-S

AD vs HC

AUCACCSEN

SPECPREF1

(a)

AUCACCSEN

SPECPREF1

70

75

80

85

90

95

100EMCI vs LMCI

AD

NI-

C

AD

NI-

S

ARW

IBO

-C

ARW

IBO

-S

NRC

D-C

NRC

D-S

NAC

C-C

NAC

C-S

(b)

078

083

088

093

098AD vs HC

KappaJaccard

AD

NI-

C

AD

NI-

S

ARW

IBO

-C

ARW

IBO

-S

NRC

D-C

NRC

D-S

NAC

C-C

NAC

C-S

(c)

KappaJaccard

07

075

08

085

09

095EMCI vs LMCI

AD

NI-

C

AD

NI-

S

ARW

IBO

-C

ARW

IBO

-S

NRC

D-C

NRC

D-S

NAC

C-C

NAC

C-S

(d)

Figure 4 Classification reports of four datasetsmdashADNI-C (cortical) ADNI-S (subcortical) ARWIBO-C (cortical) ARWIBO-S (sub-cortical) NRCD-C (cortical) NRCD-S (subcortical) NACC-C (cortical) and NACC-S (subcortical)mdashwith measurement performance(AUC accuracy sensitivity specificity precision and F1 score) (a) AD versus HC (b) EMCI versus LMCI classification reports of fourdatasets with measurements of kappa and Jaccard (c) AD versus HC and (d) EMCI versus LMCI

Table 5 Result of four datasets for the subcortical and cortical parts for AD versus HC and EMCI versus LMCI

AD vs HC Classifier AUC ACC SEN SPEC PRE F1 Kappa JaccardADNI cortical

SVM-RBF

9167 9157 8182 100 100 90 08108 08333ADNI subcortical 9045 9048 9091 90 9091 9091 08091 08182ARWIBO cortical 8944 8947 90 8889 90 90 07889 08ARWIBO subcortical 9545 9474 100 8889 9091 9524 08939 08889NRCD cortical 975 9737 9818 9524 9687 9818 09342 09091NRCD subcortical 9571 9342 963 8636 9455 9541 08379 07917NACC cortical 9688 9524 100 8333 9375 9677 08772 08333NACC subcortical 9286 9456 9333 100 100 9655 08889 08571EMCI vs LMCI Classifier AUC ACC SEN SPEC PRE F1 Kappa JaccardADNI cortical

SVM-RBF

8175 8125 75 875 8571 80 0725 07158ADNI subcortical 8889 875 7778 100 100 875 07538 07778ARWIBO cortical 9444 9324 90 100 9566 9474 08979 08889ARWIBO subcortical 95 9487 8889 9275 100 9412 08889 09012NRCD cortical 9286 9091 9145 100 100 8889 08136 08271NRCD subcortical 9643 9545 9275 100 100 9645 09043 09186NACC cortical 9167 8947 8778 8957 9024 875 07989 08133NACC subcortical 9583 9474 9256 100 100 9333 08902 09167

Journal of Healthcare Engineering 9

Tripathi et al [13] proposed a method that used the RBFand linear SVM classifier for the automated pipeline whichdistinguishes subjects between AD LMCI EMCI and HCwith subcortical and hippocampal features gained fromspherical harmonics (SPHARM-PDM) process by utilizingthe ADNI dataset e combination of voxel and SPHARMfeatures shows 8875 accuracy 8310 sensitivity and9158 specificity for an AD versus HC group with a linearkernel SVM whereas for EMCI versus LMCI group thesame combined features show 7095 accuracy 7556sensitivity and 6547 specificity with linear SVM Likewiseanother study by Zhang et al [19] used an SVM with nestedcross-validation to distinguish the features into two groupsto gain balanced results e slow-5 frequency band shows8387 accuracy 8621 sensitivity and 8182 specificityfor EMCI versus LMCI cohort by using the ADNI datasetGorji and Naima [17] have utilized CNN deep learningalgorithms in their pipeline and utilizing it they have ob-tained a 93 accuracy 9148 sensitivity and 9482specificity for EMCI versus LMCI using sagittal featuresfrom anMRI image Furthermore Nozadi and Kadoury [15]compared the FDG and AV-45 biomarkers of PET imageand then employed RBF-SVM and RF for distinguishingAD NC EMCI and LMCI with six groups by using ADNIdataset eir approach showed accuracies of 917 and912 for AD versus NC each of RBF-SVM and RF withFDG-PET image modality On the other hand AV45 il-lustrates accuracies of 908 and 879 for AD versus NCwith RBF-SVM and RF For EMCI versus LMCI FDG-PETprovides better 539 641 accuracies compared withAV45 results in RBF-SVM and RF classifier Gupta et al [5]proposed a method utilizing the GARD dataset as a known

private dataset and the OASIS dataset was used for com-parison e four classifiers were utilized for binary andtertiary classification and achieved better results accordingto the classifier For AD versus HC the softmax classifiershowed the highest accuracy of 9934 100 specificity anda precision of 100 For the HC versus mAD case the SVMclassifier had an accuracy of 992 and a specificity andprecision of 100 e mAD versus aAD SVM providedgood results as well such as a 9777 accuracy 100sensitivity and 9795 F1 score In tertiary classificationSVM had the best accuracy of 9942 9918 sensitivityand 9999 precision in AD versus HC versus mAD In ADversus HC versus aAD the NB classifier had 9653 ac-curacy 9588 sensitivity and 9764 specificity

4 Discussion

In this research the RBF-SVM method was employed forclassification of subjects with AD versus HC and EMCIversus LMCI based on anatomical T1-weighted sMRIeproposed method is shown in Figure 3 e subjects weredivided into a ratio of 70 30 for training and testingpurposes before being passed to the classifier en thedimensionality reduction method was utilized by usingthe PCA function from the scikit-learn library Subse-quently for the SVM classifier the optimal structurevalue was regulated by employing SKF-CV and a gridsearch en the above features were utilized to instructthe SVM classifier using the training dataset and thetesting dataset was assessed using the training method todetermine the performance e proposed method showsmore than 90 accuracy for all datasets based on cortical

