Neurobiology of Disease 51 (2013) 82–92

Contents lists available at SciVerse ScienceDirect

Neurobiology of Disease

j ourna l homepage: www.e lsev ie r .com/ locate /ynbd i

Automated differentiation of pre-diagnosis Huntington's disease from healthy controlindividuals based on quadratic discriminant analysis of the basal ganglia:The IMAGE-HD study

N. Georgiou-Karistianis a,⁎, M.A. Gray a,b,h, i, J.F. Domínguez D a, A.R. Dymowski a, I. Bohanna c,L.A. Johnston d,e, A. Churchyard f, P. Chua a, J.C. Stout a, G.F. Egan a,b,d,g

a School of Psychology and Psychiatry, Faculty of Medicine Nursing & Health Sciences, Monash University, Clayton, VIC 3800, Australiab Monash Biomedical Imaging (MBI), Monash University, Melbourne, VIC 3800, Australiac School of Public Health, Tropical Medicine and Rehabilitation Sciences, James Cook University, Cairns, QLD 4870, Australiad Howard Florey Institute, Florey Neuroscience Institutes, Parkville, VIC 30105, Australiae Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, VIC 3052, Australiaf Department of Neurology, Monash Medical Centre, Clayton, VIC 3168, Australiag Centre for Neuroscience, University of Melbourne, Parkville, VIC 3010, Australiah Centre for Advanced Imaging, Gehrmann Laboratory, The University of Queensland, St Lucia, 4072 Australiai The University of Queensland Centre for Clinical Research, Royal Brisbane and Women's Hospital, Herston, 4029 Australia

⁎ Corresponding author at: Experimental NeuropsychPsychology and Psychiatry, Faculty of Medicine NursinUniversity, Clayton, VIC 3800, Australia. Fax: +61 3 990

E-mail address: nellie.georgiou-karistianis@monash.edAvailable online on ScienceDirect (www.sciencedir

0969-9961/$ – see front matter © 2012 Elsevier Inc. Allhttp://dx.doi.org/10.1016/j.nbd.2012.10.001

a b s t r a c t

a r t i c l e i n f o

Article history:Received 22 March 2012Revised 31 August 2012Accepted 3 October 2012Available online 13 October 2012

Keywords:Huntington's diseaseT1-weighted volumeMean diffusivityFractional anisotropyQuadratic discriminant function analysisMagnetic resonance imagingDiffusion weighted imagingMultivariate classificationBasal ganglia

We investigated two measures of neural integrity, T1-weighted volumetric measures and diffusion tensorimaging (DTI), and explored their combined potential to differentiate pre-diagnosis Huntington's disease(pre-HD) individuals from healthy controls. We applied quadratic discriminant analysis (QDA) to discriminatepre-HD individuals from controls andweutilised feature selection anddimension reduction to increase the robust-ness of the discrimination method. Thirty six symptomatic HD (symp-HD), 35 pre-HD, and 36 control individualsparticipated as part of the IMAGE-HD study and underwent T1-weighted MRI, and DTI using a Siemens 3 Teslascanner. Volume and DTI measures [mean diffusivity (MD) and fractional anisotropy (FA)] were calculated foreach group within five regions of interest (ROI; caudate, putamen, pallidum, accumbens and thalamus). QDAwas then performed in a stepwise manner to differentiate pre-HD individuals from controls, based initially onunimodal analysis of motor or neurocognitive measures, or on volume, MD or FA measures from within the cau-date, pallidum and putamen. We then tested for potential improvements to this model, by examiningmulti-modal MRI classifications (volume, FA and MD), and also included motor and neurocognitive measures,and additional brain regions (i.e., accumbens and thalamus). Volume, MD and FA differed across the three groups,with pre-HD characterised by significant volumetric reductions and increased FA within caudate, putamen andpallidum, relative to controls. The QDA results demonstrated that the differentiation of pre-HD from controlswas highly accurate when both volumetric and diffusion data sets from basal ganglia (BG) regions were used.The highest discriminative accuracy however was achieved in a multi-modality approach and when includingall availablemeasures: motor and neurocognitive scores andmulti-modal MRI measures from the BG, accumbensand thalamus. Our QDA findings provide evidence that combined multi-modal imaging measures can accuratelyclassify individuals up to 15 years prior to onset when therapeutic intervention is likely to have maximal effectsin slowing the trajectory of disease development.

© 2012 Elsevier Inc. All rights reserved.

Introduction

Modern neuroimaging techniques enable the examination of centralnervous system neuropathology in vivo. As analytical neuroimaging

ology Research Unit, School ofg & Health Sciences, Monash5 3948.u (N. Georgiou-Karistianis).ect.com).

rights reserved.

methods mature, the number of measures of neuronal integrity whichcan be accurately measured is increasing. With these advances thereare parallel efforts to establish increasingly sensitive and reliable imag-ing biomarkers of disease onset and progression inHuntington's disease(HD) (for reviews see Bohanna et al., 2008; Esmaeilzadeh et al., 2011;Georgiou-Karistianis, 2009; Paulsen, 2009). The period prior to HD clin-ical onset is commonly known as the preclinical phase (or pre-HD). Theability to estimate proximity to disease diagnosis (e.g., Langbehn et al.,2004) provides a significant opportunity to investigate neurobiologicalchanges during this period. There are a number of important issues in

83N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

relation to the development of reliable biomarkers. Different neuroim-aging measures may be sensitive to neurodegenerative changes at dif-ferent points in the disease trajectory. These changes may progress atdifferent rates, and may be visible within different neural structures atdifferent points in time.We aimed to establish a neuroimaging based bio-marker for HDwhich provides a reliablemetric for differentiating pre-HDfrom matched healthy controls. While our approach is cross-sectional, itprovides a methodwhich can be applied to future longitudinal investiga-tions of disease onset and progression, and which importantly may pro-vide a trajectory against which to assess pharmacological interventions.In this paper we used well-defined brain regions known to be selectivelyimpaired in HD and investigated twoMR imaging approaches tomeasur-ing neural integrity, T1-weighted volumetric measures and diffusiontensor imaging (DTI), and examined their combined potential to informa neuroimaging-based biomarker of HD. In particular, we investigatedthe ability of these measures to differentiate pre-HD individuals frommatched healthy controls.

The observation of neurodegeneration within the caudate nucleusis well supported by a large number of magnetic resonance imaging(MRI) studies reporting reduced volume in symptomaticHD (symp-HD)individuals (Brandt et al., 1995; Douaud et al., 2006; Fennema-Notestineet al., 2004; Harris et al., 1992; Henley et al., 2009; Jernigan et al., 1991;Kassubek et al., 2004; Mascalchi et al., 2004; Peinemann et al., 2005;Rosas et al., 2001; Starkstein et al., 1992). Importantly, a considerablenumber of studies also report volumetric reductions in caudate duringpre-HD, and in many cases more than a decade prior to predicted ageat onset (Aylward et al., 1994, 1996, 2000, 2004; Campodonico et al.,1998; Harris et al., 1999; Hobbs et al., 2010; Jurgens et al., 2008; Kippset al., 2005; Paulsen et al., 2006; Reading et al., 2005; Soneson et al.,2010; Stoffers et al., 2010; Thieben et al., 2002; van den Bogaard et al.,2010). Furthermore, reductions in caudate volume have been relatedto severity of cognitive decline (Bäckman et al., 1997), correlated toCAG repeat length (Ruocco et al., 2006) and years to onset (Paulsen etal., 2008), and have been found in at least one study to proceed at afaster rate within the left hemisphere (Mühlau et al., 2007). Neurode-generative changes are also frequently observed within the putamen,and like the caudate, reduced volumes have been observed in bothsymp-HD (Bäckman et al., 1997; Henley et al., 2009; Kassubek et al.,2004; Mühlau et al., 2007; Peinemann et al., 2005; Ruocco et al., 2006)and pre-HD (Campodonico et al., 1998; Harris et al., 1999; Hobbs et al.,2010; Jurgens et al., 2008; Kipps et al., 2005; Paulsen et al., 2008;Stoffers et al., 2010; Thieben et al., 2002; van den Bogaard et al., 2010).

Diffusion based neuroimaging measures have become widelyestablished over the last few years and now represent the largest bodyof neuroimaging findings, next to volumetric reports, for identifying pa-thology in caudate (Klöppel et al., 2008; Mascalchi et al., 2004; Readinget al., 2005; Rosas et al., 2006; Seppi et al., 2006; Sritharan et al., 2010)and putamen (Klöppel et al., 2008; Rosas et al., 2006; Sritharan et al.,2010). After parcellating striatal regions, based on connectivity to differ-ing cortical projection sites, Bohanna et al. (2011) demonstrated thegreatest extent of pathological volumetric and diffusion change withinstriatal motor regions that correlated with disease severity in symp-HD.Extra-striatal regions, such as the pallidum, also reveal volumetric reduc-tions in symp-HD (Douaud et al., 2006) and pre-HD (Kipps et al., 2005;Thieben et al., 2002; vandenBogaard et al., 2010), aswell as degenerativechanges via DTI in pre-HD (Klöppel, et al., 2008; Rosas, et al., 2006). Ad-ditionally, reductions in pallidum volume have been shown to correlatewith increased oculomotor deficits (Hicks et al., 2010) and increased clin-ical severity (Jurgens et al., 2008).

