Nat commun: Multicohort-longitudinal clustering of Alzheimer's disease stages and subtypes

Understanding the heterogeneity of Alzheimer's disease (AD) is important
for understanding the underlying pathophysiology of AD.
However, the AD cerebral atrophy subtype may reflect the stage of the disease or may be a subtype
of the disease.
A recent article in Nature Communications titled "Multi-cohort and longitudinal Bayesian clustering study of stage and subtype in Alzheimer's disease" was published, in which the authors used longitudinal magnetic resonance imaging data (891 AD patients; 305 healthy controls) and longitudinal cluster analysis methods evaluated the trajectory
of AD brain atrophy over time.
The authors identified five longitudinal patterns of brain atrophy, dividing previously reported AD subtypes into two atrophy pathways (midtemporal lobe and cortex).

In this study, AD subtypes were transitioned from a cross-sectional understanding to a longitudinal clustering perspective
.
The model proposed by the authors helps solve a long-standing problem
that distinguishes AD disease stages from actual disease subtypes.

Research background

Brain atrophy in Alzheimer's disease (AD) is associated
with cognitive decline and topological spread of neurofibrillary tangles (NFTs).
Neuropathological and in vivo neuroimaging studies challenge the hypothesis of AD as a single entity, supporting the hypothesis
of AD as a heterogeneous disease.
The latest research suggests that the heterogeneity of AD can be explained by two main dimensions: severity and typicality, which come
in the form of various biomarkers and clinical manifestations.
Based on regional atrophy and/or NFT diffusion, four AD subtypes have been reported in the literature: typical subtype, hippocampal preservation subtype, marginal dominant subtype, and microatrophy subtype
.

What is urgently needed to know is whether the observed heterogeneity reflects different disease stages or different subtypes, and whether these subtypes eventually converge
in the later stages of the disease.
Advances in biomarker research, data collection, and computational methods have significantly improved the ability to
study heterogeneity across different diseases.
These computational methods combine various in vivo pathophysiological markers to mimic disease heterogeneity
.
Meaningful grouping of AD patients based on neuropathology, neuroimaging, clinical, and biomarkers reveals heterogeneity
in the clinical diagnosis of AD.

However, the current study is based on cross-sectional analysis, increasing the likelihood that the
patterns identified reflect different disease stages rather than disease subtypes.
Recent studies have modeled the trajectories of subtype biomarkers in vivo from cross-sectional imaging datasets to infer disease stage
.
This is the first step
in assessing and interpreting the stage of disease.
However, it cannot be ruled out that the identified patterns may still reflect different stages of disease, as longitudinal information is not used for clustering, but only for post-hoc subtypes
.
This hypothesis is partially confirmed in models with various biomarker types (increased disease specificity), but it remains unrealistic when a clear timescale of events for each patient is not in place
.

Recent reviews present current methods for identifying heterogeneous disease subtypes and summarize existing AD subtypes in the literature, pointing out important data and methodological limitations that need to be overcome to better understand heterogeneity
in AD.
According to their conclusions, the field lacks longitudinal AD subtypes based on well-defined time scales (i.
e.
, age at measurement, age at onset of disease) in order to distinguish disease stages from disease subtypes
.

The aim of this study was to assess whether the heterogeneity of AD brain atrophy patterns stems from observations of different stages of the disease or reflects different subtypes
with specific atrophy and cognitive trajectories.
The authors modeled longitudinal data using a longitudinal Bayesian clustering framework, assessing disease staging and heterogeneity for 8 years from the onset of clinical disease (well-defined time scales) (previous studies used only cross-sectional data).

。 Multicenter cohort of structural magnetic resonance imaging (MRI) data from four continents (Alzheimer's Disease Neuroimaging Initiative, Japanese ADNI, AddNeuroMed, and the Australian Imaging, Biomarkers and Lifestyle study) were used to explore the different trajectories of AD atrophy
。 Only amyloid-positive AD patients were included to increase diagnostic specificity
.

In addition, the assessment of whether atrophic subtypes converge during the course of the disease is an important step in
understanding AD heterogeneity.
The discovered atrophy patterns were validated in an external dataset to assess the ability
of the model proposed by the authors to classify new patients.
Finally, differences between and within subtypes of cognitive decline and related disease modifiers such as APOE genotype, education, and premorbid intelligence were
assessed.

outcome

The included sample included 1196 participants (891 AD dementia patients and 305 cognitively normal patients) from 4 cohorts, and the included and validation datasets consisted of
320 and 571 AD dementia patients, respectively.
The demographics of the cohort are shown in Table 1
.