Table 7 Comparison of recently published works

EMCI vs LMCIYears Approach Dataset ACC SEN SPEC Classifier2017 Tripathi et al [13] ADNI 7029 7395 6601 RBF-SVM2018 Nozadi and Kadoury [15] ADNI 676 701 707 RBF-SVM2018 Gupta et al [5] NRCD 9555 100 909 Softmax2019 Zhang et al [19] ADNI 8387 8621 8182 RBF-SVM2019 Gorji and Naima [17] ADNI 9300 9148 9482 CNN

2019 Proposed method

NRCD 9545 9275 100

RBF-SVMNACC 9474 9256 100ARWIBO 9487 8889 9275ADNI 8750 7778 100

Table 6 Comparison of recently published works

AD vs HCYears Approach Dataset ACC SEN SPEC Classifier2017 Tripathi et al [13] ADNI 8598 7555 9030 RBF-SVM2018 Nozadi and Kadoury [15] ADNI 893 888 859 RBF-SVM

2018 Gupta et al [5] NRCD 9934 9814 100 SoftmaxOASIS 9840 9375 100

2019 Proposed method

NRCD 9737 9818 9524

RBF-SVMNACC 9524 100 8333ARWIBO 9474 100 8889ADNI 9157 8182 100

10 Journal of Healthcare Engineering

and subcortical features in the AD versus HC groups Inthe GARD dataset case cortical thickness had the bestaccuracy 9737 a sensitivity of 9818 and an F1 scoreof 9818 Moreover kappa and Jaccard statistical ana-lyses resulted in more than 080 for all datasets in ADversus HC binary classification

In the EMCI versus LMCI group GARD subcorticalvolumes had the highest accuracy among all data-setsmdash9545 accuracy 100 precision and 100 speci-ficity e Cohen kappa and Jaccard statistical volumeproduced good results as well Moreover GARD had a goodAUC for the subcortical volume and cortical thickness ofboth groups as given in Figure 5 In the ADNI dataset case(AD versus HC) the cortical thickness showed 100specificity 100 precision and an accuracy of 9157 eEMCI versus LMCI subcortical volume had a specificity andprecision of 100 In the ARWIBO dataset case (AD versusHC) the subcortical volume had 100 sensitivity 9474

accuracy and 9524 F1 score but for EMCI versus LMCIthe cortical thickness had a specificity of 100 precision of9566 and 9474 F1 score Furthermore in the NACCdataset case (AD versus HC) the subcortical volume was100 with 100 specificity and precision and a 9655 F1score Likewise the EMCI versus LMCI subcortical volumehad a specificity and precision of 100 and an accuracy of9474 In both binary classifications the proposed methodobtained good results compared to those proposed in otherworks In addition the Cohen kappa and Jaccard resultsdemonstrate excellent statistical analysis for ADNI GARDARWIBO and NACC

5 Conclusions

A novel method for automatic classification AD from MCI(early or late converted to AD) and an HC was developed byutilizing the features of cortical thickness and subcortical

ROC curve for AD vs NC NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

100804 0600 021 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09750)

(a)

ROC curve for AD vs NC NRCD_Subcortical

02 04 06 08 10001 minus specificity (false positive rate)

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

ROC_curve_APOE (AUC1 = 09571)

(b)

ROC curve for EMCI vs LMCI NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

02 04 06 08 10001 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09286)

(c)

ROC curve for EMCI vs LMCI NRCD_Subcortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

1000 06 0802 041 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09643)

(d)

Figure 5 AUC-ROC performance for the GARD dataset (a) AD versus HC cortical thickness (b) AD versus HC subcortical (c) EMCIversus LMCI cortical and (d) EMCI versus LMCI subcortical

Journal of Healthcare Engineering 11

volumes e features were extracted from a MALPEMtoolbox and classification was performed by the RBF-SVMclassifier based on the GARD dataset ey were comparedto three datasets e results of this research prove that theproposed approach is efficient for future clinical predictionbetween the subcortical view of EMCI versus LMCI and thecortical view of AD versus HC which are shown in Table 5Based on the subcortical volume and cortical thickness theproposed method obtained different accuracies according tothe different databases such as the GARD and NACCdataset AD versus HC group cortical thickness accuracies of9737 and 9524 e subcortical volumes of ARWIBOand NACC for the AD versus HC group were 9474 and9456

In this research only the subcortical volume and corticalthickness features were considered for the classificationprocedure In the future it is planned to examine classifiermultimodality features of the brain biomarkers and non-imaging biomarkers for diagnosis of AD Furthermore weare planning a future work to compare state-of-the-art andrecently published methods with different available datasetsfor early prediction of AD

Data Availability

e Gwangju Alzheimerrsquos disease and Related Dementia(GARD) dataset was used to support the findings of thisstudy e GARD is a private dataset that was generated inChosun University hospitals and it belongs to ChosunUniversity e authors cannot share it or make it availableonline for privacy reasons Moreover to compare theproposed approach with other datasets the authorsdownloaded the NACC dataset from httpswwwalzwashingtonedu the ARWIBO one from httpswwwgaaindataorg and the ADNI one from httpadniloniuscedu

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Saidjalol Toshkhujaev and Kun Ho Lee contributed equallyto this work

Acknowledgments

is work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Korea gov-ernment (MSIT) (nos NRF-2019R1A4A1029769 and NRF-2019R1F1A1060166) and this research was supported byKBRI basic research program through Korea Brain ResearchInstitute funded by Ministry of Science and ICT (20-BR-03-02) Data collection and sharing for this project was funded bythe Alzheimerrsquos Disease Neuroimaging Initiative (ADNI) (Na-tional Institutes of Health Grant U01 AG024904) and DODADNI (Department of Defense Award No W81XWH-12-2ndash0012) As such the investigators within the ADNI contributed