Further examination of HDpathology reveals that the rate of degener-ationmay not be evenly distributed throughout the basal ganglia (BG). Inearlier accounts, Vonsattel et al. (1985) observed differing degrees ofdegeneration within the BG at post-mortem including a dorso-ventralgradient of neuropathological changes, with the dorsal regions (caudateand dorsal putamen) affected before the ventral regions. More recently,Douaud et al. (2006) reported a similar pattern of neuroimaging

volumetric measures within early diagnosed individuals. The ventral re-gions of the striatum including the nucleus accumbens, have beenreported to undergopathological change fairly late in thedisease trajecto-ry. While many neuroimaging studies have not differentiated and inde-pendently examined the accumbens from the putamen, which isdorsally continuous with it, a small number have reported specific vol-ume reductions in the accumbens (Fennema-Notestine et al., 2004;Hobbs et al., 2010; van den Bogaard et al., 2010), supporting earlierpost-mortem findings (Bots and Bruyn, 1981; De la Monte et al., 1988).

Widespread structural changes have been reported in regions out-side the BG with, in particular, volume reductions of the thalamus insymp-HD (Bäckman et al., 1997; Douaud et al., 2006, 2009; Henley etal., 2009; Jernigan et al., 1991; Kassubek et al., 2005; Starkstein et al.,1992) and pre-HD (Paulsen et al., 2006; Thieben et al., 2002). DTI abnor-malities have been reported within the thalamus (Rosas et al., 2006;Seppi et al., 2006) and thalamic radiations (Stoffers et al., 2010). Thesefindings, togetherwith previous accounts of dopaminergic receptor dis-turbances in striatal projection neurons (Albin et al., 1992; Antonini etal., 1996), suggest that striatal-thalamic regionsmay usefully contributeto an imaging based biomarker of neurodegenerative change in HD.Although a number of neuroimaging studies have reported cortical neu-rodegenerative changes in both symp-HD and pre-HD (Hobbs et al.,2010; Jernigan et al., 1991; Kassubek et al., 2004; Nopoulos et al.,2007; Peinemann et al., 2005; Reading et al., 2005; Rosas et al., 2002,2003, 2008; Ruocco et al., 2006; Tabrizi et al., 2009; van den Bogaardet al., 2010), these studies have reported a high degree of variability inthe location of observed differences and in the cortical regions selectedfor study (see Henley et al., 2009; Hobbs et al., 2010).

Recently, anatomical classification analyses have been applied to dif-ferentiate pre-HD from healthy control participants, where neurode-generation maybe more subtle (Klöppel et al., 2008, 2009; Rizk-Jacksonet al., 2011). Thesemachine learning approaches differ frommass univar-iate approaches as they are inherently predictive (Pereira et al., 2009),with an analysisfirst conducted on a training-set of data tomaximally dif-ferentiate the classes, followed by prediction of groupmembership for anindependent data set. Klöppel et al. (2008) appliedmachine learning clas-sification to DTImeasures within whitematter (WM) tracts in 25 pre-HDindividuals and 20 matched healthy controls. They further extended thismethod to predictively classify 96 pre-HD from 95matched healthy con-trol participants using segmentedMRImeasures ofwhole brain greymat-ter (Klöppel et al., 2009). Multivariate classification techniques may alsobe useful in application to HD, potentially identifying MRI based bio-markers that are sensitive to disease progression.Multivariate approachesbased on discriminative models that incorporate a number of indepen-dent measures of disease progression, including volumetric andDTI-MRI derived measures have recently been reported. Rizk-Jacksonet al. (2011) applied amultivariate approach by adopting two classifica-tion methods (linear discriminant function and support vector machineanalysis) with volumetric and DTI measures and differentiated pre-HDfrom healthy controls.

Multivariate analytical (MVA) techniques (as opposed to mass-univariate approaches) are novel in HD, and in neuroimaging researchgenerally, and as outlined in a recent review of multivariate anatomicalclassifier techniques (see Ashburner and Klöppel, 2011) often sufferfrom the “curse of dimensionality”. That is, MVA techniques may beunder-determined and contain many more observations than subjects.This difficulty canbe ameliorated bydimensionality reduction via a statis-tical summary such as singular-value decomposition or independentcomponent analysis, or by feature selection such as restriction of thedataset to relevant anatomical ROI. The well-established anatomicalchanges observed in HD therefore provide a significant advantage overand above other neurodegenerative disorders since selected nuclei with-in the BG undergo neuronal degeneration that can be identified duringboth pre-HD and symp-HD stages.

In the current study we adopted a quadratic discriminant analysis(QDA) as a multivariate classifier, applied to three measures derived

Table 1Demographic, clinical, motor, neurocognitive and neuropsychiatric data betweengroups.

Mean±SD

Controls (n=36)�x±σ

Pre-HD (n=35)�x±σ

Symp-HD (n=33)�x±σ

Gender (M:F) 14:22 14:21 17:16Age (years) 42.41±13.44 42.22±9.87a 51.35±8.98b

NART (IQ) 117.99±9.71 116.90±11.40 112.96±11.26UHDRS 0.94±1.24a 18.85±12.83CAG repeats 42.20±1.92 43.27±2.50Disease burdenscore (DBS)

268.61±53.23a 381.72±72.12

Estimated YTO 15.58±7.12Duration of illness(years)

1.99±1.67

Speeded tapping (ms) 219.85±38.09 243.59±45.04a 372.32±167.10c

Self-paced tapping —

(1/SD ITI) 333 ms23.81±7.74d 19.58±7.19a 10.35±4.30c

Self-paced tapping —

(1/SD ITI) 550 ms29.13±8.25d 23.72±8.47a 11.43±5.98c

UPSIT 33.97±3.08 32.57±4.97a 26.21±6.99c

SDMT 56.28±10.08e 51.57±8.76a 36.18±11.98c

STROOP 109.75±16.57 104.46±17.80a 83.18±22.14c

SCOPI — total OCD 80.94±22.32 83.09±24.56 90.30±26.89FrSBe — total score 87.06±25.38 91.80±23.20 95.97±23.97HADS: A 4.97±2.77 6.43±3.33 5.79±3.38HADS: D 2.61±3.09 2.71±3.00 2.82±2.31BDI II 4.03±4.12d 8.71±9.95 8.70±7.05b

NART: National Adult Reading Test; UHDRS: motor subscale score Unified Huntington's Dis-ease Rating Scale (pre-HD, UHDRSb5; symp-HD, UHDRS≥5); CAG: cytosine-adenine-guanine (number of repeats >40 is full penetrance); Disease Burden Score: (CAG−35.5)∗age; YTO: years to onset; DBS: Disease Burden Score; SDMT: Symbol Digit Mo-dalities Test; STROOP: STROOP speeded word reading task (number of correctwords); FrSBe: Frontal Systems Behaviour Scale; SCOPI: Schedule of Compulsions Ob-sessions and Pathological Impulses; HADS A: Hospital Anxiety and Depression scale— anxiety sub score; HADS D: Hospital Anxiety and Depression scale — depressionsub score; BDI II: Beck Depression Inventory score Version II; UPSIT: University ofPennsylvania Smell Identification Test (score out of 40); ITI: inter-tap interval. Agewas used as a covariate in all group comparisons except for age, UHDRS, CAG and DBS.

a Pre-HD v symp-HD, pb0.001.b Control v symp-HD, pb0.05.c Control v symp-HD, pb0.001.d Control v pre-HD, pb0.01.e Control v pre-HD, pb0.05.

84 N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

from T1-weighted and DTI images. QDA is a generalisation of a lineardiscriminant function analysis thatfitsmultivariate normal distributionswith covariance estimates for each group. We utilised both dimension-ality reduction and feature selection by considering the mean values ofvolume, mean diffusivity (MD) and fractional anisotropy (FA) measuresfrom BG regions, namely, the caudate, putamen, pallidum, accumbens,and from the thalamus. We hypothesised that: 1) volumetric and DTImeasures (i.e., MD and FA) in the BG (caudate, putamen and pallidum)would differentiate healthy controls frompre-HDand symp-HD individ-uals, and 2) thatmulti-modal QDAwould best discriminate pre-HD indi-viduals frommatched healthy controls. Finally, we investigatedwhetherthe inclusion of motor and/or neurocognitive measures within the QDAwould improve the discrimination accuracy of the model.

