Introduction
The biological framework of Alzheimer’s disease (AD) recognizes beta-amyloid (Aβ), tau, and neurodegeneration as the characteristic biomarkers in disease pathogenesis [
1]. Among the AD hallmarks, spread of Aβ in the brain is rather diffuse whereas the accumulation of tau occurs in a more ordered manner [
2]. Occurrence of neurodegeneration downstream to Aβ and tau has impelled several investigations on the relationship among these biomarkers, suggesting a closer association between neurodegeneration and tau than neurodegeneration and Aβ [
3‐
6].
Biological heterogeneity in AD manifests as distinct patterns of biomarkers in the cognitively normal and prodromal stages. In contrast to the biomarker-based
subtypes which are typically found at the dementia stage and are unlikely to change, biomarker-based
patterns are more likely to evolve and change over time as the disease progresses. Neuroimaging studies have shown topographical conformity and association between tau pathology from tau positron emission tomography (tau-PET) and longitudinal brain atrophy-based neurodegeneration from magnetic resonance imaging (MRI), in cognitively unimpaired individuals [
7], prodromal AD and/or AD dementia [
4,
8,
9], and clinical subtypes of AD [
10,
11]. A critical caveat, however, is the failure to account for heterogeneity in tau-PET topography at a given disease stage (i.e., tau patterns in cognitively normal and prodromal stages or subtypes at dementia stage) [
12‐
16]. The relationship between tau-PET patterns and atrophy remains unexplored and is critical for precision medicine.
To this end, our study aims to provide two complementary perspectives on this issue (see Supplementary Figure
1 for study design): (a)
Biological perspective: we investigated the association between different tau-PET patterns and longitudinal atrophy in the AD continuum (cognitively normal, prodromal AD, AD dementia cases with Aβ pathology); and (b)
Methodological perspective: we characterized tau-PET patterns on a continuous scale inspired by the recent conceptual framework [
17], compared to and extending beyond the conventional characterization of discrete categorization [
14‐
16,
18]. This continuous-scale operationalization comprises two key dimensions including typicality (spanning from limbic predominant to cortical predominant patterns) and severity (spanning from typical AD to minimal tau patterns). Together, these dimensions represent the heterogeneity of an individual as a combination of protective factors, risk factors, and concomitant comorbid pathologies in AD.
Corresponding to these two perspectives, we hypothesized that (a) biologically, tau-PET patterns would modulate the association between baseline tau-PET and longitudinal atrophy differentially; and (b) methodologically, treating heterogeneity (i.e., the different tau-PET patterns) on a continuous scale over a discrete scale can potentially be more efficient for future research.
Discussion
We investigated the association between heterogeneity in tau-PET and longitudinal neurodegeneration (atrophy) in the AD continuum. As hypothesized: (a) from a
biological perspective, different tau-PET patterns revealed a differential association with longitudinal atrophy; and (b) from a
methodological perspective, characterizing heterogeneity on a continuous scale may be more useful than the conventional categorization of individuals into discrete patterns. Recent studies have investigated the association between tau pathology and downstream neurodegeneration in healthy, cognitively normal, prodromal AD, and AD dementia [
5,
7‐
9,
36‐
38] individuals, as well as in clinical subtypes of AD [
39‐
41]. To our knowledge, our study is the first to characterize the role of biological heterogeneity (tau-PET patterns) as a modulator of the association between tau pathology and neurodegeneration.
The four tau-PET patterns captured by the continuous and discrete scales in our study are reminiscent of the biological tau-PET AD subtypes [
14,
15,
18,
42]. With regard to the cortical predominant tau-PET pattern in particular, further analysis of different regions in the cortex revealed that this pattern had relatively lower tau burden particularly in the medial temporal regions. Compared to previous studies describing cortical predominant subtypes in tau-PET (occipital-dominant/visual variant, left hemisphere-dominant/language variant, etc.) [
16,
30,
31], the cortical predominant pattern in our sample is reflective of an amnestic phenotype. In this study, we describe heterogeneity in terms of tau-PET
patterns and not
subtypes.
Subtypes are conventionally reported in the advanced disease stage such as in AD dementia and may potentially be less likely to change into a different subtype. However, given that our cohort additionally included individuals at earlier disease stages such as at pre-dementia stages, there may be a possibility that the
pattern exhibited currently may eventually evolve and transition into a
different pattern at advanced disease stages. Thus, different tau-PET topographies which may represent current
patterns in our cohort at early disease stage may be more appropriately described as
subtypes in AD dementia. The value of identifying
patterns lies in that heterogeneity in tau pathology may be detectable at early stages of the disease. Our study confirms the findings from the recent report identifying four discrete trajectories in tau-PET within AD continuum [
16]. The novelty of our findings lies in the associations between baseline tau-PET and longitudinal atrophy across heterogeneity and the realization of heterogeneity as a continuous phenomenon.
