Multi-cohort and longitudinal Bayesian clustering study of stage and subtype in Alzheimer’s disease

Understanding Alzheimer’s disease (AD) heterogeneity is important for understanding the underlying pathophysiological mechanisms of AD. However, AD atrophy subtypes may reflect different disease stages or biologically distinct subtypes. Here we use longitudinal magnetic resonance imaging data (891 participants with AD dementia, 305 healthy control participants) from four international cohorts, and longitudinal clustering to estimate differential atrophy trajectories from the age of clinical disease onset. Our findings (in amyloid-β positive AD patients) show five distinct longitudinal patterns of atrophy with different demographical and cognitive characteristics. Some previously reported atrophy subtypes may reflect disease stages rather than distinct subtypes. The heterogeneity in atrophy rates and cognitive decline within the five longitudinal atrophy patterns, potentially expresses a complex combination of protective/risk factors and concomitant non-AD pathologies. By alternating between the cross-sectional and longitudinal understanding of AD subtypes these analyses may allow better understanding of disease heterogeneity.


ADNI
The ADNI was launched in 2003 as a public-private partnership, led by a principal investigator Michael W. Weiner, MD.The primary goal of ADNI has been to test whether serial MRI, positron-emission tomography, other biological markers, and clinical and neuropsychological assessments can be combined to measure the progression of MCI and early AD.Similar to the AddNeuroMed study, ADNI strives to reveal sensitive and specific markers of AD progression in patients from various sites to support the development of new treatments and monitor their effectiveness, as well as to reduce the expenditures of clinical trials.Informed consent was obtained from all subjects included in this study.
For AD patients, consent was obtained both from the patient and a relative.

AddNeuroMed
AddNeuroMed is an integrated project that is part of InnoMed (the Innovative Medicines Initiative) and funded by the European Union Sixth Framework program.The main objective of AddNeuroMed is to identify biomarkers or experimental models that can improve diagnosis, prediction, and monitoring of disease progression in AD.Regarding neuroimaging, AddNeuroMed uses MRI and magnetic resonance spectroscopy to extract valuable information for the identification of AD biomarkers (Westman et al., 2011).The MRI data for AddNeuroMed were collected from different centres across Europe: University of Perugia (Italy), King's College London (United Kingdom), Aristotle University of Thessaloniki (Greece), University of Kuopio (Finland), University of Lodz (Poland), and University of Toulouse (France).Informed consent was obtained from all subjects included in this study.For AD patients, consent was obtained both from the patient and a relative.The columns 2-5 show visit frequencies for each cohort (ADNI, JADNI, AIBL, AddNeuroMed) and dataset (Discovery, Validation).Columns 6-9 Show the time in years between pairs of visits (e.g. the mean time interval between the 1 st and 2 nd MRI visits of AD patients in the discovery cohort was 0.68 years).

Clustering solution
Cluster  S2.Model optimisation information Five models were optimized for each number of clusters between two and eight.This totalled thirty-five models.Initial values were altered for each different model simulation.The best model for each number of clusters is presented above.The 2-cluster solution has slightly lower percentage of means with highly autocorrelated Monte Carlo Markov chains followed by the 5-cluster solution (even those chains with higher autocorrelation compared to the ones with no autocorrelation have converged but not as optimally as the latter ones).Model deviance is slightly lower for the latter model.This shows that the 2 and 5-cluster solutions do not differ significantly in terms of model quality.**With the word likelihood of the model we refer to the observed data likelihood 1 .The order of clustering solution on the table was decided based on a formula that accounts cluster means with high autocorrelation (a), and model deviance (b), simultaneously.That is, a and b were rescaled between 0 and 1 in order to be equally weighted.Then Euclidean distance was calculated based on a and b values for each clustering solution.The solution with the smallest Euclidean value accounts for the most optimal clustering solution (lowest a and b) followed by the other solutions ordered in the same fashion.S4.Dementia duration at MRI acquisition for each AD patient cluster.

Limbic
The data in the table refer to MRI visits per cluster of patients in the different disease duration spans (<50 months, 50-99 months, and >99 months).S5.First and second order subject assignment into clusters for the discovery dataset.

Clustering solution
The clustering algorithm assigned probabilities of each subject to belong in each cluster.Information on 2 nd and 3 rd class assignment is presented for subjects that did not have high probability to belong in the 1 st cluster assignment (e.g.16 subjects that are predominantly clustered in the LPA cluster can also be clustered to the MA cluster with lower probability while only one subject of the HS cluster can be clustered to the MA cluster with a lower probability than for the HS cluster that is assigned to).S6.First and second order subject assignment into clusters for the validation dataset.