Methods

Participants

IMAGE-HD is an ongoing longitudinal multi-modal neuroimagingstudy. This paper reports on cross-sectional data acquired at study base-line (2008–2009). A total of 108 participants were originally recruited,consisting of 36 healthy control, 36 pre-HD and 36 early diagnosed HD(symp-HD) participants. One pre-HD and 3 symp-HD participants wereexcluded due to missing data or segmentation failure, leaving a total of35 pre-HD and 33 symp-HD participants. Healthy controls werematchedfor age, gender and IQ (National Adult Reading Test 2nd edition,NART-2; Nelson and Willison, 1991) to the pre-HD individuals. Pre-HDand symp-HD participants were clinically assessed (by A.C or P.C.)with the Unified Huntington's Disease Rating Scale (UHDRS) motor as-sessment (Huntington Study Group, 1996). Consistent with the criteriaemployed by Tabrizi et al. (2009), individuals with a UHDRS motorscore ≤5 were included in the pre-HD group and those with UHDRSmotor score >5 were included in the symp-HD group. Moreover, allsymp-HD participants had a clinical diagnosis of HD. CAG repeat lengthranged from 39 to 50 (42.2±1.9 for pre-HD; 43.3±2.5 for symp-HD).The pre-HD group had an average estimated years to clinical onset(Langbehn et al., 2004) of 15.6±7.1 years while the symp-HD grouphad an estimated average years since diagnosis of 2.0±1.7 years. De-mographic, clinical, motor, neurocognitive and neuropsychiatric dataare presented in Table 1.

A battery of neurocognitive tests were administered on the day ofscanning and were selected as part of the IMAGE-HD protocol basedon their sensitivity in previously detecting differences between groups(Stout et al., 2011; Tabrizi et al., 2009). The tests assessed visuomotorspeed and attention (Symbol Digit Modalities Test, SDMT; Smith,1982), speeded reading (StroopWord Test; Stroop, 1935), odour recog-nition (University of Pennsylvania Smell Identification Test, UPSIT; Dotyet al., 1984) and motor performance (speeded tapping and self-pacedtapping tasks; Stout et al., 2011). Participants completed behaviouralquestionnaires which included assessments of behaviours associatedwith frontal–striatal brain dysfunction, including executive function(Frontal Systems Behaviour Scale, FrSBe; Grace and Mallory, 2001)and psychiatric disturbances (Schedule of Obsessions, Compulsionsand Psychological Impulses, SCOPI; Watson and Wu, 2005; HospitalAnxiety and Depression Scale, HADS; Zigmond and Snaith, 1983; BeckDepression Inventory Version II, BDI-II; Beck et al., 1996). Scores onneurocognitive tests are shown in Table 1. T-tests revealed thatsymp-HD performed worse than both pre-HD (pb0.001) and control(pb0.001) participants in motor function (speeded, self-paced fastand self-paced slow tapping), UPSIT, SDMT and STROOP reading scores.Pre-HD also performed worse during self paced tapping (slow and fast)and SDMT, relative to controls. BDI II scores were higher in pre-HD(pb0.01) and symp-HD (pb0.05) relative to control participants.There were no significant between group differences in SCOPI (totalOCD), FrSBe (total score) or HADS anxiety or depression scores. Allgroup comparisons of motor, neurocognitive and neuropsychiatric

measures were corrected for differences in age. Group differencesin motor, neurocognitive and neuropsychiatric measures are alsopresented as percent change scores, relative to control participants,in Supplementary Table 1. The study was approved by the MonashUniversity and Melbourne Health Human Research Ethics Committeesand informedwritten consentwas obtained from each participant priorto testing in accord with the Helsinki Declaration.

Procedure

Acquisition of MR imagesAll participants underwent T1-weighted and diffusion tensor imaging

(DTI) on a Siemens Magnetom Trio Tim System 3 Tesla scanner with a32-channel head coil (Siemens AG, Erlangen, Germany) at the MurdochChildren's Research Institute (Royal Children's Hospital, Victoria,Australia). High-resolution T1-weighted images were acquired (192slices, slice thickness of 0.9 mm, 0.8 mm 0.8 mm in-plane resolution320 320 field of view, TI=900 ms ,TE=2.59 ms, TR=1900 ms, flipangle=9°). Diffusion weighted data were also acquired using a doublespin echo diffusion weighted EPI sequence (TR=8200, TE=89 ms,Flip=90°, 64 contiguous slices with 2 mm isotropic voxels, acquisitionmatrix 128 128). Diffusion-sensitising encoding gradients were appliedin 60 directions using a b value of 1200 s/mm2, and 10 images withoutdiffusion weighting (b=0 s/mm2).

85N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

Whole-brain segmentationFMRIB's Software Library (FSL 4.1.6 l, Smith et al., 2004)was used for

whole brain segmentation. T1 images were first skull stripped using theBrain Extraction Tool (BET) and were all visually checked to ensure ac-curacy. Next, FMRIB's Automated Segmentation Tool (FAST) generatedglobal tissue classifications of grey matter (GM), white matter (WM)and cerebrospinal fluid (CSF). Total Intracranial Volume (TIV=GM+WM+CSF) and Total Tissue Volume (TTV) estimates were calculatedfrom these outputs.

ROI volumesThe accuracy of fully automated identification of ROIs (i.e., caudate

nucleus, putamen, globus pallidus, nucleus accumbens, thalamus) canbe challenging in neurodegenerative populations. We adopted a semi-customised procedure, based on Bayesian tissue segmentation, and im-plemented in SPM8's “Segment” routine. This approach applies estimat-ed tissue probability maps within a unified model which additionallyaccounts for individual variability in neural structure and B0 field inho-mogeneities (Ashburner & Friston, 2005). Individual structural scanswere co-registered toMNI space, and normalisation parameters derivedusing SPM8 segment routine, with permissive warping regularisation(regularisation factor 0.01, discrete cosine transform warp frequencycutoff of 22). After visually checking for accurate T1 normalisations,we applied standardised subcortical masks (see Harvard Oxford Atlas)to identify ROIs. Particular emphasis was placed on accurate identifica-tion of peri-ventricular CSF boundaries, since this boundary is highlysensitive to HD pathology. Inaccuracies in identifying these boundariescan occur with fully automated segmentation methods typically usedfor identifying subcortical structures, especially in symp-HD. Standardspace ROI masks, altered to include CSF voxels at the caudate boundaryand to avoid additional regions such as the fornix, were then applied toeach participant's normalised structural scan. Estimated ROIs for cau-date, putamen, accumbens and thalamus were then eroded by remov-ing voxels identified as CSF in segmented probabilistic classifications.Caudate ROIs were additionally corrected by removing voxels probabi-listically identified as WM. Corrected ROIs were then transformed into

Fig. 1. Automated identification of subcortical regions of interest across participant groupssymp-HD participant for illustrative purposes. A) Structural T1 weighted images wereperiventricular boundaries were applied. C) Peri-ventricular boundaries were then correct(red), putamen (green), accumbens (blue) and thalamus (purple). The illustrative image intion, Culture, Sports, and Technology(MEXT) Integrated Database Project, and is licensed b

native space via the inverse normalisation transformation, and theresulting structures visually inspected for accuracy before volume esti-mates were calculated. See Fig. 1.

DTI measuresIn order to avoid partial volume effects, that may result from larger

voxel sizes in diffusion weighted scans, and misalignment betweenstructural and diffusion images, we utilised independent ROI identifica-tion in diffusion images. Diffusion weighted images (DWI) were brainextracted (BET) and corrected for movement and eddy current induceddistortions via affine registration to a reference B=0 volume (FMRIB'sDiffusion Toolbox, FDT). Calculation of diffusion tensors and principleeigenvectors allowed derivation of voxel wise MD and FA maps. MDand FA measures were then computed for each subcortical ROI usingstructure specific masks. First, each participant's brain-extracted T1imagewas non-linearly registered to its corresponding FAmap (FMRIB'sNonlinear Image Registration Tool, FNIRT). Segmentation of subcorticalstructures was performed on the non-linearly registered T1 image withFIRST (FSL's automated segmentation and registration tool that usesBayesian shape and appearance models). We applied the default bound-ary correction methods included in FIRST segmentations (Patenaude etal., 2011); caudate boundaries were eroded using intensity basedGaussian mixture-models identifying CSF, GM and WM and appliedvia Markov Random Field. Putamen, pallidum and thalamus erosionswere achieved by fitting surfaces to the manual models after erosionby 0.5 mm from the outer boundary. We did not boundary correctaccumbens' volumes due to the small initial diameter of this structure.The masks generated were then binarised and used to obtain FA andMD measures from the respective DTI maps. Segmentations and regis-trations were visually inspected to ensure accuracy. We were satisfiedthat this method yielded adequate masks enabling structure-widevoxel samples.

Statistical analyses: whole brain and ROI measuresWe initially characterised our sample by testing for univariate

group differences in whole brain volumes, and differences in volume,

. All participants were processed. We provide a representation of one control and onenon-linearly normalised to MNI space where B) hand traced masks enlarged ated before D) returning to subject space; coronal and axial sections illustrate caudatepanel B was generated from BodyParts3D/Anatomography website, Ministry of Educa-y them for free use under the Creative Commons Attribution 2.1 Japan license.