Table 1 Demographics of participants included and validated for cohort
.
Note: 1n (%); 2median (median absolute distance); 3 mean (standard deviation); The 4% denominator refers to the sum of APOE records that are not lost; The length of education is divided into 4 levels (1 = 0-8 years; 2 = 9-13 years; 3 = 13-15 years; 4>15 years).

The median MMSE score at baseline MRI visit in the AD group was 23 (first quartile: 21, third quartile: 25).

Longitudinal gray matter patterns in the cognitively impaired (CU) group and AD group showed that gray matter deteriorated with age in the CU group (Figure 1A
).
As expected, the AD group showed broader shrinkage (Figure 1B).

Applying the correction method of the AD dataset (normalization of gray matter in each AD patient relative to the CU model in Figure 1A) shows that AD shows different patterns
of atrophy at the population level depending on the age of the patient.
Patients younger than 65 years of age typically have more posterior cortical atrophy, while patients older than 75 years present with a typical pattern of temporal lobe atrophy in AD (Figure 1C).

Figure 1 Brain atrophy patterns
in CU and AD groups.
To calculate CU and AD atrophy patterns for different ages (A, B), z-value conversion
was performed on the data.
A mixed-effects multivariate model was used to visualize two diagnostic labels (red, more atrophy; yellow, less atrophy).

The color legend at the top right refers to the standard deviation of the sample mean (0 corresponds to the mean of the AD and CU sample values).

At age 55, AD appears to have similar levels of atrophy to CU populations and shows differences
with age.
To visualize the correction of AD data based on CU sample (C), two separate mixed-effect multivariate models (one for CU samples and one for AD samples)
were used.
AD data is normalized
based on CU data.
Thus, a lower color legend shows the standard deviation of the AD population below the CU population (w values, 0 corresponds to the mean of the CU sample values).

Younger AD patients (between 55 and 65 years of age) showed more posterior cortical atrophy compared to controls, while older AD patients (over 75 years of age) showed more pronounced midtemporal lobe and hippocampal atrophy
.

Cluster evaluation

Longitudinal cluster analysis showed that the 2-cluster and 5-cluster models were optimal with minimal
differences.
2 - Cluster models are more suitable for one clustering criterion (fewer random-effects parameters in the MCMC sample, higher autocorrelation), while 5 - Cluster models are more favorable for another clustering criterion (lower model bias).

Other clustering solutions have poor combinations of quality scores (many autocorrelated MCMC samples or models have high bias).

The 2-clustering solution separates the set of findings only in cortical severity (high vs.
low cerebral atrophy), while the 5-clustering solution (Figure 2, fitted values) reveals spatially distinct atrophy subtypes
.
Since the different spatial atrophy subtypes are more important from an exploratory perspective, and taking into account the previous literature on AD subtypes, here the authors interpret the results of the 5-clustering scheme
.

Figure 2 Fitted values of cortical thickness and subcortical volume for different longitudinal atrophy patterns since the onset of AD
.
Atrophy fit at the onset of clinical AD
.
Each row represents a group of patients
with a corresponding pattern of atrophy.
Color scale showing cortical thinning and subcortical volume loss with Aβ-negative, cognitively impaired (CU) individuals (red, more atrophy; yellow, less atrophy) comparatively
.
The data are converted by w values, so the colors indicate the standard deviation
of the CU group below the aging control.
The fitted values for intracranial volume and MRI scanner field strength are fixed
.

Cluster shrinkage patterns and discriminant characteristics

In these datasets, the authors found that five groups of patients showed gradual or dramatic longitudinal atrophy progression (Figure 2).

Compared to the CU group, the largest cluster, microatrophy (MA, 59.
1%), had very small midtemporal lobe atrophy at the onset of clinical AD (Figure 2, below 1.
6 standard deviations below the CU population).