MA
The classification algorithm (post clustering) assigned probabilities of each subject to belong in each cluster.Information on 1 nd and 2 rd class assignment is presented for subjects that did not have high probability to belong in the 1 st cluster assignment.Only 5 out of the 571 AD patients did not receive a certain classification to any of the clusters (e.g.first row: one patient that IS predominantly clustered in the MA cluster can also be clustered to the LPA cluster with lower probability and one subject of the MA cluster can be clustered to the HS cluster with a lower probability.S8.Concordance between the clustering in the whole AD dataset analysis and the separate cohort analyses.This figure shows the 1 st , 2 nd , and 3 rd principal component analysis coordinates for the patient cluster probabilities matrix of the discovery dataset.Only four out of the five hundred seventy one principal components were above near zero values.In the figure, each dot represents one patient of the discovery dataset.Any coordinate within the pyramidal plot, represents different patient cluster probability compositions, e.g., closest to the lower right corner, we can observe patients that are certainly clustered at the minimal atrophy cluster of patients.Since only the first three principal components are visualized, the corner of certain classification for the hippocampal sparing cluster of patients is not visualized.The first three principal components are enough to summarize information that separates the hippocampal sparing cluster from the other clusters of patients.Only 5 out of 571 patients are not certainly classified to one of the five clusters (see Table S5).The minimal and limbic predominant atrophy patterns show the greatest similarity (many AD patient probabilities trace dots lie on the borders between the two clusters).For the 1 st , 2 nd , 3 rd , 6 th and 7 th years after the AD onset, the validation dataset had enough data to provide median atrophy images per cluster.These new observations were classified to each cluster, and median disease duration was calculated.Median atrophy maps (group median atrophy) for the new data of each cluster are presented in the left column of each year's category.Then atrophy fitted values at the median disease duration of each cluster were calculated through the clustering model (right column of each year's category).The colourscale of the cortical maps reflects AD atrophy levels compared to a multicohort dataset of  negative cognitively unimpaired (CU) individuals.Data are w-value transformed and therefore colours represent standard deviations bellow the CU group controlled for aging.Fitted values are fixed for intracranial volume and MRI scanner field strength.Yellow and red represent less and more atrophy respectively.S8.The quantitative analysis for the assessment of similarities between the ADNI and J-ADNI/AIBL longitudinal atrophy patterns, showed that ADNI atrophy pattern 1 is more similar to J-ADNI/AIBL pattern 3, ADNI atrophy pattern 2 is more similar to J-ADNI/AIBL pattern 2, ADNI atrophy pattern 3 is more similar to J-ADNI/AIBL pattern 2, ADNI atrophy pattern 4 is more similar to J-ADNI/AIBL pattern 4, and ADNI atrophy pattern 5 is more similar to J-ADNI/AIBL pattern 4. J-ADNI/AIBL pattern 1

Figure S8. Discovery set participants
In the J-ADNI dataset 117 individuals were excluded because only 69 of them were AD or CU.28 out of 69 participants were CU, and only 18 remained CU (With no imaging data) for all cognitive assessments while the rest progressed to mild cognitive impairment.41 out of 69 participants were diagnosed with AD and only 3 had repeated measurements imaging data that did not pass the manual FreeSurfer quality control assessment.
from volunteers.The AIBL team wishes to thank the following clinicians who referred patients with AD and/or MCI to the study: Professor David Ames, Associate Professor Brian Chambers, Professor Edmond Chiu, Dr Roger Clarnette, Associate Professor David Darby, Dr Mary Davison, Dr John Drago, Dr Peter Drysdale, Dr Jacqui Gilbert, Dr Kwang Lim, Professor Nicola Lautenschlager, Dr Dina LoGiudice, Dr Peter McCardle, Dr Steve McFarlane, Dr Alastair Mander, Dr John Merory, Professor Daniel O'Connor, Professor Christopher Rowe, Dr Ron Scholes, Dr Mathew Samuel, Dr Darshan Trivedi, and Associate Professor Michael Woodward.We thank all those who participated in the study for their commitment and dedication to helping advance research into the early detection and causation of AD.

Figure
Figure S1.2-cluster solution atrophy fitted value maps Atrophy fitted values after the AD onset.Each row represents one cluster of patients with the corresponding pattern of atrophy.The colourscale reflects to AD atrophy compared to a multicohort dataset of  negative cognitively unimpaired (CU) individuals.Data are w-value transformed and therefore colours represent standard deviations bellow the CU group controlled for aging.Fitted values are fixed for intracranial volume and MRI scanner field strength.Yellow and red represent less and more atrophy respectively.

Figure S2 .
Figure S2.Fitted values for cortical thickness and subcortical volumes for the different longitudinal patterns of atrophy from AD onset.Atrophy fitted values from

Figure S3 .
Figure S3.Cluster-specific intercept and slope covariance matrixThe data are presented as correlations instead of variance/covariance in heatmaps form.A, C, E, G, I show cluster intercepts while B, D, F, H, J represent slopes.On the right side of each heatmap we present the topological order the relationship between variables with edges between brain regions coloured according to correlation.The colour scale spans from blue to red with low and high correlations, respectively.

Figure S5 .
Figure S5.Atrophy fitted values after the AD onset for the trained clustering model versus the new validation dataset data.

Figure S6 .Figure S7 .
Figure S6.Atrophy fitted values after the AD onset for ADNI (left panel) and J-ADNI/AIBL (right panel).Each row represents one cluster of patients with the corresponding pattern of atrophy.The colour scale reflects to AD atrophy compared to a multicohort dataset of  negative cognitively unimpaired (CU) individuals.Data are w-value transformed and therefore colours represent standard deviations bellow the CU group controlled for aging.Fitted values are fixed for intracranial volume and MRI scanner field strength.Yellow and red represent less and more atrophy respectively.Imaging threshold is set to -1.6 standard deviations below the CU sample normative values.

Table S7 .
Demographical and cognition, characteristics of the separate cohort analyses.Mini mental state examination; CDR: Clinical dementia rating: CDR: CDR sum of boxes; statistical testing: for ADNI only clusters 1, 2, and 3 were assessed because clusters 4 and 5 have low counts to test hypotheses.For J-ADNI,/AIBL only differences between clusters 2 and 3 were assessed because clusters 1 and 4 have low counts to test hypotheses; a categorical variables comparisons with  2 (two sided) and Fisher's exact test (two sided); b group comparisons with Kruskal-Wallis rank sum test (nonparametric); The Hochberg method was employed for post hoc multiple comparisons corrections.Confidence level for statistical comparisons was set to a = 0.05. c