86 N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

FA and MD within ROIs. One-way ANCOVAs assessing the effect ofgroup on TTV, GM, WM and CSF included TIV and age as covariates.Likewise, one-wayANCOVA's assessing groupdifferences in ROI volumealso included TIV and age as covariates, whereas one-way ANOVAs forgroup differences in ROI FA and MD included only age as a covariateof no interest. All ANCOVAs were performed in SPSS 16.0.

Quadratic discriminant analysis (QDA)We next examined multivariate discrimination. As with previous HD

classification papers our primary focus was to discriminate between pre-HDparticipants andmatched healthy controls, where neurological differ-ences are likely to be more subtle. For completeness we also present theresults of themultivariate discrimination analysis between symp-HD andcontrols (see Supplementary Table 2). Bartlett's test revealed homosce-dasticity inmultivariate covariancematrices, therefore quadratic discrim-inant analysis was performed rather than the simpler linear discriminantanalysis. Additionally, we applied diagonalised quadratic covariancematrices to avoid violations by singular covariance matrices, and to en-sure robustmultivariate discrimination. To demonstrate which features(variables) provided the best group discriminations, we adopted a step-wise approach. First, we considered eachmodality separatelywithin ourmodel (i.e., motor, neurocognitive, and the following MRI measures —volume, MD and FAmeasures fromwithin the BG). Next, we tested dis-criminative accuracy for combinedmodalities; that is, motor and neuro-cognitive measures combined with MRI measures from within the BG.Finally, we investigated the potential for additional improvements in dis-criminative accuracy by combining motor and neurocognitive and MRImeasures within the BG together with additional regions (i.e., nucleusaccumbens and thalamus). We reported discrimination accuracy basedon an iterative “leave one out” cross validation, where discriminationsare trained on N-1 participants, and tested with the remaining partici-pant. This is repeated for all possible combinations of training and testingdatasets, with the reported accuracy reflecting the total number of mis-classifications across all permutations. All multivariate normality and dis-crimination analyses was carried out using Matlab (Mathworks, version7.14.0.739 R2012a with the statistics toolbox).

Results

Whole brain volume analyses (i.e., GM,WM, CSF, TTV) and ROI anal-yses for individual measures of volume, MD and FA (i.e., left and rightaccumbens, caudate, pallidum, putamen, and thalamus) are presentedbelow, followed by the QDA and QDA cross validation in pre-HD andmatched control individuals using the volumetric, MD and FAmeasures.

Whole brain volume

One-way ANCOVAs revealed that symp-HDparticipants differed fromboth control and pre-HD participants, with reduced volume in measuresof TTV, GM and CSF. WM volume was also reduced in symp-HD, relativeto controls. There were no significant differences between pre-HD and

Table 2Whole brain volumes and statistical differences between groups.

Group Group effect C

TTV F=25.97 pb0.001 1GM volume F=8.30 pb0.001WM volume F=3.64 p=0.030CSF volume F=24.31 pb0.001

TTV: total tissue volume (GM+WM); GM: grey matter; WM: white mater; CSF: cerebrospAll measurements reported in cm3.Group effects: One-way ANCOVA for the effect of group, F(2, 99), covaried for age and TIV.

a Pre-HD vs symp-HD pb0.001.b Control vs symp-HD pb0.001.c Control vs symp-HD pb0.01.

controls. Whole brain GM, WM and CSF volume data are presented inTable 2.

ROI analyses

ROI volume by groupOne-way ANCOVAs of regional volume revealed significant group

effects bilaterally within the caudate, putamen and pallidum. Withineach of these structures post hoc comparisons revealed significantlyreduced volume in symp-HD relative to both controls and pre-HD partic-ipants. Likewise, volume was significantly reduced within caudate, puta-men and pallidum in pre-HD relative to control participants. Volumemeasures of accumbens and thalamus, however, did not discriminatebetween groups. Mean group volumes and statistical differences arepresented in Table 3, and illustrated in Fig. 2.

ROI FA by groupOne-way ANCOVAs of FA revealed significant group effects for cau-

date, putamen and pallidum bilaterally, and for the right accumbens.Post hoc comparisons revealed bilateral group differences within cau-date and putamen, with FA increasing from control to pre-HD andfrompre-HD to symp-HD. FAwas also significantly increased in bilateralpallidum in pre-HDand symp-HD relative to controls. Additionally, rightaccumbens FAwas reduced in symp-HD relative to pre-HD and controls.Thalamic FA did not differ between groups (see Table 3).

ROI MD by groupOne-way ANCOVAs revealed significant differences in MD across

groups within bilateral caudate, putamen, pallidum and accumbens,but not thalamus. Post hoc comparisons revealed increased symp-HDMD relative to controls, in bilateral caudate, putamen and accumbens,and in symp-HD relative to pre-HD in bilateral putamen and right cau-date. Pre-HD MD was also increased relative to controls in bilateral pu-tamen accumbens, and left caudate. Conversely, symp-HD MD wasreduced bilaterally in the pallidum relative to pre-HD and controls,and also in the right hemisphere in pre-HD relative to controls. ThalamicMD did not differ between groups (see Table 3).

Diagonalised quadratic discriminant analysis (QDA)Considering each modality individually, classification accuracy was

relatively poor, with onlymotor and BGMDmeasures performing signif-icantly above chance, although not surviving Bonferroni correction formultiple QDA comparisons (66.2% accuracy in each case, see Table 4).Discrimination accuracy with either neurocognitive, volume or FA mea-sures was non-significant. Classification accuracy considerably improvedwhen considering multiple modalities, and all models discriminated sig-nificantly above chance following correction for multiple comparisons.The inclusion ofmotor and neurocognitivemeasures discriminated accu-racy by 69.0%. Discrimination accuracy based on volume, FA and MDwithin the BG reached 70.4%. Adding neurocognitive measures to thismodel did not improve accuracy (70.4%), whereas adding motor im-proved accuracy to 74.6%. Adding multi-modal MRI measures from the

ontrols Pre-HD Symp-HD

179.0±124.6 1131.0±141.0a 1054.0±133.03b

629.8±64.6 609.6±80.3a 553.4±84.7b

549.2±65.2 521.4±69.9 500.7±65.0c

277.5±37.9 281.7±35.1a 326.2±40.0b

inal fluid.

Significance of individual differences (covaried for age and TIV) are:

Table 3Statistical group differences within regions of interest (volume, FA and MD).

ROI Side Volume ω FA×10−2 ψ MD×10−5 ψ

Group Control Pre-HD Symp-HD Group Control Pre-HD Symp-HD Group Control Pre-HD Symp-HD

F p �X SEM �X SEM �X SEM F p �X SEM �X SEM �X SEM F p �X SEM �X SEM �X SEM

Caudate L 33.4 0.001 3.22a±0.07 2.88b±0.10 2.21c±0.08 20.8 0.001 14.08d±0.26 15.82b±0.50 19.29c±0.71 4.0 0.020 85.32d±0.92 90.31±2.20 93.46e±2.08R 39.1 0.001 3.22a±0.07 2.84b±0.09 2.16c±0.08 15.3 0.001 13.86d±0.32 14.99b±0.48 18.43c±0.77 7.2 0.001 93.50±1.80 100.50f±3.34 112.07c±3.51

Putamen L 31.1 0.001 4.39g±0.09 3.84b±0.10 3.39c±0.08 29.7 0.001 14.92±0.35 16.22b±0.44 20.14c±0.48 34.3 0.001 86.78g±0.52 92.15b±1.14 100.62c±1.48R 28.4 0.001 4.08g±0.09 3.55h±0.09 3.15c±0.07 24.0 0.001 15.30d±0.37 16.52b±0.48 20.78c±0.62 41.5 0.001 85.53g±0.59 90.17b±0.95 98.35c±1.25

Pallidum L 18.9 0.001 0.81g±0.02 0.72h±0.02 0.64c±0.02 14.0 0.001 40.05±1.43 43.07b±1.79 50.73c±1.31 5.6 0.005 81.96±1.28 82.76h±1.81 77.35e±1.39R 16.7 0.001 0.77a±0.02 0.68h±0.02 0.61c±0.02 27.8 0.001 43.15±1.45 47.46b±1.63 55.98c±1.53 10.1 0.001 81.67d±1.19 78.11f±1.19 74.70c±1.32

Accumbens L 1.5 0.220 0.25±0.01 0.23±0.01 0.24±0.01 0.1 0.863 21.15±0.40 20.99±0.50 21.47±0.46 8.6 0.001 93.19g±0.58 100.55±2.04 99.94c±1.23R 0.7 0.512 0.21±0.01 0.20±0.01 0.20±0.00 3.1 0.050 21.03±0.53 20.36f±0.35 19.86i±0.40 3.6 0.032 97.29d±1.51 102.46±2.01 102.64i±1.68

Thalamus L 1.0 0.358 6.87±0.14 6.72±0.11 6.76±0.12 0.1 0.937 29.15±0.32 29.07±0.30 29.89±0.39 0.6 0.545 90.00±0.60 90.93±0.56 91.10±0.77R 1.0 0.390 6.64±0.13 6.46±0.11 6.47±0.11 0.2 0.815 29.83±0.35 29.79±0.34 30.67±0.39 1.2 0.294 88.86±0.61 89.47±0.65 90.31±0.73

FA: fractional anisotropy; MD: mean diffusivity.ω One-way ANCOVA for effect of group, F(2, 99), covaried for age and TIV.ψ One-way ANCOVA for effect of group, F(2, 100), covaried for age.Significance of individual group differences are:

a Con v pre-HD pb0.01.b Pre-HD v symp-HD pb0.001.c Con v symp-HD pb0.001.d Con v pre-HD pb0.05.e Con v symp-HD pb0.01.f Pre-HD v symp-HD pb0.05.g Con v pre-HD pb0.001.h Pre-HD v symp-HD pb0.01.i Con v symp-HD pb0.05.