The type progresses slowly, with enolinal and hippocampal involvement extending into other temporal lobe regions
.
The second largest cluster, marginal preponderance atrophy (LPA, 29.
1%), presents with entorhinal cortical atrophy at the beginning of the clinic, and later affects other temporal lobe regions
, including the hippocampus.
The third cluster, LPA+ (7.
2%), is spatially similar to the LPA cluster, but exhibits more atrophy
in the entorhinal cortex at the onset of AD.
Atrophy gradually extends to the temporal lobe and then further to the rest of the cortex
.

The authors also found a cluster, diffuse atrophy (DA, 1.
6%), which already had temporal and frontal lobe involvement at the onset of AD, and atrophy spread
rapidly over the course of the disease.
The last aggregate, hippocampus-sparing (HS, 3.
1%), had parietal lobe atrophy at onset without endotemporal structural involvement, but atrophy progressed rapidly
.
The MA and LPA patterns focus on generalized temporal lobe atrophy, while LPA+ concentrates on DA
7 years after disease onset.
Over time, the least typical atrophy pattern, HS, also progressed to a more diffuse atrophy pattern, but less often involving the hippocampus
.
Cluster
names are determined based on the atrophy pattern at the beginning of AD.
Table 2 provides a four-dimensional representation of each subtype to illustrate how atrophy and cognitive patterns evolve over time (Table 3, Figure 3).

Cluster intercepts (AD onset) show that HS and DA clusters are significantly thinner in the cortical parietal lobe than the other three clusters (figures 2 and 4).

LPA clusters have less entorhinal atrophy than LPA+
.
With regard to cluster slope (evolution of atrophy over time), posterior cingulate gyrus, operculum, orbit, and insula distinguish DA and HS from the other three clusters (Figures 2, 4).

HS clusters have the steepest contraction slope, followed by DA and LPA+ clusters
.

Table 2 Summary
of longitudinal atrophy and cognitive trajectories of four groups of AD patients.

Table 3 Cluster cluster characteristics
.
Note: 1n (%); 2median (median absolute distance); 3 mean (standard deviation); Estimate at the beginning of 4AD (estimated standard error); Estimate at the beginning of 5AD (estimated annual change); The number of years of education is divided into 4 levels (1 = < 0-8 years; 2 = 9-13 years; 3 = 13-15 years; 4 > 15 years old).

Adjustments
for multiple comparisons were evaluated using the Holms-Sidak method.
ANART National Adult Reading Test, MMSE Mini-Mental State Test, CDR Clinical Dementia Score, CDR SOB CDR Box Sum, GDS Geriatric Depression Scale
.
The baseline difference
between Cluster 1 and the rest of the population.
Vertical differences
between cluster 1 and other clusters.
cBaseline or longitudinal differences
between discovery datasets and validation datasets.

Figure 3 Cluster-specific cognitive trajectories
after clinical onset of dementia.
Trajectory estimation using mixed-effects models to account for intra-participant and cohort variability
.
MMSE Mini-Mental State Test, ADAS Alzheimer's Disease Assessment Scale
.
The dotted line represents the 95% confidence interval
.
Figure 4 Longitudinal clustering model clustering mean intercept and slope atrophy coefficient
.
Each row of the heatmap is grouped according to neuroanatomical spatial location (red, more atrophy; yellow, less atrophy).

Columns representing different clusters are grouped
based on similarity between clusters.
The vertical line within the cell represents the average ROI value of the cluster area (the vertical dashed line represents the value 0, which is not different from the CU sample).

Diffuse atrophy clusters have the lowest intercept and are not grouped
with any other clusters.
Diffuse atrophy and cluster slope of hippocampal retention clusters are
combined.
The marginal dominant subtype and the microatrophic subtype are grouped together
.

Compared to the 2-cluster protocol, five longitudinal atrophy patterns (Figure 2) reveal a fine grouping that includes changes
in the typical distribution of AD patients' atrophy stages.
In Table 3, the authors summarize the longitudinal patterns of atrophy to show the different features of the five longitudinal patterns and the patient characteristics
associated with them.
Following the main cluster analysis, the cluster-specific atrophy intercept and slope of the post-hoc hierarchical clustering (Figure 4, slope dendrogram and legend) quantitatively showed that MA, LPA, and LPA+ had similar spatial distributions of atrophy over time (however, different levels of atrophy at the onset of AD and different rates of atrophy progression) starting in the middle temporal lobe and spreading further into the neocortex
.
The HS pattern follows another spatial atrophy distribution, starting from the cortical region
.
DA clusters were quantitatively grouped along with the HS pattern, but expressed both progressive atrophy patterns at the same time, as the authors observed it in the advanced stages of the disease, where it
had been extensively atrophied.