to the design and implementation of ADNI andor provideddata but did not participate in analysis orwriting of this report Acomplete listing of ADNI investigators can be found at httpadniloniusceduwp-contentuploadshow_to_applyADNI_Acknowledgment_Listpdf ADNI is funded by theNational Institute on Aging the National Institute ofBiomedical Imaging and Bioengineering and throughgenerous contributions from the following AbbVie Alz-heimerrsquos Association Alzheimerrsquos Drug Discovery Foun-dation Araclon Biotech BioClinica Inc Biogen Bristol-Myers Squibb Company CereSpir Inc Cogstate EisaiInc Elan Pharmaceuticals Inc Eli Lilly and CompanyEuroImmun F Hoffmann-La Roche Ltd and its affiliatedcompany Genentech Inc Fujirebio GE HealthcareIXICO Ltd Janssen Alzheimer Immunotherapy Researchamp Development LLC Johnson amp Johnson PharmaceuticalResearch amp Development LLC Lumosity LundbeckMerck amp Co Inc Meso Scale Diagnostics LLC NeuroRxResearch Neurotrack Technologies Novartis Pharma-ceuticals Corporation Pfizer Inc Piramal Imaging Serv-ier Takeda Pharmaceutical Company and Transitionerapeuticse Canadian Institutes of Health Research isproviding funds to support ADNI clinical sites in CanadaPrivate sector contributions are facilitated by the Foun-dation for the National Institutes of Health (httpwwwfnihorg) e grantee organization is the Northern Cal-ifornia Institute for Research and Education and the studyis coordinated by the Alzheimerrsquos erapeutic ResearchInstitute at the University of Southern California ADNIdata are disseminated by the Laboratory for Neuroimagingat the University of Southern California Data used inpreparation of this article were obtained from the NationalAlzheimerrsquos coordination center (NACC) database funded byNIANIH Grant U01 AG016976 NACC data are contributedby the NIA funded ADCs P30 AG019610 (PI Eric ReimanMD) P30 AG013846 (PI Neil Kowall MD) P50 AG008702(PI Scott Small MD) P50 AG025688 (PI Allan Levey MDPhD) P50 AG047266 (PI Todd Golde MD PhD) P30AG010133 (PI Andrew Saykin PsyD) P50 AG005146 (PIMarilyn Albert PhD) P50 AG005134 (PI Bradley HymanMD PhD) P50 AG016574 (PI Ronald Petersen MD PhD)P50 AG005138 (PI Mary Sano PhD) P30 AG008051 (PIomas Wisniewski MD) P30 AG013854 (PI Robert VassarPhD) P30 AG008017 (PI Jeffrey Kaye MD) P30 AG010161(PI David Bennett MD) P50 AG047366 (PI Victor Hen-derson MD MS) P30 AG010129 (PI Charles DeCarli MD)P50 AG016573 (PI Frank LaFerla PhD) P50 AG005131 (PIJames Brewer MD PhD) P50 AG023501 (PI Bruce MillerMD) P30 AG035982 (PI Russell Swerdlow MD) P30AG028383 (PI Linda Van Eldik PhD) P30 AG053760 (PIHenry Paulson MD PhD) P30 AG010124 (PI John Troja-nowski MD PhD) P50 AG005133 (PI Oscar Lopez MD)P50 AG005142 (PI Helena Chui MD) P30 AG012300 (PIRoger Rosenberg MD) P30 AG049638 (PI Suzanne CraftPhD) P50 AG005136 (PI omas Grabowski MD) P50AG033514 (PI Sanjay Asthana MD FRCP) P50 AG005681(PI John Morris MD) and P50 AG047270 (PI StephenStrittmatter MD PhD) ARWiBo data (httpwwwarwiboit)was obtained from NeuGRID4You initiative funded by the

12 Journal of Healthcare Engineering

Tripathi et al [13] proposed a method that used the RBFand linear SVM classifier for the automated pipeline whichdistinguishes subjects between AD LMCI EMCI and HCwith subcortical and hippocampal features gained fromspherical harmonics (SPHARM-PDM) process by utilizingthe ADNI dataset e combination of voxel and SPHARMfeatures shows 8875 accuracy 8310 sensitivity and9158 specificity for an AD versus HC group with a linearkernel SVM whereas for EMCI versus LMCI group thesame combined features show 7095 accuracy 7556sensitivity and 6547 specificity with linear SVM Likewiseanother study by Zhang et al [19] used an SVM with nestedcross-validation to distinguish the features into two groupsto gain balanced results e slow-5 frequency band shows8387 accuracy 8621 sensitivity and 8182 specificityfor EMCI versus LMCI cohort by using the ADNI datasetGorji and Naima [17] have utilized CNN deep learningalgorithms in their pipeline and utilizing it they have ob-tained a 93 accuracy 9148 sensitivity and 9482specificity for EMCI versus LMCI using sagittal featuresfrom anMRI image Furthermore Nozadi and Kadoury [15]compared the FDG and AV-45 biomarkers of PET imageand then employed RBF-SVM and RF for distinguishingAD NC EMCI and LMCI with six groups by using ADNIdataset eir approach showed accuracies of 917 and912 for AD versus NC each of RBF-SVM and RF withFDG-PET image modality On the other hand AV45 il-lustrates accuracies of 908 and 879 for AD versus NCwith RBF-SVM and RF For EMCI versus LMCI FDG-PETprovides better 539 641 accuracies compared withAV45 results in RBF-SVM and RF classifier Gupta et al [5]proposed a method utilizing the GARD dataset as a known

private dataset and the OASIS dataset was used for com-parison e four classifiers were utilized for binary andtertiary classification and achieved better results accordingto the classifier For AD versus HC the softmax classifiershowed the highest accuracy of 9934 100 specificity anda precision of 100 For the HC versus mAD case the SVMclassifier had an accuracy of 992 and a specificity andprecision of 100 e mAD versus aAD SVM providedgood results as well such as a 9777 accuracy 100sensitivity and 9795 F1 score In tertiary classificationSVM had the best accuracy of 9942 9918 sensitivityand 9999 precision in AD versus HC versus mAD In ADversus HC versus aAD the NB classifier had 9653 ac-curacy 9588 sensitivity and 9764 specificity

4 Discussion

In this research the RBF-SVM method was employed forclassification of subjects with AD versus HC and EMCIversus LMCI based on anatomical T1-weighted sMRIeproposed method is shown in Figure 3 e subjects weredivided into a ratio of 70 30 for training and testingpurposes before being passed to the classifier en thedimensionality reduction method was utilized by usingthe PCA function from the scikit-learn library Subse-quently for the SVM classifier the optimal structurevalue was regulated by employing SKF-CV and a gridsearch en the above features were utilized to instructthe SVM classifier using the training dataset and thetesting dataset was assessed using the training method todetermine the performance e proposed method showsmore than 90 accuracy for all datasets based on cortical

Table 7 Comparison of recently published works

EMCI vs LMCIYears Approach Dataset ACC SEN SPEC Classifier2017 Tripathi et al [13] ADNI 7029 7395 6601 RBF-SVM2018 Nozadi and Kadoury [15] ADNI 676 701 707 RBF-SVM2018 Gupta et al [5] NRCD 9555 100 909 Softmax2019 Zhang et al [19] ADNI 8387 8621 8182 RBF-SVM2019 Gorji and Naima [17] ADNI 9300 9148 9482 CNN