87N.G

eorgiou-Karistianis

etal./

Neurobiology

ofDisease

51(2013)

82–92

Fig. 2. Boxplots illustrating volume, FA and MD in control, pre-HD and symp-HD groups in each ROI. Significant group differences are illustrated; FA group differences here are il-lustrated separately for hemisphere. Figures illustrate raw values for volume, FA and MD, whereas statistical comparisons include age (and total intracranial volume for volume ROImeasures) as a covariate. Pre-HD: pre-symptomatic Huntington's disease, symp-HD: symptomatic Huntington's disease.

88 N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

Table 4Diagonialised quadratic discriminant analyses of volume, FA and MD within the regions of interest in pre-HD and controls.

Measurement Neural regions of interest Discrimination accuracy leave one out cross validation

% correct 95% CI Z score p Value

Single modalitiesNeurocog only – 59.2% 46.8–70.7% 1.42 >0.05Motor only – 66.2% 54.0–77.0% 2.61 0.01Volume Caudate, putamen, pallidum 62.0% 50.0–73.2% 1.90 >0.05FA Caudate, putamen, pallidum 57.7% 45.4–69.4% 1.19 >0.05MD Caudate, putamen, pallidum 66.2% 54.0–77.0% 2.61 0.01

Multiple-modalitiesMotor and neurocog – 69.0% 56.9–79.5% 3.09 0.002a

vol FA MD Caudate, putamen, pallidum 70.4% 58.4–80.7% 3.32 0.0009a

Vol FA MD Caudate, putamen, pallidum, accumbens, thalamus 71.8% 59.9–81.9% 3.35 0.0004a

Vol FA MD+neurocog Caudate, putamen, pallidum 70.4% 58.4–80.7% 3.32 0.0009a

Vol FA MD+motor Caudate, putamen, pallidum 74.6% 62.9–84.2% 4.04 0.0001a

Vol FA MD+motor and neurocog Caudate, putamen, pallidum 74.6% 62.9–84.2% 4.04 0.0001a

Vol FA MD+motor and neurocog Caudate, putamen, pallidum, accumbens, thalamus 77.5% 66.0–86.5% 4.51 0.00001a

Controls n=36, pre-HD n=35. Confidence intervals (CI), binomial z-ratios and p scores (two tailed) calculated from exact binomial probabilities. Vol: volume. FA: fractional an-isotropy; MD: mean diffusivity. Motor scores: speeded tapping and fast self-paced tapping. Neurocognitive scores: Stroop reading score, Symbol Digit Modalities Test (DSMT) andUniversity of Pennsylvania Smell Identification Test (UPSIT).

a Significant after Bonferroni correction for multiple QDA comparisons (α=0.004).

89N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

accumbens and thalamus improved accuracy to 71.8%. Discrimination ac-curacy based on multi-modal MRI measures within the BG was furtherimproved by the consideration of motor or motor and neurocognitivescores (74.6% in each case). The highest discriminative accuracy howeverwas achieved when including all available measures: motor and neuro-cognitive scores andmulti-modalMRImeasures from the BG, accumbensand thalamus (77.5%, see Table 4). QDA discrimination was also per-formed between symp-HD and control participants; discrimination washighly accurate in all models, ranging from 79.7% to 97.1% (see Supple-mentary Table 2).

Discussion

Our aim was to investigate two measures of neural integrity, T1-weighted volumetric and diffusion based measures, and to examine,using QDA, their combined potential as a neuroimaging based biomarkerof HD. In the first instance we reported on volumetric and DTI measuresacross groups. Our findings demonstrate that whole brain GM, WM andTTVwere significantly reduced in symp-HD relative to controls, with GMand TTV also differing from pre-HD; none of the whole brain measuresdifferentiated pre-HD from controls. In support of our first hypothesis,BG volumetric, MD and FA measures differed across all groups. Volumeand FA measures differentiated groups most clearly, with stepwise re-ductions in volume and increases in FA with all regions of the BG fromcontrol to pre-HD and from pre-HD to symp-HD. Similarly, caudateand putamen MD was consistently increased in symp-HD relative tocontrols. Diffusion also revealed reductions relative to controls; pallidalMD and right accumbens FA were reduced in symp-HD relative topre-HD and controls. Thalamic volume and diffusion measures did notdiffer between groups. In accordance with our second hypothesis, QDAbased on multi-modality measures consistently differentiated pre-HDfrom matched control participants. Classifications based on MRI mea-sures outperformed those based on only motor and neurocognitivemeasures, and the addition of motor scores improved accuracy basedon multi-modal BG MRI measures alone. The highest discriminative ac-curacy was obtained by furnishing the QDA model with all availablemeasures motor, neurocognitive and multi-modal MRI measures fromthe BG, accumbens and thalamus.

Neuroimaging studies over the last decade have played a significantrole in characterising structural, microstructural and functional alter-ations during various stages of the HD continuum. Large-scale multi-modal imaging studies (TRACK-HD, PREDICT-HD, IMAGE-HD) are

currently searching for sensitive and reliable biomarkers of diseaseonset and progression. Although many imaging measures individuallyhave detected a range of abnormalities during early stage pre-HD, asidentified in this study and previously reviewed (Bohanna et al., 2008;Esmaeilzadeh et al., 2011; Georgiou-Karistianis, 2009; Paulsen, 2009;Rizk-Jackson et al., 2011) to date no single imaging modality has beenvalidated as an optimal marker. Our whole brain, and in particular BG,measures provide further evidence of the utility of MRI in identifyingcross-sectional neurodegenerative change in individuals relatively farfrom onset (Langbehn et al., 2004). Novel alternatives to these univari-ate approaches, such as multivariate or combined multi-modal imagingapproaches, may offer the potential to provide a more sensitive neuro-imaging based biomarker of HD. Once established in cross-sectionalstudies, these approaches may be applied in longitudinal designs, withthe aim of identifying within-subject neurodegenerative change overshort time-spans (e.g., 12 months).

Themultivariate model presented in this paper is biologically repre-sentative and sensitive to both cell loss (i.e., volume reductions) anddegradation of neuronal membrane integrity (i.e., FA and MD) withinbrain regions that have previously shown reliable univariate evidenceof brain pathology in HD. Specifically, within individual modalities ourfindings are supported by previous reports of decreased volume withincaudate, putamen and pallidum (Aylward et al., 2004; Paulsen et al.,2008; Soneson et al., 2010; Stoffers et al., 2010; Tabrizi et al., 2009;van den Bogaard et al., 2010) and increased FA and MD within caudateand putamen (Bohanna et al., 2011; Douaud et al., 2009; Mascalchi etal., 2004; Sánchez-Castañeda et al., in press; Seppi et al., 2006). Interest-ingly, we also observed decreased pallidumMD and right accumbens FAin the pre-HD sample. Our pallidal ROIs encompassed internal and ex-ternal segments, and themedialmedullary laminaWMbandwhich sep-arates them. Our observation of reduced pallidal MD may conceivablyreflect WM (i.e., medial medullary lamina) neurodegeneration in HD,typically associated with reduced microstructural diffusion (Bohannaet al., 2011; Della Nave et al., 2010; Rosas et al., 2006). The relationshipbetween diffusion based MRI and specific microstructural alteration isindirect however, and further research is needed to confirm thesefindings.

