Article Text

Original research
Prediction of dementia using diffusion tensor MRI measures: the OPTIMAL collaboration
  1. Marco Egle1,
  2. Saima Hilal2,3,
  3. A M Tuladhar4,
  4. Lukas Pirpamer5,
  5. Edith Hofer5,6,
  6. Marco Duering7,8,
  7. James Wason9,10,
  8. Robin G Morris11,
  9. Martin Dichgans7,8,12,
  10. Reinhold Schmidt5,
  11. Daniel Tozer1,
  12. Christopher Chen2,
  13. Frank-Erik de Leeuw4,
  14. Hugh S Markus1
  1. 1 Stroke Research Group, Department of Clinical Neurosciences, University of Cambridge, Cambridge, UK
  2. 2 Memory Aging and Cognition Centre, Department of Pharmacology, Yong Loo Lim School of Medicine, National University of Singapore, Singapore
  3. 3 Saw Swee Hock School of Public Health, National University of Singapore and National University Health System of Singapore, Singapore
  4. 4 Radboud University Medical Center, Donders Institute for Brain, Cognition and Behaviour, Department of Neurology, Nijmegen, The Netherlands
  5. 5 Department of Neurology, Medical University Graz, Graz, Austria
  6. 6 Institute for Medical Informatics, Statistics and Documentation, Medical University of Graz, Graz, Austria
  7. 7 Institute for Stroke and Dementia Research, University Hospital, Ludwig Maximilian University Munich, Munich, Germany
  8. 8 Munich Cluster for Systems Neurology (SyNergy), Munich, Germany
  9. 9 MRC Biostatistics Unit, School of Clinical Medicine, University of Cambridge, Cambridge Institute of Public Health, Cambridge, UK
  10. 10 Population Health Sciences Institute, Newcastle University, Baddiley-Clark Building, Newcastle upon Tyne, UK
  11. 11 Department of Psychology (R.G.M), King's College London, Institute of Psychiatry, Psychology and Neuroscience, London, UK
  12. 12 German Center for Neurodegenerative Diseases (DZNE), Munich, Germany
  1. Correspondence to Professor Hugh S Markus, Department of Clinical Neurosciences, University of Cambridge, Cambridge, Cambridgeshire, UK; hsm32{at}


Objectives It has been suggested that diffusion tensor imaging (DTI) measures sensitive to white matter (WM) damage may predict future dementia risk not only in cerebral small vessel disease (SVD), but also in mild cognitive impairment. To determine whether DTI measures were associated with cognition cross-sectionally and predicted future dementia risk across the full range of SVD severity, we established the International OPtimising mulTImodal MRI markers for use as surrogate markers in trials of Vascular Cognitive Impairment due to cerebrAl small vesseL disease collaboration which included six cohorts.

Methods Among the six cohorts, prospective data with dementia incidences were available for three cohorts. The associations between six different DTI measures and cognition or dementia conversion were tested. The additional contribution to prediction of other MRI markers of SVD was also determined.

Results The DTI measure mean diffusivity (MD) median correlated with cognition in all cohorts, demonstrating the contribution of WM damage to cognition. Adding MD median significantly improved the model fit compared to the clinical risk model alone and further increased in all single-centre SVD cohorts when adding conventional MRI measures. Baseline MD median predicted dementia conversion. In a study with severe SVD (SCANS) change in MD median also predicted dementia conversion. The area under the curve was best when employing a multimodal MRI model using both DTI measures and other MRI measures.

Conclusions Our results support a central role for WM alterations in dementia pathogenesis in all cohorts. DTI measures such as MD median may be a useful clinical risk predictor. The contribution of other MRI markers varied according to disease severity.

  • dementia
  • alzheimer's disease
  • cerebrovascular disease
  • MRI
  • vascular dementia

Data availability statement

Data are available on reasonable request.

