Article Text

Download PDFPDF

Research paper
The identification of cognitive subtypes in Alzheimer's disease dementia using latent class analysis
  1. Nienke M E Scheltens1,
  2. Francisca Galindo-Garre2,
  3. Yolande A L Pijnenburg1,
  4. Annelies E van der Vlies1,
  5. Lieke L Smits1,
  6. Teddy Koene3,
  7. Charlotte E Teunissen4,
  8. Frederik Barkhof5,
  9. Mike P Wattjes5,
  10. Philip Scheltens1,
  11. Wiesje M van der Flier1,2
  1. 1Department of Neurology, Alzheimer Center, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands
  2. 2Department of Epidemiology and Biostatistics, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands
  3. 3Department of Medical Psychology, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands
  4. 4Neurochemistry Laboratory and Biobank, Department of Clinical Chemistry, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands
  5. 5Department of Radiology and Nuclear Medicine, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands
  1. Correspondence to Nienke Scheltens, Department of Neurology, Alzheimer Center, Neuroscience Campus Amsterdam, VU University Medical Center, P.O. Box 7057, Amsterdam 1007 MB, the Netherlands; n.scheltens{at}vumc.nl

Abstract

Objective Alzheimer's disease (AD) is a heterogeneous disorder with complex underlying neuropathology that is still not completely understood. For better understanding of this heterogeneity, we aimed to identify cognitive subtypes using latent class analysis (LCA) in a large sample of patients with AD dementia. In addition, we explored the relationship between the identified cognitive subtypes, and their demographical and neurobiological characteristics.

Methods We performed LCA based on neuropsychological test results of 938 consecutive probable patients with AD dementia using Mini-Mental State Examination as the covariate. Subsequently, we performed multinomial logistic regression analysis with cluster membership as dependent variable and dichotomised demographics, APOE genotype, cerebrospinal fluid biomarkers and MRI characteristics as independent variables.

Results LCA revealed eight clusters characterised by distinct cognitive profile and disease severity. Memory-impaired clusters—mild-memory (MILD-MEM) and moderate-memory (MOD-MEM)—included 43% of patients. Memory-spared clusters mild-visuospatial-language (MILD-VILA), mild-executive (MILD-EXE) and moderate-visuospatial (MOD-VISP) —included 29% of patients. Memory-indifferent clusters mild-diffuse (MILD-DIFF), moderate-language (MOD-LAN) and severe-diffuse (SEV-DIFF) —included 28% of patients. Cognitive clusters were associated with distinct demographical and neurobiological characteristics. In particular, the memory-spared MOD-VISP cluster was associated with younger age, APOE e4 negative genotype and prominent atrophy of the posterior cortex.

Conclusions Using LCA, we identified eight distinct cognitive subtypes in a large sample of patients with AD dementia. Cognitive clusters were associated with distinct demographical and neurobiological characteristics.

  • ALZHEIMER'S DISEASE
  • COGNITION
  • COGNITIVE NEUROPSYCHOLOGY
  • DEMENTIA
  • STATISTICS
View Full Text

Statistics from Altmetric.com

Introduction

With an estimated worldwide prevalence of over 35 million patients, dementia is a major healthcare challenge.1 The most common cause of dementia is Alzheimer's disease (AD). There is still no cure for AD. In the search for treatment, people have so far predominantly searched for one single therapeutic agent that could target the entire AD spectrum (magic bullet).2 A key factor in the failure to find effective therapy may be the fact that AD is a highly heterogeneous disorder and a single effective therapy for the entire AD entity may never be found.2 ,3 Rather, specific therapeutic strategies may benefit subgroups of patients. Identification of meaningful subtypes in AD may be a necessary first step towards personalised medicine.