Importantly, these findings extend recent work using support vectormachines and linear discriminant analysis (see Klöppel et al., 2008,2009; Rizk-Jackson et al., 2011) by focusing onmulti-modal neurodegen-erative changes within specific ROIs. Our findings have demonstratedvery good discriminative accuracy in pre-HD from the simultaneous

90 N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

consideration of both structural andmicrostructuralMRImeasures. How-ever, increased multivariate discrimination accuracy was achieved fromcombined MRI, motor (speeded tapping, self-paced tapping) andneurocognitive measures (SDMT, STROOP, UPSIT). While self-paced tap-ping and SDMTmeasures significantly differed between pre-HD and con-trol participants, QDA discriminations based on motor or neurocognitivemeasures separately performed only modestly. This suggests that inpre-HD participants, even though motor and neurocognitive changesare likely to bemore subtle, increased sensitivitymaybegenerated inneu-roimaging studies with small sample sizes (i.e., nb35) by multivariateand multi-modal analysis.

Our MVA results contribute importantly to previous findings. Forexample, two previous studies have examined pre-HD classificationfrom T1 volume or diffusion measures when considered separately(Klöppel et al., 2008, 2009), with diffusion data providing greaterclassification accuracy (76% and 82%, respectively). More recent results(Rizk-Jackson et al., 2011) have enabled the comparison of classificationmethodology (support vector machine [SVM] and linear discriminantanalysis [LDA]) and MRI modality (GM volume, cortico-spinal FA andBOLD response). These studies have reported that the greatest accuracyin distinguishing pre-HD from controls (83% correct) was based on LDAof regional volume in the BG. In our sample, inhomogeneity inmultivar-iate covariances has suggested that LDA may not be reliable; however,ourQDAfindings, based on individual and combinedMRImodality clas-sification accuracy, suggest the potential benefit from adopting a multi-modal approach. Considering only single modality discrimination, ourresults suggest that MD discrimination in BG ROIs was more accuratethan for models based only on volume or FA. This finding is in accordwith earlier suggestions that DTI may be a particularly sensitive markerfor assessing neurodegenerative change in pre-HD individuals far fromonset (Klöppel et al., 2008).

Our QDA findings were derived using an iterated leave one out crossvalidation procedure. This represents a conservative approach, which insmall samples may slightly under-estimate discriminative accuracy(Hastie et al., 2009; Steyerberg, 2009; van der Schaaf, et al., in press). Itis also important to note that our cross validation (iterated leave oneout) is only one alternative from a family of available k-fold andbootstrapping methods (Chernick, 2007; Golub et al., 1979; Tinslet andBrown, 2000). It would therefore be informative for future research to di-rectly contrast differentMVA and cross validation approaches (for exam-ple LDA, QDA, SVM) with larger sample sizes (see Misaki et al., 2010) todetermine how thesemethods compare. Moreover, changes in classifica-tion accuracy across different points of the longitudinal time courseshould also be examined because there may be different combinationsof imaging modalities sensitive at particular stages due to the variationin neuropathological characteristics at particular points in the diseasecontinuum. These changes are likely to progress at different rates, andmay show regional specificity, being visible within different neural struc-tures at different points in time (van den Bogaard et al., 2010). Forexample Mühlau et al. (2007) suggested a likely left sided asymmetryin the progression of atrophy in symp-HD,while Rosas et al. (2002) dem-onstrated early cortical ribbon thinning which proceeded from posteriorto anterior cortical regions with disease progression. Differences in opti-mal sensitivity across imaging modalities, and neural regions, providestrong motivation for further multi-modal imaging studies aimed at bio-marker discovery in HD.

A strength of the method adopted in this paper is the relative sim-plicity of the approach. Unlike recentmachine learning classification ap-proaches in HD (see Klöppel et al., 2008, 2009; Rizk-Jackson et al., 2011),in which many millions of data points are required to inform the statis-ticalmodel, our QDAmethod restricted inputs to between 3 and 18 vari-ables and therefore may represent a more parsimonious multivariatemodel of neurodegenerative changes in pre-HD individuals. High di-mensionality inmultivariate discriminative analyses increases thepossi-bility of over-fitting classifiers. Our approach of dimension reduction,and regional feature extraction, produced very good classification

accuracy while limiting the model complexity. Direct comparison ofROI linear discrimination and support vector machine classificationfromwhole brain volumes also highlights the strengths of informed fea-ture selection (Rizk-Jackson et al., 2011). This approach is possible in HDbecause of evidence of localized neurodegeneration within the BG,whereas additional changes elsewhere in the brain have yet to be con-sistently and repeatedly demonstrated. While feature selection, derivedfrom voxel based morphometric differences between pre-HD and con-trols, may allow the weighting of input variables toward BG structures(Klöppel et al., 2009), while leaving selection open to other regional dif-ferences, the inclusion of isolated butmathematically informative voxelsmay be difficult to explain neuropathologically. Alternatively, indepen-dent component analysis may also provide useful dimension reductionand/or feature selection, while imposing minimum spatial sparsity (orcontiguity) requirements during automated, voxel based feature selectionwhich may increase biological plausibility (see Ashburner and Klöppel,2011). Moreover, ourmethod incorporated a fully automated approachfor the identification of subcortical structures. Achieving highly sensi-tive, fully automated measures of BG structures is a recognised chal-lenge (Morey et al., 2010), compounded in disorders such as HD(Douaud et al., 2006). We explored a range of available automatedapproaches before developing a novel method which combined proba-bilistic Bayesian estimation of individual deformations (Ashburner &Friston 2005), with individualised erosion of periventricular boundariesto provide accurate measures of regional volume and diffusion.

The current IMAGE-HD study investigated volumetric and diffusionchanges in symp-HD, pre-HD and controls, and adopted a QDA, as amultivariate classifier, applied to three measures (volume, FA, MD) todifferentiate pre-HD from control individuals. Firstly, the findings dem-onstrated structural and microstructural signs of neurodegeneration inboth symp-HD and pre-HD individuals compared with controls, andpre-HD differences that were characterised by reduced volume and in-creased FA within the BG. Secondly, multivariate QDA best discriminat-ed pre-HD participants when it was based on multiple-modality data.While MD provided the most accurate discrimination based on a singleMRI modality, our findings demonstrate the advantages provided by amulti-modality approach. Our analysis also combined dimension reduc-tion (mean values) and specific feature selection (ROIs) with automatedidentification of subcortical structures, andprovides evidence to supportviability of multivariate models with restricted input predictors. Impor-tantly, we have shown that QDA is able to classify individuals patho-logically up to 15 years prior to disease onset. These findings provideevidence that combined multi-modal imaging measures maybe a sensi-tivemeans bywhich to develop a neuroimaging based biomarker that isable to accurately classify pre-HD participants from controls very earlyin the course of the disease trajectory. Multivariate approaches offernew and exciting opportunities for identifying early neurodegenerativechange, with clear utility for longitudinal clinical trials many years priorto disease manifestation, where therapeutic potential maybe greatest.

Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.nbd.2012.10.001.

Acknowledgements

Wewould like to acknowledge the contribution of all the participantswho took part in this study.We are also grateful to the CHDI Foundation,Inc. (USA), for their support in funding this research. We also thank theRoyal Children's Hospital for the use of their 3 T MR scanner. GFE is aPrincipal NHMRC Research Fellow.

References

Albin, R.L., Reiner, A., Anderson, K.D., Dure Iv, L.S., Handelin, B., Balfour, R., et al., 1992.Preferential loss of striato-external pallidal projection neurons in presymptomaticHuntington's disease. Ann. Neurol. 31 (4), 425–430.

91N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

Antonini, A., Leenders, K.L., Spiegel, R., Meier, D., Vontobel, P., Weigell-Weber, M., et al.,1996. Striatal glucose metabolism and dopamine D2 receptor binding in asymptom-atic gene carriers and patients with Huntington's disease. Brain 119 (6), 2085–2095.

Ashburner, J., Friston, K.J., 2005. Unified segmentation. NeuroImage 26 (3), 839–851.Ashburner, J., Klöppel, S., 2011. Multivariate models of inter-subject anatomical vari-

ability. NeuroImage 56 (2), 422–439.Aylward, E.H., Brandt, J., Codori, A.M., Mangus, R.S., Barta, P.E., Harris, G.J., 1994. Re-

duced basal ganglia volume associated with the gene for Huntington's disease inasymptomatic at-risk persons. Neurology 44 (5), 823–828.

Aylward, E.H., Codori, A.M., Barta, P.E., Pearlson, G.D., Harris, G.J., Brandt, J., 1996. Basalganglia volume and proximity to onset in preclinical Huntington's disease. Arch.Neurol. 53, 1293–1296.

Aylward, E.H., Codori, A.M., Rosenblatt, A., Sherr, M., Brandt, J., Stine, O.C., et al., 2000.Rate of caudate atrophy in presymptomatic and symptomatic stages ofHuntington's disease. Mov. Disord. 15 (3), 552–560.

Aylward, E.H., Sparks, B.F., Field, K.M., Yallapragada, V., Shpritz, B.D., Rosenblatt, A., etal., 2004. Onset and rate of striatal atrophy in preclinical Huntington disease. Neu-rology 63 (1), 66–72.

Bäckman, L., Robins-Wahlin, T.B., Lundin, A., Ginovart, N., Farde, L., 1997. Cognitive def-icits in Huntington's disease are predicted by dopaminergic PET markers and brainvolumes. Brain 120 (12), 2207–2217.