Cluster characteristics

The percentage of patients in each of the five clusters was similar (table 3).

In the datasets found, MA had the highest APOE e4 carrier frequency (75%) and HS the lowest (40%)
.
Patients in the DA and HS clusters had a higher level of education (> 15 years of age), followed by the MA, LPA, and LPA+ groups (≤ 15 years of age).

Using MA (the largest cluster in the dataset) as a reference group, the authors found a significant reduction in US National Adult Reading Test (ANART) scores for LPA+ and HS (p < 0.
05).

For LPA, the Mini-Mental Status Examination (MMSE) at the onset of AD is significantly worse (p < 0.
05) (figure 3).

Longitudinally, the MMSE of LPA+ and HS decreased the fastest (p < 0.
05).

Regarding the Alzheimer's Disease Assessment Scale (ADAS-cog) subscale, memory (word recall) is initially lower in LPA, while LPA+ declines fastest
over time in this area.
At the time of AD episodes, HS is significantly worse in language (following commands) and practice (structural) than in other clusters
.
At the onset of AD, LPA+ orientation (ADAS) is worse
.
In model validation, no differences
were found between Aβ clustering states.
Information about a patient's medical history is available for the Alzheimer's Neuroimaging Program (ADNI) and the Japanese Alzheimer's Neuroimaging Program (J-ADNI), but not for the Australian Imaging, Biomarkers and Lifestyle Study (AIBL) or AddNeuroMed cohort
.

Intercept and slope covariance matrix

MA has the largest total node strength and is used as a reference group
for pairwise cluster comparisons of intercept and slope.
LPA and LPA+ have low node strengths, with few exceptions (Figure 5).

DA has high node strength only in a few medial (frontal, temporal, and occipital) brain regions (intercept and slope), while HS has high node strength
at intercept in some ventromedial prefrontal and medial temporal lobe regions.

Figure 5 Cluster-specific covariance matrix compared
to node strength.
The cluster-specific intercepts (A, C, E, and G) and slope (B, D, F, and H) covariance matrices are compared
with network theory.
The sphere diameter shows the node strength
for each region.
Regions where the node strength of the smallest shrinking clusters is higher than that of other clusters is shown in red, while blue is the opposite
.

Model validation

The author's model is validated
in two ways.
First, the authors used a separate unseen patient MRI dataset to assess whether these five longitudinal atrophy patterns could reasonably classify
the new dataset.
In addition, the authors performed separate clustering analysis on ADNI and J-ADNI/AIBL datasets
.
The clustering probability indicates that few patients have multiple clusters belonging to the discovery dataset, and even fewer in the validation dataset (0.
009% of the dataset).

Finally, in the validation dataset, the median cortex and hippocampal atrophy of each cluster at the median course of disease are highly similar to the model fit values for the same disease stage (Figure 6).

Figure 6 Comparison
of model fit values and atrophy levels of the validation dataset.
The atrophy fit of the trained clustering model after AD start with the newly validated dataset
.
Each group of new observations was classified and median disease duration
calculated.
The atrophy fit for the median duration of disease for each cluster is then calculated from the clustering model (middle column
).
The median atrophy plot (median atrophy in the group) for the new data for each cluster is shown in the left column
.
The right column shows the hippocampal volume for each cluster (boxplot color: green, minimum; n = 420, olive: edge advantage; n = 283, orange: marginal advantage +; n = 12, blue: diffuse; n = 13, purple: hippocampus; n = 8）
。 New observations (including repeated measures) and model fit values hippocampal atrophy (green vertical line).

Furthermore, when clustering is applied to the ADNI and J-ADNI/AIBL datasets, respectively, the former shows five different atrophy patterns and the latter shows four different atrophy patterns
.
The atrophy patterns found in the independent cohorts were similar to the overall discovery dataset, including MA, LPA, LPA+, DA, and HS.