2019 Proposed method

NRCD 9545 9275 100

RBF-SVMNACC 9474 9256 100ARWIBO 9487 8889 9275ADNI 8750 7778 100

Table 6 Comparison of recently published works

AD vs HCYears Approach Dataset ACC SEN SPEC Classifier2017 Tripathi et al [13] ADNI 8598 7555 9030 RBF-SVM2018 Nozadi and Kadoury [15] ADNI 893 888 859 RBF-SVM

2018 Gupta et al [5] NRCD 9934 9814 100 SoftmaxOASIS 9840 9375 100

2019 Proposed method

NRCD 9737 9818 9524

RBF-SVMNACC 9524 100 8333ARWIBO 9474 100 8889ADNI 9157 8182 100

10 Journal of Healthcare Engineering

and subcortical features in the AD versus HC groups Inthe GARD dataset case cortical thickness had the bestaccuracy 9737 a sensitivity of 9818 and an F1 scoreof 9818 Moreover kappa and Jaccard statistical ana-lyses resulted in more than 080 for all datasets in ADversus HC binary classification

In the EMCI versus LMCI group GARD subcorticalvolumes had the highest accuracy among all data-setsmdash9545 accuracy 100 precision and 100 speci-ficity e Cohen kappa and Jaccard statistical volumeproduced good results as well Moreover GARD had a goodAUC for the subcortical volume and cortical thickness ofboth groups as given in Figure 5 In the ADNI dataset case(AD versus HC) the cortical thickness showed 100specificity 100 precision and an accuracy of 9157 eEMCI versus LMCI subcortical volume had a specificity andprecision of 100 In the ARWIBO dataset case (AD versusHC) the subcortical volume had 100 sensitivity 9474

accuracy and 9524 F1 score but for EMCI versus LMCIthe cortical thickness had a specificity of 100 precision of9566 and 9474 F1 score Furthermore in the NACCdataset case (AD versus HC) the subcortical volume was100 with 100 specificity and precision and a 9655 F1score Likewise the EMCI versus LMCI subcortical volumehad a specificity and precision of 100 and an accuracy of9474 In both binary classifications the proposed methodobtained good results compared to those proposed in otherworks In addition the Cohen kappa and Jaccard resultsdemonstrate excellent statistical analysis for ADNI GARDARWIBO and NACC

5 Conclusions

A novel method for automatic classification AD from MCI(early or late converted to AD) and an HC was developed byutilizing the features of cortical thickness and subcortical

ROC curve for AD vs NC NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

100804 0600 021 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09750)

(a)

ROC curve for AD vs NC NRCD_Subcortical

02 04 06 08 10001 minus specificity (false positive rate)

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

ROC_curve_APOE (AUC1 = 09571)

(b)

ROC curve for EMCI vs LMCI NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

02 04 06 08 10001 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09286)

(c)

ROC curve for EMCI vs LMCI NRCD_Subcortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

1000 06 0802 041 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09643)

(d)

Figure 5 AUC-ROC performance for the GARD dataset (a) AD versus HC cortical thickness (b) AD versus HC subcortical (c) EMCIversus LMCI cortical and (d) EMCI versus LMCI subcortical

Journal of Healthcare Engineering 11

volumes e features were extracted from a MALPEMtoolbox and classification was performed by the RBF-SVMclassifier based on the GARD dataset ey were comparedto three datasets e results of this research prove that theproposed approach is efficient for future clinical predictionbetween the subcortical view of EMCI versus LMCI and thecortical view of AD versus HC which are shown in Table 5Based on the subcortical volume and cortical thickness theproposed method obtained different accuracies according tothe different databases such as the GARD and NACCdataset AD versus HC group cortical thickness accuracies of9737 and 9524 e subcortical volumes of ARWIBOand NACC for the AD versus HC group were 9474 and9456

In this research only the subcortical volume and corticalthickness features were considered for the classificationprocedure In the future it is planned to examine classifiermultimodality features of the brain biomarkers and non-imaging biomarkers for diagnosis of AD Furthermore weare planning a future work to compare state-of-the-art andrecently published methods with different available datasetsfor early prediction of AD

Data Availability

e Gwangju Alzheimerrsquos disease and Related Dementia(GARD) dataset was used to support the findings of thisstudy e GARD is a private dataset that was generated inChosun University hospitals and it belongs to ChosunUniversity e authors cannot share it or make it availableonline for privacy reasons Moreover to compare theproposed approach with other datasets the authorsdownloaded the NACC dataset from httpswwwalzwashingtonedu the ARWIBO one from httpswwwgaaindataorg and the ADNI one from httpadniloniuscedu

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Saidjalol Toshkhujaev and Kun Ho Lee contributed equallyto this work

Acknowledgments

is work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Korea gov-ernment (MSIT) (nos NRF-2019R1A4A1029769 and NRF-2019R1F1A1060166) and this research was supported byKBRI basic research program through Korea Brain ResearchInstitute funded by Ministry of Science and ICT (20-BR-03-02) Data collection and sharing for this project was funded bythe Alzheimerrsquos Disease Neuroimaging Initiative (ADNI) (Na-tional Institutes of Health Grant U01 AG024904) and DODADNI (Department of Defense Award No W81XWH-12-2ndash0012) As such the investigators within the ADNI contributed