Statistics from


Cerebral small vessel disease (SVD) causes a quarter of all strokes and is the most common pathology underlying vascular cognitive impairment and dementia.1 Conventional MRI shows characteristic features including lacunar infarcts, white matter hyperintensities (WMH), cerebral microbleeds (CMB) and enlarged perivascular spaces.2

It is hypothesised that cognitive impairment results from disruption of neuronal networks due to damage to WM tracts providing their connections, and that this either disrupts efficient processing of information between neurocognitive systems or damages control mechanisms coordinating cognition.3 A key role of WM damage is suggested by population-based studies showing that WMH burden predicts stroke and dementia.4 A more sensitive method of identifying WM ultrastructural damage and tract alterations is diffusion tensor imaging (DTI). Studies have shown altered structural integrity within the WM and in areas outside WMH.5 Cross-sectional studies in SVD show that DTI measures correlate more strongly with cognition than WMH or other conventional SVD markers.6–8 DTI measures have been shown to predict progression to dementia,9 but previous prediction studies have been largely limited to single-centre cohorts.

The use of DTI measures may provide better prediction of who will progress to dementia, allowing those at higher risk to be targeted for therapeutic interventions. Recent image analysis techniques allow rapid assessment of DTI measures to be calculated, making it a realistic biomarker for clinical prediction.8 However, prior to its more widespread use, it is important to demonstrate that DTI measures indeed predict dementia across populations, including patients with varying severity of SVD. Furthermore other markers such as lacunes, WMH, atrophy and CMB have also been associated with cognition in SVD.7 10 It is important to determine the relative importance of these markers. A combination of DTI and other MRI measures may be needed reflecting the disease heterogeneity beyond microstructural WM damage.

To assess the predictive value of DTI measures across the full range of SVD severity, we formed a collaboration in six cohorts to determine correlations between DTI measures and cognition, and whether baseline and changes in DTI measures predicted dementia conversion. We further determined whether adding conventional MRI markers increased prediction.


The OPtimising mulTImodal MRI markers for use as surrogate markers in trials of Vascular Cognitive Impairment due to cerebrAl small vesseL disease (OPTIMAL) collaboration was established to determine whether DTI measures predicted dementia in SVD across multiple cohorts.11


Six cohorts with differing degrees of SVD severity were included

  1. Severe symptomatic SVD (SCANS)

  2. Severe symptomatic SVD (PRESERVE)

  3. Moderate SVD (RUN DMC).

  4. Mild cognitive impairment (MCI) (HARMONISATION).

  5. Elderly stroke free population based cohort (ASPS-Fam).

  6. Monogenic SVD (CADASIL).

Informed consent was obtained from participants. Details were published previously and are summarised in table 1.

Table 1

Overview about the cohort studies included in the OPTIMAL project

Severe symptomatic SVD cohort (SCANS)

MRI scanning took place at baseline and over 3 yearly follow-up sessions on a 1.5 T General Electric Signa HDxt MRI system.9 Acquisition parameters are presented in online supplemental table 1. Cognition was measured using standardised tests sensitive to SVD.12 Age-standardised test scores were used to form a measure of Global Cognition (online supplemental table 2).

Supplemental material

Multicentre severe symptomatic SVD cohort (PRESERVE)

MRI acquisition took place on eight 3-Tesla MRI scanners (3 Philips Achieva TX, 1 Philips Achieva, 1 Philips Ingenia, 1 Siemens Verio, 1 Siemens Prisma, 1 Siemens Magnetom Prisma fit)13 as part of the PRESERVE trial. Acquisition parameters across scanners were harmonised (online supplemental table 1). Neuropsychological scores were age standardised to create Global Cognition (online supplemental table 2).