In quantitative terms, MA is more similar to ADNI Cluster 3 and JADNI/AIBL Cluster 3 (in terms of intercept and slope), LPA is more similar to ADNI Cluster 2 and J-ADNI/AIBL Cluster 2, LPA+ is more similar to ADNI Cluster 1 and J-ADNI/AIBL Cluster 3, DA is more similar to ADNI Cluster 4 and J-ADNI/AIBL Cluster 4, and finally HS is more similar to ADNI Cluster 2 and J-ADNI/AIBL Cluster 1
。

discuss

A major contribution of this study is the transition of AD subtypes from a cross-sectional understanding to a longitudinal clustering perspective
.
Some previously reported AD subtypes appear to reflect different stages of the disease, which can be observed
in the five longitudinal atrophy patterns in this paper.
Thus, the authors' data help solve the long-term problem
of separating disease stages from actual disease subtypes.
This was achieved by using a well-defined time scale, i.
e.
longitudinal data modeled in a large multi-ethnic cohort of 891 AD dementia cases from four continents within 8 years from the onset of the disease
.
Another important finding is that AD subtypes with significantly different atrophy trajectories may converge
later in the disease.
This is a new understanding of AD neurodegeneration, which, combined with knowledge of neuropathology and clinical heterogeneity, can lay the foundation
for future personalized prediction of biological alterations and cognitive decline in AD.

In the simulated onset of clinical disease, the authors' approach successfully identified the same patterns of atrophy (typical, hippocampus-preserved, borderline dominance, and microatrophy) previously identified in neuropathological and neuroimaging subtype studies
.
The authors' results reveal two main atrophy pathways
.
The authors introduced the term pathway to describe AD patients who exhibited a similar spatial distribution
of atrophied brain regions over time.
Patients may progress faster (LPA+) than others (LPA and MA) in the same atrophy pathway, but their spatial distribution of atrophy is similar
over time.
This pathway contrasts with a second different atrophy pathway in AD, which has a different spatial distribution, mainly cortical atrophy
over time.
The difference in the rate of progression also reflects the rate of
cognitive decline in patients.
Understanding the underlying factors influencing differences in progress within and between pathways is a very important research direction
for the future.

In certain stages of the disease in the MA, LPA, or LPA+ longitudinal atrophy clusters, subtypes
of AD are identified with mild atrophy (limited to atrophy of the entorhinal cortex), dominant marginal atrophy (atrophy mainly in the marginal region), and typical atrophy (extensive atrophy of the hippocampus, temporal, parietal, and frontal lobes).
MA is the most representative cluster in the dataset being studied and has the highest variability
within the cluster.
It is important to emphasize that the MA cluster includes patients grouped in a slight and marginal dominant atrophy pattern, as well as AD patients
who may have been typically reported in some early literature.
Because atrophy trajectories from the onset of disease are simulated in this paper, longitudinal structural changes
in CU Aβ-negative subjects are illustrated.

Connect atrophy patterns in the literature by simulating atrophy trajectories to unambiguously stage the disease
.
The MA and LPA clusters may belong to the same AD subtype observed at two different stages, as patients with MA reach LPA levels (baseline)
two years after the onset of AD.
Differences in cognitive intercepts (MMSE and ADAS word recall) between MA and LPA clusters support the idea
that they reflect different disease stages.
LPA+ clusters appear to be on the same atrophy pathway, but have a faster
rate of atrophy compared to MA and LPA clusters.
Patients in the LPA+ cluster had the most dramatic cognitive decline, including memory and orientation, among the five identified clusters
.
Patients with LPA+ have APOE e4, education, and disease episodes
similar to MA and LPA.
However, as a representative of cognitive reserve, LPA+ has significantly higher premorbid intelligence than MA and LPA
.
The authors suggest that due to higher cognitive reserve, LPA+ patients can achieve higher levels of brain atrophy than MA and LPA clusters while maintaining similar clinical severity until they reach the onset
of AD.
Brain atrophy dynamics differ
over time in MA, LPA, and LPA+ clusters.

However, current data seem to indicate that these three longitudinal atrophy clusters belong to the same atrophic pathway in AD, the midtemporal lobe atrophy pathway
.
This atrophy in well-documented pathways has been shown to be associated
with autoanatomical neurofibrillary tangle pathology.
Although these three clusters (MA, LPA, and LPA+) belong to the same atrophy pathway, their rates of atrophy and the rate of cognitive decline vary widely, which may be clinically important
.
These observed differences may be due to a combination of protective and risk factors and potential concomitant non-AD brain lesions
.
For example, Ferreira and colleagues found that the location and frequency of markers of small-vessel disease differed
between AD subtypes.