to the design and implementation of ADNI andor provideddata but did not participate in analysis orwriting of this report Acomplete listing of ADNI investigators can be found at httpadniloniusceduwp-contentuploadshow_to_applyADNI_Acknowledgment_Listpdf ADNI is funded by theNational Institute on Aging the National Institute ofBiomedical Imaging and Bioengineering and throughgenerous contributions from the following AbbVie Alz-heimerrsquos Association Alzheimerrsquos Drug Discovery Foun-dation Araclon Biotech BioClinica Inc Biogen Bristol-Myers Squibb Company CereSpir Inc Cogstate EisaiInc Elan Pharmaceuticals Inc Eli Lilly and CompanyEuroImmun F Hoffmann-La Roche Ltd and its affiliatedcompany Genentech Inc Fujirebio GE HealthcareIXICO Ltd Janssen Alzheimer Immunotherapy Researchamp Development LLC Johnson amp Johnson PharmaceuticalResearch amp Development LLC Lumosity LundbeckMerck amp Co Inc Meso Scale Diagnostics LLC NeuroRxResearch Neurotrack Technologies Novartis Pharma-ceuticals Corporation Pfizer Inc Piramal Imaging Serv-ier Takeda Pharmaceutical Company and Transitionerapeuticse Canadian Institutes of Health Research isproviding funds to support ADNI clinical sites in CanadaPrivate sector contributions are facilitated by the Foun-dation for the National Institutes of Health (httpwwwfnihorg) e grantee organization is the Northern Cal-ifornia Institute for Research and Education and the studyis coordinated by the Alzheimerrsquos erapeutic ResearchInstitute at the University of Southern California ADNIdata are disseminated by the Laboratory for Neuroimagingat the University of Southern California Data used inpreparation of this article were obtained from the NationalAlzheimerrsquos coordination center (NACC) database funded byNIANIH Grant U01 AG016976 NACC data are contributedby the NIA funded ADCs P30 AG019610 (PI Eric ReimanMD) P30 AG013846 (PI Neil Kowall MD) P50 AG008702(PI Scott Small MD) P50 AG025688 (PI Allan Levey MDPhD) P50 AG047266 (PI Todd Golde MD PhD) P30AG010133 (PI Andrew Saykin PsyD) P50 AG005146 (PIMarilyn Albert PhD) P50 AG005134 (PI Bradley HymanMD PhD) P50 AG016574 (PI Ronald Petersen MD PhD)P50 AG005138 (PI Mary Sano PhD) P30 AG008051 (PIomas Wisniewski MD) P30 AG013854 (PI Robert VassarPhD) P30 AG008017 (PI Jeffrey Kaye MD) P30 AG010161(PI David Bennett MD) P50 AG047366 (PI Victor Hen-derson MD MS) P30 AG010129 (PI Charles DeCarli MD)P50 AG016573 (PI Frank LaFerla PhD) P50 AG005131 (PIJames Brewer MD PhD) P50 AG023501 (PI Bruce MillerMD) P30 AG035982 (PI Russell Swerdlow MD) P30AG028383 (PI Linda Van Eldik PhD) P30 AG053760 (PIHenry Paulson MD PhD) P30 AG010124 (PI John Troja-nowski MD PhD) P50 AG005133 (PI Oscar Lopez MD)P50 AG005142 (PI Helena Chui MD) P30 AG012300 (PIRoger Rosenberg MD) P30 AG049638 (PI Suzanne CraftPhD) P50 AG005136 (PI omas Grabowski MD) P50AG033514 (PI Sanjay Asthana MD FRCP) P50 AG005681(PI John Morris MD) and P50 AG047270 (PI StephenStrittmatter MD PhD) ARWiBo data (httpwwwarwiboit)was obtained from NeuGRID4You initiative funded by the

12 Journal of Healthcare Engineering

and subcortical features in the AD versus HC groups Inthe GARD dataset case cortical thickness had the bestaccuracy 9737 a sensitivity of 9818 and an F1 scoreof 9818 Moreover kappa and Jaccard statistical ana-lyses resulted in more than 080 for all datasets in ADversus HC binary classification

In the EMCI versus LMCI group GARD subcorticalvolumes had the highest accuracy among all data-setsmdash9545 accuracy 100 precision and 100 speci-ficity e Cohen kappa and Jaccard statistical volumeproduced good results as well Moreover GARD had a goodAUC for the subcortical volume and cortical thickness ofboth groups as given in Figure 5 In the ADNI dataset case(AD versus HC) the cortical thickness showed 100specificity 100 precision and an accuracy of 9157 eEMCI versus LMCI subcortical volume had a specificity andprecision of 100 In the ARWIBO dataset case (AD versusHC) the subcortical volume had 100 sensitivity 9474

accuracy and 9524 F1 score but for EMCI versus LMCIthe cortical thickness had a specificity of 100 precision of9566 and 9474 F1 score Furthermore in the NACCdataset case (AD versus HC) the subcortical volume was100 with 100 specificity and precision and a 9655 F1score Likewise the EMCI versus LMCI subcortical volumehad a specificity and precision of 100 and an accuracy of9474 In both binary classifications the proposed methodobtained good results compared to those proposed in otherworks In addition the Cohen kappa and Jaccard resultsdemonstrate excellent statistical analysis for ADNI GARDARWIBO and NACC

5 Conclusions

A novel method for automatic classification AD from MCI(early or late converted to AD) and an HC was developed byutilizing the features of cortical thickness and subcortical

ROC curve for AD vs NC NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

100804 0600 021 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09750)

(a)

ROC curve for AD vs NC NRCD_Subcortical

02 04 06 08 10001 minus specificity (false positive rate)

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

ROC_curve_APOE (AUC1 = 09571)

(b)

ROC curve for EMCI vs LMCI NRCD_Cortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

02 04 06 08 10001 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09286)

(c)

ROC curve for EMCI vs LMCI NRCD_Subcortical

0001020304050607080910

Sens

itivi

ty (t

rue p

ositi

ve ra

te)

1000 06 0802 041 minus specificity (false positive rate)

ROC_curve_APOE (AUC1 = 09643)

(d)

Figure 5 AUC-ROC performance for the GARD dataset (a) AD versus HC cortical thickness (b) AD versus HC subcortical (c) EMCIversus LMCI cortical and (d) EMCI versus LMCI subcortical

Journal of Healthcare Engineering 11

volumes e features were extracted from a MALPEMtoolbox and classification was performed by the RBF-SVMclassifier based on the GARD dataset ey were comparedto three datasets e results of this research prove that theproposed approach is efficient for future clinical predictionbetween the subcortical view of EMCI versus LMCI and thecortical view of AD versus HC which are shown in Table 5Based on the subcortical volume and cortical thickness theproposed method obtained different accuracies according tothe different databases such as the GARD and NACCdataset AD versus HC group cortical thickness accuracies of9737 and 9524 e subcortical volumes of ARWIBOand NACC for the AD versus HC group were 9474 and9456

In this research only the subcortical volume and corticalthickness features were considered for the classificationprocedure In the future it is planned to examine classifiermultimodality features of the brain biomarkers and non-imaging biomarkers for diagnosis of AD Furthermore weare planning a future work to compare state-of-the-art andrecently published methods with different available datasetsfor early prediction of AD