Moderate SVD cohort (RUN DMC)

MRI acquisition was performed on a 1.5 T Siemens Magnetom Avanto MRI machine,14 using parameters listed in online supplemental table 1). Due to scanner updates, only data from 2011 and 2015 were used for longitudinal analysis, as DTI is sensitive to scanner changes. 258 dementia-free patients with DTI and conventional MRI were still enrolled in 2011. Cognition was assessed using validated cognitive tasks. Age-standardised scores were used to compute a Global Cognition score (online supplemental table 2).

MCI cohort (Harmonisation)

Imaging data were acquired on a 3T Siemens Magnetom Trio Tim system (online supplemental table 1). Baseline and 2 years follow-up DTI data were available.15 Patients were tested on neuropsychological test batteries previously validated. Scores were standardised to the mean and SD to form a measure of Global Cognition. More details regarding the test can be found in online supplemental table 2.

Elderly stroke free population-based cohort (ASPS-Fam)

Magnetic resonance acquisition was performed on a 3T Tim Trio whole body scanner (online supplemental table 1). Only baseline DTI measures were used.16 Global Cognition was created with age-standardised test scores (online supplemental table 2).

Monogenic cohort (CADASIL)

Imaging data were based on a 1.5 T GE Signa system in Munich (online supplemental table 1).8 To measure the main cognitive deficit in these patients, that is processing speed, the Trail-making test–B was used. The main outcome score was transformed to z-scores using normative data from the literature (online supplemental table 2).17

MRI processing

Brain volume and WMH volume

SCANS: Brain volume (BV) was computed from native T1 images as an estimate of brain size relative to the skull size with SIENAX as part of the FMRIB software library.7 WMH and brain tissues segmentations were carried out using the methods previously described.18 A measure of WMH, called SVDp, was calculated by taking the ratio of WMH volume to the total cerebral volume composed of the sum of grey matter (GM), WM and WMH.

RUN DMC: Tissue probability maps were calculated employing Statistical Parametric Mapping 12 (SPM12) unified segmentation routine on the T1 MPRAGE images.19 The GM and WM volumes derived from the sum of the all voxel volumes were added creating a measure of BV. WMH volumes were calculated by a WMH segmentation method20 and the same measure as in SCANS was used (SVDp).

PRESERVE: T1W scans were segmented into GM, WM and cerebrospinal fluid (CSF) tissue probability maps employing SPM12b ( Soft segmentations of the GM and WM tissue probability maps were used to compute BV and an estimate of brain size relative to the skull size with SIENAX was calculated.13 A semiautomated programme JIM (Xinapse Systems; was used to segment WMH regions. Brain lesion maps were used to compute lesion load as the percentage of WMH lesion volume against BV.13

HARMONISATION: Image preprocessing and the tissue classification algorithm have been described elsewhere.21 Briefly, a k-nearest-neighbour technique was used to classify voxels into CSF, GM and normal appearing WM and volumes were calculated from these measurements. WMH volumes were detected using an adapted threshold technique.21 Intracranial volume was the sum of the CSF, GM, normal WM and WMH.

ASPS-FAM: Intracranial volume was estimated by FreeSurfer V.6.0. A custom-written Interactive Data Language programme was used to create WMH maps. Segmentation was performed on Fluid-attenuated inversion recovery (FLAIR) images by 2 two raters.16

CADASIL: BV was computed from native T1 images as an estimate of size of the brain relative to the skull size with SIENAX as part of the FMRIB software library.8 WMH segmentations were generated by a semiautomatic pipeline and corrected by trained raters.8

Lacune count and CMB count

Lacune and CMB were identified by trained raters blinded to the data.

In SCANS, RUN DMC, HARMONISATION and PRESERVE lacunes were defined as CSF filled cavity within WM or subcortical areas between 3 and 15 mm in diameter on T1-weighted and FLAIR images.2 In ASPS-Fam lacunes were graded as focal lesions with maximum diameter of 10 mm. In CADASIL lacunes were detected on T1-weighted images with a signal identical to CSF, sharp delineation and diameter >2 mm.

in SCANS, RUN DMC and PRESERVE CMB were identified on T2* weighted Gradient echo (GRE) images as focal spots up to 10 mm in diameter.7 10 In HARMONISATION rating of the CMB was done on susceptibility-weighted images employing the Microbleed Anatomical Rating Scale.22 23 In ASPS-Fam CMB were defined as homogeneous rounded lesions with a diameter of 2–5 mm. In CADASIL CMB were graded on GRE sequences as rounded foci <5 mm in diameter.