An approach to identify AD subtypes is to subdivide patient groups based on neurobiological characteristics, including age-at-onset, apolipoprotein E (APOE) genotype, cerebrospinal fluid (CSF) biomarkers or imaging and subsequently exploring relationships with cognitive profiles. From earlier studies it became clear that specific biological features seem to be related to variability in the profile of cognitive impairment.4–6

An alternative approach is to classify patients based on cognitive features. Generally, episodic memory loss is the most salient symptom in AD.7 Several clinical variants of AD with relative sparing of memory have been described, such as logopenic progressive aphasia, the dysexecutive subtype and posterior cortical atrophy.8–10 Cognitive heterogeneity, however, is not restricted to these extreme, atypical variants, but is also present in patients in the normal spectrum of AD. Several previous studies used a more data-driven and bottom-up approach to capture cognitive heterogeneity in neurodegenerative diseases by performing statistical clustering methods based on neuropsychological data, resulting in the identification of distinct cognitive subtypes.11–16 Generalisability, however, is hampered in these previous studies by either a small sample size of patients with AD dementia or a limited neuropsychological test protocol. We extended on these former findings by using a large sample of patients with AD dementia in combination with an extensive neuropsychological test battery as input for the statistical clustering. We chose latent class analysis (LCA) to identify cognitive subtypes since LCA yields a smaller misclassification rate than the other clustering methods because it is a probability-based approach.17 Following the clustering, we explored relationships between the identified cognitive subtypes and demographical and neurobiological features, that is, APOE genotype, CSF biomarkers and MRI characteristics.

Methods

Patients

We selected 1006 consecutive patients with probable AD dementia from the Amsterdam Dementia Cohort.18 Patients visited the outpatient memory clinic of the VU University Medical Center (VUmc) Alzheimer Center between February 2001 and May 2013. All patients were offered a standardised one-day assessment, including medical history, informant-based history, physical and neurological examination, blood tests, neuropsychological assessment and MRI of the brain. We recorded the subjectively reported duration of symptoms at first visit. Level of education was defined according to a rating scale ranging from 1 (low, primary school not finished) to 7 (high, university degree).19 Mini-Mental State Examination (MMSE) was obtained as a measurement of disease severity.20 Patients with MMSE ≤10 were excluded (ie, 18 patients because of unavailable MMSE results, 50 patients because of MMSE ≤10) to limit the influence of floor effects on cognitive tests, resulting in a study sample of 938 patients.

Probable AD dementia diagnosis was made in a multidisciplinary meeting according to the NINCDS-ADRDA criteria.21 In addition, all patients met the core clinical criteria of the National Institute on Aging-Alzheimer's Association (NIA-AA) workgroups.7 The local Medical Ethics Committee approved the study. All patients provided written informed consent to use their clinical data for research purposes.

Neuropsychological testing

The standard neuropsychological test battery was composed to assess the major cognitive functions. Memory was assessed with the Dutch version of the Rey Auditory Verbal Learning Test (RAVLT) and the Visual Association Test (VAT).22 ,23 From the RAVLT, we used the total number of words remembered on the five learning trials (RAVLT learning, score 0–75) and the delayed recall (score 0–15). From the VAT, we used the total score of the first two trials (range 0–12). Language was assessed with VAT naming (score 0–12), naming of 20 images from the Arizona Battery for Communication Disorders of Dementia (ABCD naming) and category fluency (animals).24 ,25 Executive functioning was assessed with letter fluency (D-A-T), digit span backward, Trail Making Test B (TMT-B), Frontal Assessment Battery (FAB) and comparative questions.26–29 For the digit span backward, an extended version was used with three trials per sequence length, starting with a two-digit sequence (score 0–21). For the TMT-B, the time required for completion was recorded. The six Dutch comparative questions contained comparative elements and could only be answered by ‘Yes’ or ‘No’. Attention was assessed with TMT-A and digit span forward.27 ,28 For TMT-A, the time required to complete was recorded. The digit span forward was assessed using the extended version, with three trials per sequence length starting with a two-digit sequence (0–21). Visuospatial functioning was assessed with three tests from the Visual Object and Space Perception Battery, that is, dot counting (score 0–10), number location (score 0–10) and incomplete letters (score 0–20).30 Finally, we used an inhouse developed money counting test. Patients are offered three combinations of coins and are asked to report the total amount of money. Correct answers (money counting score) and the time to report these amounts (money counting time) are recorded. The composition of the neuropsychological test battery gradually evolved over the past 12 years. As a consequence, not every test was assessed in every patient (ranging from 891 (VAT) to 321 (money counting)).

APOE genotype

DNA was isolated from 10 mL blood samples and collected in EDTA tubes. APOE genotype (available for 86% of the sample) was determined at the Department of Clinical Chemistry of the VUmc with Light Cycler APOE mutation detection method (Roche Diagnostics GmbH, Mannheim, Germany). APOE genotype was dichotomised according to the presence or absence of one or more APOE e4 alleles.