Data Availability

e Gwangju Alzheimerrsquos disease and Related Dementia(GARD) dataset was used to support the findings of thisstudy e GARD is a private dataset that was generated inChosun University hospitals and it belongs to ChosunUniversity e authors cannot share it or make it availableonline for privacy reasons Moreover to compare theproposed approach with other datasets the authorsdownloaded the NACC dataset from httpswwwalzwashingtonedu the ARWIBO one from httpswwwgaaindataorg and the ADNI one from httpadniloniuscedu

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Saidjalol Toshkhujaev and Kun Ho Lee contributed equallyto this work

Acknowledgments

is work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Korea gov-ernment (MSIT) (nos NRF-2019R1A4A1029769 and NRF-2019R1F1A1060166) and this research was supported byKBRI basic research program through Korea Brain ResearchInstitute funded by Ministry of Science and ICT (20-BR-03-02) Data collection and sharing for this project was funded bythe Alzheimerrsquos Disease Neuroimaging Initiative (ADNI) (Na-tional Institutes of Health Grant U01 AG024904) and DODADNI (Department of Defense Award No W81XWH-12-2ndash0012) As such the investigators within the ADNI contributed

to the design and implementation of ADNI andor provideddata but did not participate in analysis orwriting of this report Acomplete listing of ADNI investigators can be found at httpadniloniusceduwp-contentuploadshow_to_applyADNI_Acknowledgment_Listpdf ADNI is funded by theNational Institute on Aging the National Institute ofBiomedical Imaging and Bioengineering and throughgenerous contributions from the following AbbVie Alz-heimerrsquos Association Alzheimerrsquos Drug Discovery Foun-dation Araclon Biotech BioClinica Inc Biogen Bristol-Myers Squibb Company CereSpir Inc Cogstate EisaiInc Elan Pharmaceuticals Inc Eli Lilly and CompanyEuroImmun F Hoffmann-La Roche Ltd and its affiliatedcompany Genentech Inc Fujirebio GE HealthcareIXICO Ltd Janssen Alzheimer Immunotherapy Researchamp Development LLC Johnson amp Johnson PharmaceuticalResearch amp Development LLC Lumosity LundbeckMerck amp Co Inc Meso Scale Diagnostics LLC NeuroRxResearch Neurotrack Technologies Novartis Pharma-ceuticals Corporation Pfizer Inc Piramal Imaging Serv-ier Takeda Pharmaceutical Company and Transitionerapeuticse Canadian Institutes of Health Research isproviding funds to support ADNI clinical sites in CanadaPrivate sector contributions are facilitated by the Foun-dation for the National Institutes of Health (httpwwwfnihorg) e grantee organization is the Northern Cal-ifornia Institute for Research and Education and the studyis coordinated by the Alzheimerrsquos erapeutic ResearchInstitute at the University of Southern California ADNIdata are disseminated by the Laboratory for Neuroimagingat the University of Southern California Data used inpreparation of this article were obtained from the NationalAlzheimerrsquos coordination center (NACC) database funded byNIANIH Grant U01 AG016976 NACC data are contributedby the NIA funded ADCs P30 AG019610 (PI Eric ReimanMD) P30 AG013846 (PI Neil Kowall MD) P50 AG008702(PI Scott Small MD) P50 AG025688 (PI Allan Levey MDPhD) P50 AG047266 (PI Todd Golde MD PhD) P30AG010133 (PI Andrew Saykin PsyD) P50 AG005146 (PIMarilyn Albert PhD) P50 AG005134 (PI Bradley HymanMD PhD) P50 AG016574 (PI Ronald Petersen MD PhD)P50 AG005138 (PI Mary Sano PhD) P30 AG008051 (PIomas Wisniewski MD) P30 AG013854 (PI Robert VassarPhD) P30 AG008017 (PI Jeffrey Kaye MD) P30 AG010161(PI David Bennett MD) P50 AG047366 (PI Victor Hen-derson MD MS) P30 AG010129 (PI Charles DeCarli MD)P50 AG016573 (PI Frank LaFerla PhD) P50 AG005131 (PIJames Brewer MD PhD) P50 AG023501 (PI Bruce MillerMD) P30 AG035982 (PI Russell Swerdlow MD) P30AG028383 (PI Linda Van Eldik PhD) P30 AG053760 (PIHenry Paulson MD PhD) P30 AG010124 (PI John Troja-nowski MD PhD) P50 AG005133 (PI Oscar Lopez MD)P50 AG005142 (PI Helena Chui MD) P30 AG012300 (PIRoger Rosenberg MD) P30 AG049638 (PI Suzanne CraftPhD) P50 AG005136 (PI omas Grabowski MD) P50AG033514 (PI Sanjay Asthana MD FRCP) P50 AG005681(PI John Morris MD) and P50 AG047270 (PI StephenStrittmatter MD PhD) ARWiBo data (httpwwwarwiboit)was obtained from NeuGRID4You initiative funded by the

12 Journal of Healthcare Engineering

volumes e features were extracted from a MALPEMtoolbox and classification was performed by the RBF-SVMclassifier based on the GARD dataset ey were comparedto three datasets e results of this research prove that theproposed approach is efficient for future clinical predictionbetween the subcortical view of EMCI versus LMCI and thecortical view of AD versus HC which are shown in Table 5Based on the subcortical volume and cortical thickness theproposed method obtained different accuracies according tothe different databases such as the GARD and NACCdataset AD versus HC group cortical thickness accuracies of9737 and 9524 e subcortical volumes of ARWIBOand NACC for the AD versus HC group were 9474 and9456

In this research only the subcortical volume and corticalthickness features were considered for the classificationprocedure In the future it is planned to examine classifiermultimodality features of the brain biomarkers and non-imaging biomarkers for diagnosis of AD Furthermore weare planning a future work to compare state-of-the-art andrecently published methods with different available datasetsfor early prediction of AD

Data Availability

e Gwangju Alzheimerrsquos disease and Related Dementia(GARD) dataset was used to support the findings of thisstudy e GARD is a private dataset that was generated inChosun University hospitals and it belongs to ChosunUniversity e authors cannot share it or make it availableonline for privacy reasons Moreover to compare theproposed approach with other datasets the authorsdownloaded the NACC dataset from httpswwwalzwashingtonedu the ARWIBO one from httpswwwgaaindataorg and the ADNI one from httpadniloniuscedu