DTI processing

Six WM histogram measures were computed: mean diffusivity (MD) median (MD median), peak height (PH) and peak location (pkval) and fractional anisotropy (FA) median (FA median), FAPH and FA pkval. The image analysis pipeline has been described previously13 24 25 and key steps are summarised.

In SCANS, PRESERVE, HARMONISATION and CADASIL the DTI analyses were carried using a standardised protocol either at the central site (SCANS. PRESERVE, HARMONISATION, CADASIL) or under guidance by the central site (ASPS-Fam). The eddy correct software from ‘FDT’, FMRIB’s Diffusion Toolbox, ( was employed for DTI preprocessing. In RUN DMC diffusion data analysis was carried prior to OPTIMAL and the imaging data were preprocessed employing in-house developed iteratively reweighted least squares algorithm at the Radboud University Medical Center in Nijmegen.26 Susceptibility distortions were unwarped by normalising images to the T1 images in the phase-encoding direction. For all datasets MD and FA maps were created with ‘DTIFIT’. FLAIR to T1 weighted and T1 weighted to b0 registrations were conducted and the affine transformation matrices were concatenated to generate a FLAIR-to-DTI transformation.27 Tissue probability maps were registered into DTI space using these transformations. Hard segmentations were used creating maps of tissue classes. Histogram analysis was conducted on the MD and FA maps in in all WM regions. Summary histogram measures were derived from normalised histograms with 1000 bins.

Statistical analysis

The statistical analysis was conducted using R software (V.3.6.3).

Analyses of cross-sectional baseline data

The relationship between the DTI measures and cognition was tested using linear regression controlling for clinical markers (age, sex and premorbid IQ or education). In PRESERVE, the associations were controlled for studysite-related differences. Using a model decomposition method, the amount of variance in cogntion explained was determined.28 In the CADASIL sample, one patient was excluded as an outlier and the dependent variable was power transformed to meet the statistical assumptions. The statistical assumptions of the linear model were normality of the residuals, homoscedasticity, linearity of the relationship and no outlier values which were checked through visual plot inspection.

To determine the best model fit, Akaike information criterion (AIC) together with the adjusted variance (Adj R2) were computed for the following linear regression models: clinical markers only (Clinical), clinical markers plus DTI measures (Clinical-DTI), clinical markers plus conventional MRI markers (BV, WMH, Lacune and CMB) (Clinical-MRI) and clinical markers plus both conventional MRI and DTI measures (Clinical-MRI-DTI). Again, the linear models’ assumptions were checked by visual inspection. WMH, lacune and CMB were log10 transformed. Multiple regression models were checked for multicollinearity with the variance inflation factor (VIF) being below a VIF value threshold (VIF <10).29 To estimate the relative importance of each predictor in the clinical-MRI-DTI model in explaining the overall variance in cognition, a model decomposition method was employed.28

Analyses of longitudinal data

The association between baseline DTI measures and dementia conversion was tested using a Cox regression model in SCANS, RUN DMC and HARMONISATION. The clinical markers were added as confounders. The area under the curve (AUC) was computed.30 To compare the predictive value of DTI measures to cognitive measures, a separate Cox regression model with the clinical markers and global cognition as covariates (Clinical-Global) was computed. The AUC of the Clinical-DTI model was compared with the Clinical-Global model in each cohort.30

To determine change in the markers, the difference between two time points was determined using a paired t-test in RUN DMC and HARMONISATION. In SCANS, characterised by four time points, change was estimated by a linear mixed model.31 The intercept and slope of each participant’s linear trajectory were allowed to vary with both fixed and random effects. Fixed effect variation was accounted for by time, and random effect variation allowed for remaining interindividual differences. Change in lacunes and CMB were categorised as dichotomous; that is, change versus no change in counts. Follow-up observations postdementia diagnosis were removed prior to the model’s computation.