CSF biomarkers

CSF was collected in 10 mL polypropylene tubes (Sarstedt, Nümbrecht, Germany). Part of the sample was used for measuring the number of leucocytes, total protein and glucose. Within 2 h after collection, the remainder was centrifuged at 1800g for 10 min at 4°C and transferred into a second polypropylene tube, and stored at −20°C. Within 2 months after lumbar puncture, analysis of amyloid-β 1–42 (Aβ1–42), total τ (τ) and τ phosphorylated at threonine-181 (pτ) was performed using sandwich ELISAs (Innotest β-Amyloid(1–42), Innotest hTAU-Ag and Innotest Phosphotau(181P); Innogenetics, Gent, Belgium). CSF was available for 662 patients (71% of the sample).

MRI characteristics

MRI of the brain was performed according to our dementia MRI protocol, including the following sequences: three-dimensional (3D) T1-weighted, axial T2-weighted, axial T2*-weighted and axial fluid-attenuated inversion recovery (FLAIR) images. Visual rating of MRIs was performed according to qualitative rating scales, where higher scores reflect more severe abnormalities. Medial temporal lobe atrophy (MTA, 0–4) was rated on the oblique coronal multiplanar reconstructions (MPR) of T1-weighted gradient-echo 3D sequences perpendicular to the long axis of the hippocampus.31 Atrophy of the posterior cortex (PA, 0–3) was rated on a combination of FLAIR, sagittal and coronal MPR T1-weighted images.32 Global cortical atrophy (GCA, 0–3) was rated on axial FLAIR images.33 White matter hyperintensities (WMH, 0–3) were rated on FLAIR images.34 In addition, we determined the presence of lacunes and microbleeds (0 vs ≥1). MRI scans were available for 752 patients (80% of the sample).

Statistical analyses

For the clustering, neuropsychological test results were transformed into z-scores to allow comparison of tests within patients. TMT-A, TMT-B and money counting time were log-transformed, since they were not normally distributed and inverted so that a lower score implied more impairment. We used LCA for the clustering of neuropsychological data, with Latent Gold V.4.5.35 LCA yields a smaller misclassification rate compared to other clustering methods and it allows for missing data.17 This last property is important because it allowed us to include all consecutive patients, resulting in a representative sample. In Latent Gold, latent class models are estimated with a full information maximum likelihood or maximum posterior mode algorithm that uses all the available information for each individual to compute parameter estimates. This method produces unbiased parameter estimates and SEs under missing at random and missing completely at random assumptions.36 We identified the optimal number of clusters using the Akaike Information Criterion (AIC), AIC with penalty factor of 3 (AIC3) and Bayesian Information Criterion (BIC).37 MMSE was used as covariate. We used SPSS for Mac V.21.0 (IBM Corp, Armonk, New York, USA) to explore relationships between cluster membership and demographics, APOE genotype, CSF biomarkers and MRI characteristics. Differences between clusters were examined using χ2 tests for categorical variables, analysis of variance (ANOVA) with post hoc Scheffe's analyses for normally distributed variables and Kruskal-Wallis analyses for not normally distributed variables. When group differences were observed with χ2 tests or Kruskal-Wallis analyses, we used ANOVA with Tamhane's T2 post hoc analyses, in which equal variances are not assumed.

Subsequently, we performed a multinomial logistic regression analysis with cluster membership as dependent variable (the most typical, memory-impaired subtype as reference cluster) and dichotomised demographics, APOE genotype, CSF biomarkers and MRI characteristics as independent variables. Cut-off values are given in table 5. First, we used unadjusted models. Second, we adjusted for dichotomised covariates age, sex, duration of symptoms and education. Although the present study has a bottom-up (and thus, hypothesis-free) design, we corrected for multiple testing using the Benjamini-Hochberg correction.38 Data are presented as ORs with 95% CI. The significance level was set at p<0.05.

Results

Cohort characteristics are summarised in table 1.

Table 1

Cohort characteristics

Latent class analysis

According to AIC, AIC3 and BIC, a model with eight clusters provided the best fit (presented in table 2). The classification error in the final model was 9%, implying that if patients are assigned to the cluster having the highest membership probability, 9% will be misclassified.