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Saidjalol Toshkhujaev and Kun Ho Lee contributed equallyto this work

Acknowledgments

is work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Korea gov-ernment (MSIT) (nos NRF-2019R1A4A1029769 and NRF-2019R1F1A1060166) and this research was supported byKBRI basic research program through Korea Brain ResearchInstitute funded by Ministry of Science and ICT (20-BR-03-02) Data collection and sharing for this project was funded bythe Alzheimerrsquos Disease Neuroimaging Initiative (ADNI) (Na-tional Institutes of Health Grant U01 AG024904) and DODADNI (Department of Defense Award No W81XWH-12-2ndash0012) As such the investigators within the ADNI contributed

to the design and implementation of ADNI andor provideddata but did not participate in analysis orwriting of this report Acomplete listing of ADNI investigators can be found at httpadniloniusceduwp-contentuploadshow_to_applyADNI_Acknowledgment_Listpdf ADNI is funded by theNational Institute on Aging the National Institute ofBiomedical Imaging and Bioengineering and throughgenerous contributions from the following AbbVie Alz-heimerrsquos Association Alzheimerrsquos Drug Discovery Foun-dation Araclon Biotech BioClinica Inc Biogen Bristol-Myers Squibb Company CereSpir Inc Cogstate EisaiInc Elan Pharmaceuticals Inc Eli Lilly and CompanyEuroImmun F Hoffmann-La Roche Ltd and its affiliatedcompany Genentech Inc Fujirebio GE HealthcareIXICO Ltd Janssen Alzheimer Immunotherapy Researchamp Development LLC Johnson amp Johnson PharmaceuticalResearch amp Development LLC Lumosity LundbeckMerck amp Co Inc Meso Scale Diagnostics LLC NeuroRxResearch Neurotrack Technologies Novartis Pharma-ceuticals Corporation Pfizer Inc Piramal Imaging Serv-ier Takeda Pharmaceutical Company and Transitionerapeuticse Canadian Institutes of Health Research isproviding funds to support ADNI clinical sites in CanadaPrivate sector contributions are facilitated by the Foun-dation for the National Institutes of Health (httpwwwfnihorg) e grantee organization is the Northern Cal-ifornia Institute for Research and Education and the studyis coordinated by the Alzheimerrsquos erapeutic ResearchInstitute at the University of Southern California ADNIdata are disseminated by the Laboratory for Neuroimagingat the University of Southern California Data used inpreparation of this article were obtained from the NationalAlzheimerrsquos coordination center (NACC) database funded byNIANIH Grant U01 AG016976 NACC data are contributedby the NIA funded ADCs P30 AG019610 (PI Eric ReimanMD) P30 AG013846 (PI Neil Kowall MD) P50 AG008702(PI Scott Small MD) P50 AG025688 (PI Allan Levey MDPhD) P50 AG047266 (PI Todd Golde MD PhD) P30AG010133 (PI Andrew Saykin PsyD) P50 AG005146 (PIMarilyn Albert PhD) P50 AG005134 (PI Bradley HymanMD PhD) P50 AG016574 (PI Ronald Petersen MD PhD)P50 AG005138 (PI Mary Sano PhD) P30 AG008051 (PIomas Wisniewski MD) P30 AG013854 (PI Robert VassarPhD) P30 AG008017 (PI Jeffrey Kaye MD) P30 AG010161(PI David Bennett MD) P50 AG047366 (PI Victor Hen-derson MD MS) P30 AG010129 (PI Charles DeCarli MD)P50 AG016573 (PI Frank LaFerla PhD) P50 AG005131 (PIJames Brewer MD PhD) P50 AG023501 (PI Bruce MillerMD) P30 AG035982 (PI Russell Swerdlow MD) P30AG028383 (PI Linda Van Eldik PhD) P30 AG053760 (PIHenry Paulson MD PhD) P30 AG010124 (PI John Troja-nowski MD PhD) P50 AG005133 (PI Oscar Lopez MD)P50 AG005142 (PI Helena Chui MD) P30 AG012300 (PIRoger Rosenberg MD) P30 AG049638 (PI Suzanne CraftPhD) P50 AG005136 (PI omas Grabowski MD) P50AG033514 (PI Sanjay Asthana MD FRCP) P50 AG005681(PI John Morris MD) and P50 AG047270 (PI StephenStrittmatter MD PhD) ARWiBo data (httpwwwarwiboit)was obtained from NeuGRID4You initiative funded by the

12 Journal of Healthcare Engineering

European Commission (FP72007ndash2013) under grantagreement no 283562 e overall goal of ARWiBo is tocontribute thorough synergy with neuGRID (httpsneugrid2eu) to global data sharing and analysis in orderto develop effective therapies prevention methods and a curefor Alzheimerrsquo and other neurodegenerative diseases

References

[1] C P Ferri et al ldquoGlobal prevalence of dementia a Delphiconsensus studyrdquo e Lancet vol 366 no 9503 pp 2112ndash2117 2005

[2] B C Riedel M Daianu G Ver Steeg et al ldquoUncoveringbiologically coherent peripheral signatures of health and riskfor Alzheimerrsquos disease in the aging brainrdquo Frontiers in AgingNeuroscience vol 10 p 390 2018

[3] P B Rosenberg and C Lyketsos ldquoMild cognitive impairmentsearching for the prodrome of Alzheimerrsquos diseaserdquo WorldPsychiatry vol 7 no 2 pp 72ndash78 2008

[4] C Ledig et al ldquoMulti-class brain segmentation using atlaspropagation and EM-based refinementrdquo in Proceedings of the2012 9th IEEE International Symposium On Biomedical Im-aging (ISBI) pp 896ndash899 Barcelona Spain May 2012

[5] Y Gupta K H Lee K Y Choi J J Lee B C Kim andG-R Kwon ldquoAlzheimerrsquos disease diagnosis based on corticaland subcortical featuresrdquo Journal of Healthcare Engineeringvol 2019 Article ID 2492719 13 pages 2019

[6] Y Gupta K H Lee K Y Choi J J Lee B C Kim andG R Kwon ldquoEarly diagnosis of Alzheimerrsquos disease usingcombined features from voxel-based morphometry and cor-tical subcortical and hippocampus regions of MRI T1 brainimagesrdquo PLoS One vol 14 no 10 Article ID e0222446 2019