Employing a Cox or logistic regression the association between the change in DTI measures and dementia was tested while accounting for the clinical markers. Using AIC the DTI model was compared with the clinical model, to the clinical-MRI model and to clinical-MRI-DTI model. The AUC was computed for each model and compared within cohorts.30


The cohorts’ demographic characteristics are shown in table 2.

Table 2

Demographic, imaging and cognition measures in each cohort studies

Cross-sectional association between imaging markers and cognition

The DTI measure MD median showed the most consistent associations with the largest effect size, being related to cognition in all cohorts (table 3), and was taken forward as the DTI marker used in further analyses. The Clinical-DTI model explained 0.454–0.516 of variance in the single non-monogenic centre cohorts, but less variance in the multicentre cohort (adj R2=0.373) (table 3). The adjusted variance in the CADASIL cohort was 0.235 due the low variance explained by the clinical model in this cohort. DTI’s normalised R2 contribution was highest in CADASIL (figure 1).

Figure 1

Normalised contribution of the individual marker to the unadjusted explained variance in cognition in the Clinical-DTI model. In the sporadic SVD and MCI cohorts normalised DTI’s contribution to the overall explained variance varied around 20%. DTI’s contribution to the model was significant higher for CADASIL. Across all non-monogenic cohorts clinical markers explained a significant proportion of the model’s explained variance. Panels: A=SCANS, B=RUN DMC, C=HARMONISATION, D=PRESERVE, E=ASPS Fam, F=CADASIL. Lines refer to 95% CIs after bootstrapping. Overall unadjusted R2 variance explained by the Clinical-DTI model: SCANS=0.496, RUN DMC=0.458, HARMONISATION=0.474, PRESERVE=0.409, ASPS-Fam=0.524, CADASIL=0.299 MD median=median of the normalised mean diffusivity histogram, NART= premorbid IQ, Educ= education, R2= explained variance of the model, DTI= diffusion tensor imaging; MCI= mild cognitive impairment; MD= mean diffusivity; SVD= small vessel disease.

Table 3

Cross-sectional analysis between DTI measures and Global Cognition or TMT-B (CADASIL) independently of the clinical markers

The variance in cognition explained by the different models is shown in table 4. The additional variance explained by adding conventional MRI markers to the DTI-clinical model varied but was greatest in SCANS and CADASIL, while little or no additional variance was added in HARMONISATION and ASPS-Fam. In all models VIF was below 3. Relative contributions of the different imaging markers of SVD to cognition in the Clinical-DTI-MRI model were determined (figure 2). This varied according to SVD severity. In SCANS and in CADASIL, this was greatest for BV, lacunes and the DTI measure. In contrast, in milder SVD (ASPS-FAM and HARMONISATION), the DTI measure contributed most, and BV and lacunes contributed little (figure 2). Of the clinical markers, age was the most important predictor in the ASPS-Fam, whereas education or premorbid IQ explained most variance in the non-monogenic cohorts.