The HS clustering in this article is similar to the hippocampus-sparing subtype described in previous neuropathological and neuroimaging subtype
studies.
Compared with other AD subtypes, this subtype is often characterized
by cortical atrophy.
Some of the features of HS clustering in this article include steep atrophy trajectories, lower frequency of APOE e4 alleles, high premorbid intelligence, higher educational attainment, and early-onset AD, which are consistent with
hippocampal reserved subtype-related features reported in previous studies.
This cluster has the lowest frequency, which is also consistent
with previous studies.
Significantly affected tectonics and conceptual practices are a key feature of hippocampal retention subtypes, which were also confirmed
in this study.
A comparison between MA and HS clustering covariance patterns reveals network differences
between the two groups.
In MA, anatomical differences due to disease are mainly confined to the medial temporal lobe and cortical regions
that act as a network combination at the onset of AD.
On the other hand, differences in HS clustering networks at the onset of AD also involve basal ganglia
.
In addition, HS clustering has higher nodal strength
at intercepts of some ventromedial prefrontal lobes and medial temporal regions from MA clusters.
Based on all these results, the authors suggest that the HS-type atrophy pattern represents a unique atrophy pathway in AD, the cortical pathway
.

Interpreting the atrophy trajectories of DA clusters is challenging because excessive frontal and temporal lobe atrophy
is already present at the time of clinical onset.
Data from this paper show that in the late stages of the atrophied midtemporal lobe and cortical pathways, AD patients may show similar levels
of atrophy to DA clusters.
Therefore, this group of patients may belong to either of the two pathways of atrophy
.
Similar to LPA+, cognitive reserve (average education over 15 years) in DA clusters can explain greater levels of atrophy (at the onset of dementia).

The atrophy pattern of DA clustering is similar to, but less frequent, typical AD atrophy subtypes reported in the literature
.
In a recent cross-sectional clustering study using tau PET, which mainly included preclinical AD, none of the clusters had a spatial distribution of tau similar to a typical AD atrophy pattern, but cortical and medial-temporal patterns
of tau were observed.
In addition, two other studies on prodromal AD found that people with decreased glucose metabolism in the temporoparietal lobe or increased temporoparietal lobe atrophy (typical AD pattern), but with a lower sample frequency, are consistent with
the findings of this paper.

This study addresses some important methodological challenges as the first time AD atrophy subtypes
have been identified based on longitudinal biomarker trajectory modeling.
An immediate advantage of the longitudinal clustering approach overcomes the assumption that subjects in clustering (cross-sectional analysis) remain in the same cluster as the disease progresses, which is unrealistic
.
Previous studies have employed arbitrary timescales to mimic the progression of
biomarkers.
The assessment herein is based on a well-defined timescale, i.
e.
, the time
from the onset of clinical onset.
This approach helps track abnormal changes
early in the course of each cluster.
Previously, longitudinal explanations could not be directly correlated with existing data because they were not anchored to a specific time frame
.
In this paper, the atrophy w value of each patient was calculated based on the dataset of longitudinal Aβ-negative CU individuals, correcting for the effect of
age in brain morphology.

The studies also had some limitations
.
Atrophy trajectories were modeled
only in the context of AD heterogeneity.
Pre-AD scanning was not included, reducing the ability to
infer atrophy patterns before diagnosing AD dementia.
In addition, the addition of biomarkers of non-AD pathology to cluster study designs in the future will help to understand the role of
comorbidities in AD subtypes.
In addition, the AD patients included in the studies included here had a shorter follow-up time, and future reestimates of atrophy trajectories will include more MRIs per patient for better assessment
.

In summary, based on a large multi-ethnic cohort of AD dementia patients, the authors identified five longitudinal patterns of brain atrophy that divided previously reported AD subtypes into two atrophy pathways (midtemporal lobe and cortex).

The authors transferred the cross-sectional understanding of AD subtypes to longitudinal clustering
.
The authors' study is a step toward answering the pressing question of whether the heterogeneity observed in AD reflects disease stages or different biological subtypes
.
With the help of the model proposed by the authors, it is possible to unravel heterogeneity in AD, enabling precision medicine and potentially improving the treatment of
diseases in the future.