[7] K R Gray R Wolz R A Heckemann P AljabarA Hammers and D Rueckert ldquoMulti-region analysis oflongitudinal FDG-PET for the classification of Alzheimerrsquosdiseaserdquo NeuroImage vol 60 no 1 pp 221ndash229 2012

[8] H Hanyu T Sato K Hirao H Kanetaka T Iwamoto andK Koizumi ldquoe progression of cognitive deterioration andregional cerebral blood flow patterns in Alzheimerrsquos disease alongitudinal SPECT studyrdquo Journal of the Neurological Sci-ences vol 290 no 1-2 pp 96ndash101 2010

[9] J Barnes J W Bartlett L A van de Pol et al ldquoA meta-analysis of hippocampal atrophy rates in Alzheimerrsquos diseaserdquoNeurobiology of Aging vol 30 no 11 pp 1711ndash1723 2009

[10] C R Jack D A Bennett K Blennow et al ldquoNIA-AA Re-search Framework toward a biological definition of Alz-heimerrsquos diseaserdquo Alzheimerrsquos amp Dementia vol 14 no 4pp 535ndash562 2018

[11] E Braak and H Braak ldquoAlzheimerrsquos disease transientlydeveloping dendritic changes in pyramidal cells of sector CA1of the Ammonrsquos hornrdquo Acta Neuropathologica vol 93 no 4pp 323ndash325 1997

[12] B Magnin L Mesrob S Kinkingnehun et al ldquoSupport vectormachine-based classification of Alzheimerrsquos disease fromwhole-brain anatomical MRIrdquo Neuroradiology vol 51 no 2pp 73ndash83 2009

[13] S Tripathi S H Nozadi M Shakeri and S Kadoury ldquoldquoSub-cortical shape morphology and voxel-based features forAlzheimerrsquos disease classificationrdquo in Proceedings of the 2017IEEE 14th International Symposium on Biomedical Imaging(ISBI 2017) Melbourne Australia April 2017

[14] S Liu S Liu W Cai et al ldquoMultimodal neuroimaging featurelearning for multiclass diagnosis of Alzheimerrsquos diseaserdquo IEEETransactions on Biomedical Engineering vol 62 no 4pp 1132ndash1140 2015

[15] S H Nozadi and S Kadoury ldquoClassification of Alzheimerrsquos andMCI patients from semantically parcelled PET images a com-parison between AV45 and FDG-PETrdquo International Journal ofBiomedical Imaging vol 2018 Article ID 1247430 13 pages 2018

[16] Y Gupta R K Lama and G-R Kwon ldquoPrediction andclassification of Alzheimerrsquos disease based on combinedfeatures from apolipoprotein-E genotype cerebrospinal fluidMR and FDG-PET imaging biomarkersrdquo Frontiers inComputational Neuroscience vol 13 p 72 2019

[17] T H Gorji and K Naima ldquoA deep learning approach fordiagnosis of mild cognitive impairment based on MRI im-agesrdquo Brain Sciences vol 9 no 9 p 217 2019

[18] D Chyzhyk M Grantildea A Savio and J Maiora ldquoHybriddendritic computing with kernel-LICA applied to Alzheimerrsquosdisease detection in MRIrdquo Neurocomputing vol 75 no 1pp 72ndash77 2012

[19] T Zhang Z Zhao C Zhang J Zhang Z Jin and L LildquoClassification of early and late mild cognitive impairmentusing functional brain network of resting-state fMRIrdquoFrontiers in Psychiatry vol 10 p 572 2019

[20] R Cuingnet E Gerardin J Tessieras et al ldquoAutomaticclassification of patients with Alzheimerrsquos disease fromstructural MRI a comparison of ten methods using the ADNIdatabaserdquo NeuroImage vol 56 no 2 pp 766ndash781 2011

[21] S Farhan M A Fahiem and H Tauseef ldquoAn ensemble-of-classifiers based approach for early diagnosis of Alzheimerrsquosdisease classification using structural features of brain im-agesrdquoComputational andMathematical Methods inMedicinevol 2014 Article ID 862307 11 pages 2014

[22] Y Cho J-K Seong Y Jeong and S Y Shin ldquoIndividualsubject classification for Alzheimerrsquos disease based on in-cremental learning using a spatial frequency representation ofcortical thickness datardquo NeuroImage vol 59 no 3pp 2217ndash2230 2012

[23] R Wolz V Julkunen J Koikkalainen et al ldquoMulti-methodanalysis of MRI images in early diagnostics of Alzheimerrsquosdiseaserdquo PLoS One vol 6 no 10 Article ID e25446 2011

[24] C Ledig R A Heckemann A Hammers et al ldquoRobustwhole-brain segmentation application to traumatic braininjuryrdquoMedical Image Analysis vol 21 no 1 pp 40ndash58 2015

[25] K Van Leemput F Maes D Vandermeulen and P SuetensldquoAutomated model-based tissue classification of MR imagesof the brainrdquo IEEE Transactions on Medical Imaging vol 18no 10 pp 897ndash908 1999

[26] C Ledig ldquoSegmentation of MRI brain scans usingrdquo in Pro-ceedings of the MICCAI 2012 Grand Challenge and Workshopon Multi-Atlas Labeling Nice France September 2012

[27] A H Andersen D M Gash and M J Avison ldquoPrincipalcomponent analysis of the dynamic response measured byfMRI a generalized linear systems frameworkrdquo MagneticResonance Imaging vol 17 no 6 pp 795ndash815 1999

[28] J Cohen ldquoA coefficient of agreement for nominal scalesrdquoEducational and Psychological Measurement vol 20 no 1pp 37ndash46 1960

[29] M L McHugh ldquoInterrater reliability the kappa statisticrdquoBiochemia Medica vol 22 pp 276ndash282 2012

[30] F Deng S Siersdorfer and S Zerr ldquoEfficient jaccard-baseddiversity analysis of large document collectionsrdquo in Pro-ceedings of the 21st ACM International Conference on

Journal of Healthcare Engineering 13

Information And Knowledge Management-CIKM rsquo12 p 1402Maui HI USA 2012

[31] A Patil and N S Upadhyay ldquoRestaurantrsquos feedback analysissystem using sentimental analysis and data mining techniquesrdquoin Proceedings of the 2018 International Conference on CurrentTrends Towards Converging Technologies (ICCTCT) IEEECoimbatore Tamilnadu India pp 1ndash4 March 2018

14 Journal of Healthcare Engineering