Figure 2

Normalised contribution of the individual marker to the unadjusted explained variance in cognition in the Clinical-MRI-diffusion tensor imaging (DTI) model. In single-centre small vessel disease (SVD) () cohorts (SCANS, RUN DMC, CADASIL), brain volume (BV) and DTI explained a significant variance among all imaging markers. The importance of lacune count as factor varied according to the vascular disease severity with being high in SCANS, PRESERVE and CADASIL and low in RUN DMC, HARMONISATION and ASPS-Fam. In all but ASPS-Fam premorbid IQ explained most of the variance among the clinical markers. The importance of DTI was high and of clinical markers low in CADASIL. Cerebral microbleed (CMB) was not included in the model for ASPS-Fam. Panels: A=SCANS, B=RUN DMC, C=HARMONISATION, D=PRESERVE, E=ASPS Fam, F=CADASIL. Lines refer to 95% CIs after bootstrapping. CMB was excluded in the ASPS-Fam Clinical-MRI-DTI model. Overall unadjusted R2 variance explained by the Clinical-MRI-DTI model: SCANS=0.595, RUN DMC=0.475, HARMONISATION=0.483, PRESERVE=0.453, ASPS-Fam=0.538, CADASIL=0.426. MD median=median of the normalised MD histogram, Lac=lacune count, NART=premorbid IQ score, R2=explained variance of the model. MD= mean diffusivity; WMH= white matter hyperintensity.

Table 4

Predictive model comparisons at baseline

Longitudinal association between baseline imaging marker and dementia conversion

Dementia incidence together with DTI measures was available for three studies: SCANS, 5 years follow-up, 20 (17%) dementia cases, all vascular dementia (VD); RUN DMC, 9 years follow-up, 65 (13%) dementia cases, (21 VD), 30 Alzheimer’s disease (AD), 10 AD/VD, 1 Frontotemporal dementia, 1 Lewy body, 1 progressive supranuclear palsy, 1 unknown) and HARMONISATION, 2 years follow-up, 23 (18 %) dementia cases, (3 VD, 20 AD).

In all three cohorts, higher MD median was associated with a higher risk of dementia after controlling for the clinical markers (SCANS: HR (95% CI)=2.048 (1.438 to 2.918), p=7.1e-05, AUC=0.794; RUN DMC: HR (95% CI)=1.364 (1.060 to 1.755), p=0.016, AUC=0.825; HARMONISATION: HR (95% CI)=1.784 (1.085 to 2.935), p=0.023, AUC=0.757) (table 5). Associations for the other DTI measures are shown in online supplemental table 3.

Table 5

Higher median MD is related to higher risk of dementia while accounting for clinical markers

The baseline Clinical-Global model predicted dementia to a roughly similar degree to the Clinical-DTI model in all cohorts (online supplemental table 4). There were no significant differences in AUC between the Clinical-Global and Clinical-DTI models on formal statistical testing (online supplemental figure 1).

Longitudinal association between change in imaging marker and dementia conversion

Complete longitudinal data with repeat conventional MRI and DTI were available for SCANS (N=99), RUN DMC (N=257) and HARMONISATION (N=120) and for incident dementia cases: SCANS 18; RUN DMC 12; HARMONISATION 21. There was a change in DTI in all cohorts (online supplemental table 5).

Change in MD median over 3 years predicted dementia conversion over 5 years in SCANS (HR (95% CI)=2.588 (1.663 to 4.027), p=2.49e-05, AUC=0.785) but not in RUN DMC (OR (95% CI)=0.935 (0.498 to 1.667), p=0.825, AUC=0.924) or in HARMONISATION (OR (95% CI)=1.573 (0.998 to 2.597), p=0.056, AUC=0.738) (table 5). Although adding MD median increased the AUC from 0.684 to 0.738 in HARMONISATION, this increase was not significant. There was no increase in prediction in RUN-DMC (figure 3). Associations between other DTI markers and dementia conversion are in online supplemental table 6. In SCANS, adding the conventional MRI measures added prediction over the clinical-DTI model with the AUC increasing from 0.785 to 0.872 (p=0.054) (figure 3).