Table 2

Parameters of fit of latent class analysis

Neuropsychological cluster characteristics

Raw neuropsychological data by cluster are given in table 3. Neuropsychological profiles are illustrated in figure 1. Two clusters were characterised by most prominent memory-impairment and together these included 43% of the cohort. Three clusters were characterised by a relatively memory-spared profile and together these included 29% of the cohort. Three clusters did not reveal a distinct memory impaired or memory spared profile; therefore, we called them memory-indifferent.

Table 3

Raw neuropsychological test results by clusters

Figure 1

Mean neuropsychological z-scores by cluster. The x axis show the neuropsychological tests, the y axis the mean z-scores for each cluster. Scores for TMT-A, TMT-B and money counting time were inverted so that on all tests higher z-scores indicate better performance. Abbreviations clusters: MILD-MEM, mild-memory; MOD-MEM, moderate-memory; MILD-VILA, mild-visuospatial-language; MILD-EXE,mild-executive; MOD-VISP, moderate-visuospatial; MILD-DIFF, mild-diffuse; MOD-LAN, moderate-language; SEV-DIFF, severe-diffuse.

Two clusters were characterised by most prominent memory-impairment. The first memory-impaired cluster had a mean MMSE of 24 and we called this cluster mild-memory (MILD-MEM). The other memory-impaired cluster had a mean MMSE of 19; therefore, we called it moderate-memory (MOD-MEM). MOD-MEM had also relatively low scores on tests assessing executive functioning. The first memory-spared cluster showed both low scores on the language tests, especially VAT naming and the visuospatial tests. We called this cluster mild-visuospatial-language (MILD-VILA). The second memory-spared cluster was characterised by most prominent impairment in executive functioning and had a mean MMSE of 23. We called this cluster mild-executive (MILD-EXE). The last memory-spared cluster was characterised by remarkably low scores on the visuospatial tests, and the TMT-A and TMT-B; tests with a prominent visuospatial element. This cluster had a mean MMSE of 19 and we called it moderate-visuospatial (MOD-VISP). The first memory-indifferent cluster had global cognitive impairment and a mean MMSE of 21; we called this cluster mild-diffuse (MILD-DIFF). The second memory-indifferent cluster had most prominent language impairment and a mean MMSE of 20. We called this cluster moderate-language (MOD-LAN). The third memory-indifferent cluster had also a diffuse profile and had a mean MMSE of 14; we called this cluster severe-diffuse (SEV-DIFF).

Demographical and neurobiological cluster characteristics

Table 4 shows raw data of demographics, APOE genotype, CSF biomarkers and MRI characteristics by cluster. We observed group differences for age, duration of symptoms, education, MTA, GCA and presence of lacunes. Post hoc differences are given in table 4.

Table 4

Demographics, APOE genotype, cerebrospinal fluid (CSF) biomarkers and MRI characteristics by cluster

Table 5 shows OR's and 95% CI's estimated by unadjusted multinomial logistic regression analyses with cluster membership as dependent variable (with MILD-MEM as reference) and dichotomised demographics, APOE genotype, CSF biomarkers and MRI characteristics as independent variables.

Table 5

Multinomial logistic regression analysis with cluster membership as dependent variable

Compared to MILD-MEM, younger patients were six times more likely to be classified in MOD-VISP, three times likely to be classified in SEV-DIFF and two times likely to be classified in MILD-EXE. Females were twice as likely to be classified in MILD-DIFF. Patients with longer duration of symptoms were twice as likely to be a member of SEV-DIFF. Less educated patients were three times more likely to be classified in MOD-MEM and SEV-DIFF, and twice as likely to be classified in MILD-EXE and MILD-DIFF. APOE e4 negative patients were two times more likely to be members of MOD-VISP. Patients with higher τ concentrations were twice as likely to be classified in MOD-LAN. Patients with higher pτ concentrations were half as likely to be member of MILD-EXE and MILD-DIFF. Patients with prominent MTA were half as likely to be classified in MILD-EXE. Prominent PA and GCA were predisposed for membership of MOD-VISP and SEV-DIFF. Finally, patients with more WMH were three times more likely to be members of MOD-LAN.