Beck, A.T., Steer, R.A., Brown, G.K., 1996. Manual for the Beck Depression Inventory—II.Psychological Corporation, San Antonio.

Bohanna, I., Georgiou-Karistianis, N., Egan, G.F., 2011. Connectivity-based segmenta-tion of the striatum in Huntington's disease: vulnerability of motor pathways.Neurobiol. Dis. 42 (3), 475–481.

Bohanna, I., Georgiou-Karistianis, N., Hannan, A.J., Egan, G.F., 2008. Magnetic resonanceimaging as an approach towards identifying neuropathological biomarkers forHuntington's disease. Brain Res. Rev. 58 (1), 209–225.

Bots, T., Bruyn, G.W., 1981. Neuropathological changes of the nucleus accumbens inHuntington's chorea. Acta Neuropathol. 55 (1), 21–22.

Brandt, J., Bylsma, F.W., Aylward, E.H., Rothlind, J., Gow, C.A., 1995. Impaired sourcememory in Huntington's disease and its relation to basal ganglia atrophy. J. Clin.Exp. Neuropsychol. 17 (6), 868–877.

Campodonico, J.R., Aylward, E., Codori, A.M., Young, C., Krafft, L., Magdalinski, M., et al.,1998. When does Huntington's disease begin? J. Int. Neuropsychol. Soc. 4 (5),467–473.

Chernick, M.R., 2007. Bootstrap Methods: A Guide for Practitioners and Researchers.John Wiley and Sons, New Jersey.

De la Monte, S.M., Vonsattel, J.P., Richardson Jr., E.P., 1988. Morphometric demonstra-tion of atrophic changes in the cerebral cortex, white matter, and neostriatum inHuntington's disease. J. Neuropathol. Exp. Neurol. 47 (5), 516–525.

Della Nave, R., Ginestroni, A., Tessa, C., Giannelli, M., Piacentini, S., Filippi, M., Mascalchi,M., 2010. Regional distribution and clinical correlates of white matter structuraldamage in Huntington disease: a tract-based spatial statistics study. AJNR Am. J.Neuroradiol. 31 (9), 1675–1681.

Doty, R.L., Shaman, P., Kimmelman, C.P., Dann, M.S., 1984. University of Pennsylvaniasmell identification test: a rapid quantitative olfactory function test for the clinic.Laryngoscope 94 (2), 176–178.

Douaud, G., Behrens, T.E., Poupon, C., Cointepas, Y., Jbabdi, S., Gaura, V., et al., 2009. Invivo evidence for the selective subcortical degeneration in Huntington's disease.NeuroImage 46 (4), 958–966.

Douaud, G., Gaura, V., Ribeiro, M.J., Lethimonnier, F., Maroy, R., Verny, C., et al., 2006.Distribution of grey matter atrophy in Huntington's disease patients: a combinedROI-based and voxel-based morphometric study. NeuroImage 32 (4), 1562–1575.

Esmaeilzadeh, M., Ciarmiello, A., Squitieri, F., 2011. Seeking brain biomarkers for pre-ventive therapy in huntington disease. CNS Neurosci. Ther. 00, 1–19.

Fennema-Notestine, C., Archibald, S.L., Jacobson, M.W., Corey-Bloom, J., Paulsen, J.S.,Peavy, G.M., et al., 2004. In vivo evidence of cerebellar atrophy and cerebralwhite matter loss in Huntington disease. Neurology 63 (6), 989–995.

Georgiou-Karistianis, N., 2009. A peek inside the Huntington's brain: will functionalimaging take us one step closer in solving the puzzle? Exp. Neurol. 220 (1), 5–8.

Golub, G., Heath, M., Wahba, G., 1979. Generalized cross-validation as a method forchoosing a good ridge parameter. Technometrics 21, 215–224.

Grace, J., Mallory, P.F., 2001. Frontal Systems Behavior Scale: Professional Manual. Psy-chological Assessment Resources, Lutz.

Harris, G.J., Codori, A.M., Lewis, R.F., Schmidt, E., Bedi, A., Brandt, J., 1999. Reduced basalganglia blood flow and volume in pre-symptomatic, gene-tested persons at-riskfor Huntington's disease. Brain 122 (9), 1667–1678.

Harris, G.J., Pearlson, G.D., Peyser, C.E., Aylward, E.H., Roberts, J., Barta, P.E., et al., 1992.Putamen volume reduction on magnetic resonance imaging exceeds caudatechanges in mild Huntington's disease. Ann. Neurol. 31 (1), 69–75.

Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning: DataMining, Inference and Prediction, 2nd ed. Springer.

Henley, S.M.D., Wild, E.J., Hobbs, N.Z., Scahill, R.I., Ridgway, G.R., MacManus, D.G., et al.,2009. Relationship between CAG repeat length and brain volume in premanifestand early Huntington's disease. J. Neurol. 256 (2), 203–212.

Hicks, S., Rosas, H.D., Berna, C., Scahill, R., Durmas, E., Roos, R.A., et al., 2010. PAW36 oc-ulomotor deficits in presymptomatic and early Huntington's disease and theirstructural brain correlates. J. Neurol. Neurosurg. Psychiatry 81 (11), e33.

Hobbs, N.Z., Henley, S.M., Ridgway, G.R., Wild, E.J., Barker, R.A., Scahill, R.I., Barnes, J.,Fox, N.C., Tabrizi, S.J., 2010. The progression of regional atrophy in premanifestand early Huntington's disease: a longitudinal voxel-based morphometry study.J. Neurol. Neurosurg. Psychiatry 81 (7), 756–763.

Huntington Study Group, 1996. Unified Huntington's disease rating scale: reliabilityand consistency. Mov. Disord. 11 (2), 136–142 (Mar).

Jernigan, T.J., Salmon, D.P., Butters, N., Hesselink, J.R., 1991. Cerebral structure on MRI,part ii: specific changes in Alzheimer's and Huntington's diseases. Biol. Psychiatry29 (1), 68–81.

Jurgens, C.K., Van De Wiel, L., Van Es, A.C.G.M., Grimbergen, Y.M., Witjes-Ané, M.N.W.,Van Der Grond, J., et al., 2008. Basal ganglia volume and clinical correlates in ‘pre-clinical’ Huntington's disease. J. Neurol. 255 (11), 1785–1791.

Kassubek, J., Juengling, F.D., Ecker, D., Landwehrmeyer, G.B., 2005. Thalamic atrophy inHuntington's disease co-varies with cognitive performance: a morphometric MRIanalysis. Cereb. Cortex 15 (6), 846–853.

Kassubek, J., Juengling, F.D., Kioschies, T., Henkel, K., Karitzky, J., Kramer, B., et al., 2004.Topography of cerebral atrophy in early Huntington's disease: a voxel based mor-phometric MRI study. J. Neurol. Neurosurg. Psychiatry 75 (2), 213–220.

Kipps, C.M., Duggins, A.J., Mahant, N., Gomes, L., Ashburner, J., McCusker, E.A., 2005.Progression of structural neuropathology in preclinical Huntington's disease: atensor based morphometry study. J. Neurol. Neurosurg. Psychiatry 76 (5),650–655.

Klöppel, S., Chu, C., Tan, G.C., Draganski, B., Johnson, H., Paulsen, J.S., et al., 2009. Auto-matic detection of preclinical neurodegeneration: presymptomatic Huntingtondisease. Neurology 72 (5), 426–431.

Klöppel, S., Draganski, B., Golding, C.V., Chu, C., Nagy, Z., Cook, P.A., et al., 2008. Whitematter connections reflect changes in voluntary-guided saccades in pre-symptomatic Huntington's disease. Brain 131 (1), 196–204.

Langbehn, D.R., Brinkman, R.R., Falush, D., Paulsen, J.S., Hayden, M.R., 2004. A newmodel for prediction of the age of onset and penetrance for Huntington's diseasebased on CAG length. Clin. Genet. 65 (4), 267–277.

Mascalchi, M., Lolli, F., Della Nave, R., Tessa, C., Petralli, R., Gavazzi, C., et al., 2004.Huntington disease: volumetric, diffusion-weighted, and magnetization transferMR imaging of brain. Radiology 232 (3), 867–873.

Misaki, M., Kim, Y., Bandettini, P.A., Kriegeskorte, N., 2010. Comparison of multivariateclassifiers and response normalizations for pattern-information fMRI. NeuroImage53 (1), 103–118.

Morey, R.A., Selgrade, E.S., Wagner II, H.R., Huettel, S.A., Wang, L., McCarthy, G., 2010.Scan-rescan reliability of subcortical brain volumes derived from automated seg-mentation. Hum. Brain Mapp. 31 (11), 1751–1762.

Mühlau, M., Gaser, C., Wohlschläger, A.M., Weindl, A., Städtler, M., Valet, M., et al.,2007. Striatal gray matter loss in Huntington's disease is leftward biased. Mov.Disord. 22 (8), 1169–1173.