Figure 3

Change in multimodal imaging resulted in the largest area under the curve (AUC) only in severe small vessel disease. Values show the AUC for the Cox regression and logistic regression models: Clinical: Clinical markers alone, Clinical-DTI: Change in DTI combined with clinical markers, Clinical-MRI: Change in MRI markers combined with clinical markers, Clinical-MRI-DTI: All markers included. AUC evaluates how well the model classifies dementia conversion and no-dementia conversion at all possible cutoffs respectively. Differences in AUC between the models were tested. Clinical=model with clinical markers, Clinical-DTI=model with clinical markers and DTI marker, Clinical-MRI=model with clinical markers and conventional MRI markers, Clinical-MRI-DTI=model with clinical markers, conventional markers and DTI marker. DTI, diffusion tensor imaging; ROC, receiver operating cruve.


This analysis, across a number of cohorts at risk of dementia but with differing degrees of SVD, confirms the central importance of diffuse WM damage, measured by DTI, as a major determinant of cognitive impairment and dementia. It further validates the use of DTI measures as a marker to predict future dementia risk. The DTI measure also correlated with cognition in all cohorts demonstrating the contribution of WM damage to cognitive impairment across all grades of SVD disease severity. The relative contribution of the conventional markers varied according to disease severity. In all SVD cohorts (SCANS, RUN DMC and CADASIL), BV and the DTI measures were strong imaging predictors of cognition. In contrast, in the normal elderly (ASPS-Fam) and the MCI cohort (HARMONISATION), the DTI measure was the strongest imaging predictor, and there was little predictive value from brain atrophy and lacunes. This may partly reflect the fact that lacunes are more common in severe SVD. The DTI measure added a similar level of prediction to the simple clinical model as did inclusion of the global cognitive score.

Our results are consistent with WM damage leading to disruption of WM pathways and disconnection of brain networks underlying cognition, playing a crucial role in cognitive impairment and dementia. Data directly supporting this hypothesis comes from recent studies of brain networks. Tractography techniques can be used to generate brain networks from DTI data, and network measures have been shown to both correlate with cognition cross-sectionally and predict future dementia risk, in patients with SVD.3 32 33 Mediation analysis demonstrated that conventional markers of SVD such as WM damage, lacunes and CMB impair cognition through disruption of network integrity.3 Our data emphasise the importance of this process across various patient groups, not only those with prominent SVD, but also those with MCI in whom the predominant pathology is likely to be AD. Recent studies have emphasised the importance of WM alterations contributing to cognitive impairment in AD-like dementia.34 35

DTI measures have been proposed not only as a predictor of future dementia risk, but also a surrogate marker which could be used in clinical trials to evaluate new therapies in phase 2 studies.8 36 Cognition has been shown to be insensitive to change in longitudinal studies in SVD, and therefore, require very large sample sizes to detect change.12 In the severe SVD cohort SCANS but not in RUN DMC or HARMONISATION significant change in DTI measures predicted future dementia risk. The AUC significantly increased by adding the MRI markers. Taking all this evidence together suggests that trials using DTI measures together with conventional MRI markers as surrogate markers are likely to be more successful in more severe cohorts of SVD with the vascular factor being the primary determinant for dementia conversion.

We used a histogram analysis measuring alterations in the whole WM which has been previously used to estimate DTI parameters.7 9 36–38 We chose MD median as our primary measure as this showed consistent associations across the different cohorts, and since previous evidence has shown a high immediate reproducibility.37 It might be expected that FA would provide a more direct measure of WM track damage, but associations with cognition were less consistent with FA parameters, being strong in some cohorts, but weaker than MD in others. The reasons for this are unknown, but it suggests MD may be a more robust marker for use in clinical trials. However, all WM histogram analyses are time-consuming and require multiple MRI sequences. New methods of analysing DTI have been developed which would be quicker and potentially more robust for application in large datasets, including the fully automatic peak width of skeletonised MD,8 relying on tract-based spatial statistics and the semiautomatic DTI segmentation technique-based discrete segments describing brain microstructure of healthy and/or damaged tissue .39 These make the implementation of DTI measures in clinical and research prediction simpler. It is, however, important to determine how these operate across different SVD and MCI populations, between single versus multicentre studies and across different scanner types.