When we adjusted for dichotomised demographics age, sex, duration of symptoms and education, results remained largely comparable (differences indicated as footnote ‡ in table 5).

Finally, we adjusted for multiple testing using the Benjamini-Hochberg correction. Differences that were not significant anymore after correction are indicated as footnote † in table 5.

Discussion

In this study, we identified eight cognitive AD clusters and demonstrated their relationships with demographical and neurobiological characteristics.

Previous studies have demonstrated the eligibility of bottom-up and data-driven approaches, such as LCA, to identify cognitive subtypes in neurodegenerative diseases.11–16 One study identified 13 cognitive subtypes using hierarchical clustering of neuropsychological test results from a small sample of patients with AD dementia.11 Another study performed LCA based on subscales of two cognitive screening tests in a large sample of patients with AD dementia and identified four cognitive subtypes.12 Another study identified four cognitive subtypes in a small sample of patients with AD dementia and/or vascular dementia.15 Statistical clustering methods have also been used to identify cognitive subtypes in patients with mild cognitive impairment.13 ,14 One recently published study performed principle component analysis based on neuropsychological test results of a small sample of patients with AD dementia, followed by LCA.16 This study identified four cognitive clusters and concluded that cluster definitions remained stable when the clustering was repeated over time.

We extended on these former findings by performing LCA in a large and unselected sample of patients with AD dementia for whom we had extensive neuropsychological data available. We were able to relate the identified cognitive subtypes to a number of well-known markers for AD pathology. We found two memory-impaired clusters (43%), three memory-indifferent clusters (28%) and three memory-spared clusters (29%). Clusters were characterised by distinct demographical and neurobiological characteristics.

The memory-impaired cluster MILD-MEM was associated with higher age and MOD-MEM with more severe MTA. Furthermore, patients with APOE e4 positive genotype were more often associated with membership of MILD-MEM than MOD-VISP. These memory-impaired clusters probably represent the most typical and common subtypes of AD.5 ,7

Roughly one out of three patients was classified in one of the three memory-spared clusters MILD-VILA, MILD-EXE and MOD-VISP. Younger age was associated with membership of MILD-EXE and more prominently, with membership of MOD-VISP. Furthermore, MOD-VISP was associated with APOE e4 negative genotype and more prominently with PA and GCA. This is in accordance with earlier studies, where absence of prominent memory-impairment was associated with a younger age-at-onset and the absence of APOE e4, less MTA, and more PA and GCA.5 ,32 ,39 Early visuospatial problems have been described as the most frequent atypical presentation of AD, often encountered in younger patients who lack the APOE e4 allele.5 ,32 MILD-VILA was characterised by a relatively spared memory, and impairment in both the language and visuospatial domains. This cognitive profile did not represent a cognitive AD subtype described before. Possibly, (1) performance on language tests was influenced by visuospatial impairment, especially as the naming tests were impaired—possibly driven by impairment in perception and recognition. Another possibility is that (2) the identification of MILD-VILA was mainly driven by the remarkably good memory function, resulting in a cluster with patients having a memory-spared profile with either language or visuospatial impairment. Or (3) this cluster—with its relatively high mean MMSE of 25—represents patients with logopenic progressive aphasia and posterior cortical atrophy in an early stage together. Our identified memory-spared dysexecutive cluster MILD-EXE coincides with earlier findings of prominently impaired executive functioning in younger and APOE e4 negative patients.9 In our sample, MOD-VISP was characterised by lowest scores on visuospatial tasks, and on TMT-A and TMT-B—the tests for attention and executive functioning albeit with a prominent visuospatial component. Finally, we identified three memory-indifferent clusters MILD-DIFF, MOD-LAN and SEV-DIFF, which together included 28% of patients. MILD-DIFF was characterised by mild impairment in all cognitive domains. MOD-LAN had a memory-indifferent profile, but apparent impairment in memory tasks was possibly influenced by language impairment since verbal memory (RAVLT learning >delayed recall) was most prominently impaired. MOD-LAN was linked to more WMH, suggesting vascular involvement of this cluster and higher τ concentrations. The association between higher τ concentrations and the logopenic primary progressive aphasia variant of AD has been described before.40 The most severely impaired cluster SEV-DIFF was characterised by younger age, longer duration of symptoms, less education and more atrophy. The combination of more severe cognitive impairment, severe atrophy and longer duration of symptoms suggests that SEV-DIFF could represent a further stage of the disease. This would suggest that patients were relatively old and would have suffered the disease for a longer time. The relatively young average age (67±9 years) of this cluster, however, suggests that alternatively these patients have a more aggressive subtype.