Nelson, H.E., Willison, J., 1991. National Adult Reading Test (NART) Test Manual, 2nded. NFER Nelson, Windsor, UK.

Nopoulos, P., Magnotta, V.A., Mikos, A., Paulson, H., Andreasen, N.C., Paulsen, J.S., 2007.Morphology of the cerebral cortex in preclinical Huntington's disease. Am. J. Psy-chiatry 164 (9), 1428–1434.

Patenaude, B., Smith, S.M., Kennedy, D., Jenkinson, M., 2011. A Bayesian model of shapeand appearance for subcortical brain. NeuroImage 56 (3), 907–922.

Paulsen, J.S., 2009. Functional imaging in Huntington's disease. Exp. Neurol. 216 (2),272–277.

Paulsen, J.S., Langbehn, D.R., Stout, J.C., Aylward, E., Ross, C.A., Nance, M., et al., 2008.Detection of Huntington's disease decades before diagnosis: the Predict-HDstudy. J. Neurol. Neurosurg. Psychiatry 79 (8), 874–880.

Paulsen, J.S., Magnotta, V.A., Mikos, A.E., Paulson, H.L., Penziner, E., Andreasen, N.C., etal., 2006. Brain structure in preclinical Huntington's disease. Biol. Psychiatry 59(1), 57–63.

Peinemann, A., Schuller, S., Pohl, C., Jahn, T., Weindl, A., Kassubek, J., 2005. Executivedysfunction in early stages of Huntington's disease is associated with striatal andinsular atrophy: a neuropsychological and voxel-based morphometric study. J.Neurol. Sci. 239 (1), 11–19.

Pereira, F., Mitchell, T., Botvinick, M., 2009. Machine learning and fMRI: a tutorial over-view. NeuroImage 45 (Suppl. 1), S199–S209.

Reading, S.A.J., Yassa, M.A., Bakker, A., Dziorny, A.C., Gourley, L.M., Yallapragada, V., etal., 2005. Regional white matter change in pre-symptomatic Huntington's disease:a diffusion tensor imaging study. Psychiatry Res. 140 (1), 55–62.

Rizk-Jackson, A., Stoffers, D., Sheldon, S., Kuperman, J., Dale, A., Goldstein, J., et al., 2011. Eval-uating imaging biomarkers for neurodegeneration in pre-symptomatic Huntington'sdisease using machine learning techniques. NeuroImage 56 (2), 788–796.

Rosas, H.D., Goodman, J., Chen, Y.I., Jenkins, B.G., Kennedy, D.N., Makris, N., et al., 2001.Striatal volume loss in HD as measured by MRI and the influence of CAG repeat.Neurology 57 (6), 1025–1028.

Rosas, H.D., Koroshetz, W.J., Chen, Y.I., Skeuse, C., Vangel, M., Cudkowicz, M.E., et al.,2003. Evidence for more widespread cerebral pathology in early HD: an MRI-based morphometric analysis. Neurology 60 (10), 1615–1620 (May 27).

Rosas, H.D., Liu, A.K., Hersch, S., Glessner, M., Ferrante, R.J., Salat, D.H., et al., 2002. Re-gional and progressive thinning of the cortical ribbon in Huntington's disease. Neu-rology 58 (5), 695–701 (Mar 12).

Rosas, H.D., Salat, D.H., Lee, S.Y., Zaleta, A.K., Pappu, V., Fischl, B., et al., 2008. Cerebralcortex and the clinical expression of Huntington's disease: complexity and hetero-geneity. Brain 131 (Pt 4), 1057–1068 (Apr).

Rosas, H.D., Tuch, D.S., Hevelone, N.D., Zaleta, A.K., Vangel, M., Hersch, S.M., et al., 2006.Diffusion tensor imaging in presymptomatic and early Huntington's disease: selec-tive white matter pathology and its relationship to clinical measures. Mov. Disord.21 (9), 1317–1325.

Ruocco, H.H., Lopes-Cendes, I., Li, L.M., Santos-Silva, M., Cendes, F., 2006. Striatal andextrastriatal atrophy in Huntington's disease and its relationship with length ofthe CAG repeat. Braz. J. Med. Biol. Res. 39 (8), 1129–1136.

Sánchez-Castañeda, C., Cherubini, A., Elifani, F., Péran, P., Orobello, S., Capelli, G.,Sabatini, U., Squitieri, F., in press. Seeking Huntington disease biomarkers bymulti-modal, cross-sectional basal ganglia imaging. Hum. Brain Mapp.

92 N. Georgiou-Karistianis et al. / Neurobiology of Disease 51 (2013) 82–92

Seppi, K., Schocke,M.F.H., Mair, K.J., Esterhammer, R.,Weirich-Schwaiger, H., Utermann, B., etal., 2006. Diffusion-weighted imaging in Huntington's disease. Mov. Disord. 21 (7),1043–1047.

Smith, A., 1982. Symbol Digit Modality Test (SDMT): Manual (Revised). PsychologicalServices, Los Angeles.

Smith, S.M., Jenkinson, M., Woolrich, M.W., Beckmann, C.F., Behrens, T.E., Johansen-Berg, H., et al., 2004. Advances in functional and structural MR image analysisand implementation as FSL. Neuroimage 23 (Suppl. 1), S208–S219.

Soneson, C., Fontes, M., Zhou, Y., Denisov, V., Paulsen, J.S., Kirik, D., et al., 2010. Earlychanges in the hypothalamic region in prodromal Huntington disease revealedby MRI analysis. Neurobiol. Dis. 40 (3), 531–543.

Sritharan, A., Egan, G.F., Johnston, L., Horne, M., Bradshaw, J.L., Bohanna, I., et al., 2010.A longitudinal diffusion tensor imaging study in symptomatic Huntington's dis-ease. J. Neurol. Neurosurg. Psychiatry 81 (3), 257–262.

Starkstein, S.E., Brandt, J., Bylsma, F., Peyser, C., Folstein, M., Folstein, S.E., 1992. Neuro-psychological correlates of brain atrophy in Huntington's disease: a magnetic res-onance imaging study. Neuroradiology 34 (6), 487–489.

Steyerberg, E.W., 2009. Clinical Prediction Models: A Practical Approach to develop-ment, Validation, and Updating. Springer.

Stoffers, D., Sheldon, S., Kuperman, J.M., Goldstein, J., Corey-Bloom, J., Aron, A.R., 2010.Contrasting gray and white matter changes in preclinical Huntington disease: AnMRI study. Neurology 7 (15), 1208–1216.

Stout, J.C., Paulsen, J.S., Queller, S., Solomon, A.C., Whitlock, K.B., Campbell, J.C., et al., 2011.Neurocognitive signs in prodromal Huntington disease. Neuropsychology 25 (1), 1–14.

Stroop, J.R., 1935. Studies of interference in serial verbal reactions. J. Exp. Psychol. 18 (6),643–662.

Tabrizi, S.J., Langbehn, D.R., Leavitt, B.R., Roos, R.A., Durr, A., Craufurd, D., et al., 2009. Biologicaland clinical manifestations of Huntington's disease in the longitudinal TRACK-HD study:cross-sectional analysis of baseline data. Lancet Neurol. 8 (9), 791–801 (Sep).

Thieben, M.J., Duggins, A.J., Good, C.D., Gomes, L., Mahant, N., Richards, F., et al., 2002.The distribution of structural neuropathology in pre-clinical Huntington's disease.Brain 125 (8), 1815–1828.

Tinslet, H.E.A., Brown, S.D., 2000. Multivariate statistics and mathematical modelling.In: Tinslet, H., Brown, S. (Eds.), Handbook of Applied Multivariate Statistics andMathematical Modelling. Academic Press, California.

van den Bogaard, S.J.A., Dumas, E.M., Acharya, T.P., Johnson, H., Langbehn, D.R., Scahill,R.I., et al., 2010. Early atrophy of pallidum and accumbens nucleus in Huntington'sdisease. J. Neurol. 1–9.

van der Schaaf, A., Xu, C.-J, van Luijk, P, van't Veld, A.A, Langendijk, J.A, Schilstra, C., inpress. Multivariate modeling of complications with data driven variable selection:Guarding against overfitting and effects of data set size. Radiother. Oncol.

Vonsattel, J.P., Myers, R.H., Stevens, T.J., Ferrante, R.J., Bird, E.D., Richardson Jr., E.P.,1985. Neuropathological classification of Huntington's disease. J. Neuropathol.Exp. Neurol. 44 (6), 559–577 (Nov).

Watson, D., Wu, K.D., 2005. Development and validation of the schedule of compulsions,obsessions, and pathological impulses (SCOPI). Assessment 12 (1), 50–65.

Zigmond, A.S., Snaith, R.P., 1983. The hospital anxiety and depression scale. ActaPsychiatr. Scand. 67 (6), 361–370.