A major strength of this study is the validation of MRI markers across independent populations. The study also has limitations. Although these cohort studies were designed to analyse the association between imaging and clinical outcome measures, the underlying study design and image analysis were not the same across data. First, the different cohorts had different lengths of follow-up, and number of follow-up MR scans. In RUN DMC, we could only include time points 2011 and 2015 in the longitudinal DTI-change analysis due to scanner upgrades between 2006 and 2011. Second, different MRI scanners and different field strengths were employed (1.5 vs 3 Tesla). Third, different MRI analyses were used in quantifying the imaging measures. Fourth, the computation of cognitive scores and criteria for dementia diagnosis were different across cohorts which, however, may be strength as it shows the generalisability of the associations across a variety of methods.

In conclusion, our study across multiple cohorts with differing severity of SVD demonstrates the central role for diffuse WM microstructure alterations in cognitive impairment and future dementia. Our results demonstrate that DTI measures allow prediction of future dementia risk, and combining it with conventional markers provides additional predictive values in cohorts with severe SVD.

Data availability statement

Data are available on reasonable request.

Ethics statements

Patient consent for publication

Ethics approval

All studies were approved by the ethics committees of the respective institutions. SCANS was approved by the London–Wandsworth ethics committee (; study iD: 4577). RUN DMC was approved by the Medical Review Ethics Committee region Arnhem-Nijmegen (No. 2005/256). HARMONISATION was approved by the National Healthcare Group Domain-Specific Review Board (DSRB reference No. 2010/00017). PRESERVE was approved by the Harrow National Research Ethics Service committee (No: 11/LO/0458). ASPS-Fam was approved by the ethics committee of the Medical University of Graz, Austria (comission registration No. IRB00002556). CADASIL was approved by the LMU medical faculty (No. 299/03).


Supplementary materials

  • Supplementary Data

    This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.


  • Twitter @Camstroke

  • Contributors ME was responsible for the data analysis, interpretation of data, drafting the manuscript and critical revision of the manuscript for important intellectual content. He has full access to all of the data in the study. SH was responsible for the data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. AMT was responsible for the data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. LP was responsible for the data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. EH was responsible for the data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. MD was responsible for the study concept and design, data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. JW was responsible for the data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. RGM was responsible for the data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. MD was responsible for the study concept and design, data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. RS was responsible for the study concept and design, data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. DT was responsible for the data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. CPLHC was responsible for the study concept and design, data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. F-EdL was responsible for the study concept and design, the data acquisition and data analysis, interpretation of data and critical revision of the manuscript for important intellectual content. HSM was responsible for the study concept and design, the data acquisition and analysis, the interpretation of the data, drafting the manuscript, critical revision of the manuscript for important intellectual content and has full access to all of the data in the study.

  • Funding This work was funded by a grant from Alzheimer’s Research UK (ARUK-PG2016A-1). Additional support was provided by a Cambridge University- LMU collabative grant. ME is funded by a Priority Programme Grant from the Stroke Association (PPA 2015/02). Infrastructural support for this work was provided by National Institute of Health Research (NIHR) Cambridge Biomedical Research Centre Dementia and Neurodegeneration Theme (146281). HSM is supported by an NIHR Senior Investigator Award. CPLHC is supported by an National Medical Research Council (NMRC) of Singapore Senior Clinician-Scientist Award and the HARMONISATION study has been funded by NMRC grants. F-EdL is supported by a clinical established investigator grant from the Dutch Heart Foundation (2014 T060) and by a VIDI innovational grant from The Netherlands ZonMw (grant number 016126351). AMT has received a grant from the Junior Staff Member Dutch Heart Foundation (2016T044).

  • Disclaimer The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care. The funders and sponsors played no role in study design or analysis.

  • Competing interests None declared.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Linked Articles