The present study has been designed to identify cognitive subtypes, and explore their demographical and neurobiological characteristics. The exploration of cluster characteristics was, therefore, hypothesis-free or bottom-up. Partly due to relatively small cluster sizes, and the many demographical and neurobiological variables introduced in the analyses, several results did not remain significant after correction for multiple testing (indicated as † in table 5). It is important to validate our findings by repeating this approach in other (large) samples of patients with AD dementia .

Among the limitations of our study may be the fact that not all neuropsychological tests were assessed in every patient, largely due to slight changes in our test battery over the years. However, we feel that it may actually be a strength that we included all patients, independent of test completeness and in this way promoted generalisability for the entire AD population. Since our centre is a tertiary referral centre with specific expertise in early onset dementia, our cohort includes a relative large proportion of early onset patients and patients with less typical phenotypes visiting for a second or third opinion, which may have led to an overestimation of the relative prevalence of the non-memory subtypes. Nonetheless, this has also an advantage, as the resulting cohort had sufficient power to identify these non-memory phenotypes. A potential limitation of our study is the fact that heterogeneity in disease severity could have driven the clustering. For this reason, we excluded patients with MMSE ≤10 and we used MMSE as covariate in the clustering. Furthermore, it is a possibility that part of the patients could be classified in another cluster when clustering was performed in another disease stage. However, an earlier study demonstrated that cognitive AD clusters—identified using LCA based on neuropsychological test results—remained stable over time.16 An important strength of our study is the uniqueness of the large and unselected sample of probable patients with AD in combination with the extensive neuropsychological test battery we used for LCA. All patients were carefully diagnosed in a multidisciplinary team after diagnostic investigation according to our standard protocol and according to the NINCDS-ADRDA criteria.18 ,21 Moreover, all patients met the core clinical criteria of the NIA-AA.

Our results have important implications. The fact that a cure for AD has not been found may be due to the complexity of the disease. It is conceivable that dementia due to AD represents a wide spectrum of patients experiencing a similar end stage of a disease, that may be reached via (partially) different pathways. If this is true, then one magic bullet will never be found, but different therapeutic agents may benefit different subgroups of patients. To achieve such a future of personalised medicine, it is necessary to make more differentiated diagnoses than is now customary. The identification of cognitive AD subtypes can be a first step towards such a differentiated diagnosis. Furthermore, the distinct neurobiological characteristics of each subtype that we found provides support for the notion that partially different pathological pathways underlie these different subtypes and this may help guide the search for new therapy.

Acknowledgments

Research of the VUmc Alzheimer centre is part of the neurodegeneration research programme of the Neuroscience Campus Amsterdam. The VUmc Alzheimer Center is supported by Alzheimer Nederland and Stichting VUmc Fonds. The clinical database structure was developed with funding from Stichting Dioraphte.

References

View Abstract

Footnotes

  • Twitter Follow Nienke Scheltens at @NienkeScheltens

  • Contributors All authors have made substantial contributions to the conception and design of the work, the acquisition, analysis and/or interpretation of data for the work, drafting of the article or revising it critically for important intellectual content and gave final approval of the version to be published.

  • Competing interests CET is a member of the Innogenetics International Advisory Boards of Fujirebio/Innogenetics and Roche. FB serves/has served on the advisory boards of Bayer-Schering Pharma, Sanofi-Aventis, Biogen-Idec, TEVA, Merck-Serono, Novartis, Roche, Synthon BV, Jansen Research and Genzyme. He received funding from the Dutch MS Society and EU-FP7 and has been a speaker at symposia organised by the Serono Symposia Foundation and MedScape. MPW serves as a consultant for Biogen Idec and Roche. PS has served as consultant for Wyeth-Elan, Genentech, Danone and Novartis and received funding for travel from Pfizer, Elan, Janssen and Danone Research.

  • Patient consent Obtained.

  • Ethics approval Medical ethical committee of the VU University Medical Centre.

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

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.