The prevalence of the identified tau-PET patterns differed slightly from previous reports: on the continuous scale (Fig.
1 A, B), a large proportion of the individuals exhibited intermediate values of typicality and low values of severity (lower variance in prodromal AD and cognitively normal may suggest less heterogeneity); on the discrete scale (Fig.
4), minimal tau was the most prevalent pattern (37%) and the cortical predominant pattern (18%) was more prevalent than the limbic predominant pattern (12%). This breakdown of prevalence of the patterns is different when considering AD dementia cases alone (Table
2)—typical AD pattern was the most prevalent and minimal tau pattern was the least prevalent. Thus, the discrepancy in prevalences of tau-PET patterns is likely owing to the large proportion of individuals at early disease stages (Aβ+ cognitively normal and prodromal AD), who may have not accumulated considerable amount of tau pathology, which is typical to AD. Additionally, a current tau-PET pattern at early disease stages may likely evolve into a different pattern at a later timepoint. This may explain why the demographic/clinical profiles of our tau-PET patterns (Table
2) do not entirely conform with the expected profiles previously reported in AD [
17]. Similar results have been found when characterizing heterogeneity in tau-PET in the AD continuum [
16], atrophy in prodromal AD [
43] and glucose-hypometabolism in prodromal AD [
44]. These differing prevalences may be a function of the predominant disease stage in the cohort in addition to the cutpoints used to determine abnormality in the brain regions. Altogether, heterogeneity at preclinical and prodromal stages of AD may be similar to, albeit less pronounced than, heterogeneity in AD dementia. Modeling heterogeneity on a continuous spectrum may offer an avenue to circumvent the lack of generalizability of specific prevalences of subtypes in a disease population.
Our main finding was that tau-PET patterns showed differential association with longitudinal atrophy. On the continuous scale (Table
3), typicality was significantly associated with longitudinal atrophy in the entorhinal cortex but not the neocortex. This result highlights that tau pathology in the entorhinal cortex (signature of a limbic predominant pattern) can be tracked by longitudinal atrophy in the region, with greater atrophy seen in the highest extreme of typicality (limbic predominant pattern) [
45] compared to the lowest extreme (cortical predominant pattern). However, tau pathology in the neocortex (signature of a cortical predominant pattern) cannot necessarily be tracked by longitudinal atrophy in the region, with comparable atrophy seen in the limbic predominant and cortical predominant patterns. On the other hand, severity was significantly associated with longitudinal atrophy in both the entorhinal cortex and the neocortex. This result highlights that greater tau burden in the entorhinal cortex and neocortex (signature of typical AD pattern) can be tracked with greater atrophy in these regions in the highest extreme of severity (typical AD pattern) compared to the lowest extreme (minimal tau pattern).
On the discrete scale (Table
4), baseline tau-PET patterns were associated with greater longitudinal atrophy for typical AD and limbic predominant patterns but not the cortical predominant pattern in the entorhinal cortex. Baseline tau-PET pattern was associated with greater longitudinal atrophy for the typical AD pattern only in the neocortex. This result highlights a region-specific differential association between tau-PET patterns and atrophy. Typical AD and limbic predominant patterns showed increasing topographical correspondence between baseline tau-PET and atrophy over time while cortical predominant and minimal tau did not (visualized in Fig.
4). The two latter patterns showed marked atrophy in brain regions non-specific to the tau-PET patterns (e.g., entorhinal atrophy in cortical predominant; cortical atrophy in minimal tau), indicating that atrophy may not always regionally follow the different tau-PET patterns. Conversely, topographical correspondence has been reported between tau-PET and MRI in atrophy-based AD subtypes [
46]. Combining findings from this study with ours may imply that heterogeneity of a downstream event (atrophy) may be reflected in an upstream event (tau pathology) but not vice versa. Downstream contributions of other neuropathologies towards atrophy may play a role in determining heterogeneity [
47] and need to be considered as biomarkers for those pathologies become available. Altogether, considering tau pathology as a sole or main driver of neurodegeneration may be a simplification and understanding of disease heterogeneity requires a more unifying approach [
48].
Across the continuous- and discrete-scale characterizations of tau-PET patterns, longitudinal atrophy associated with baseline tau pathology supports the hypothesis of tau pathology as a possible driver of atrophy [
4,
7,
36,
49], observed across some but not necessarily all tau-PET patterns. Although findings from both the characterizations are consistent, the continuous-scale approach was significantly better than the discrete-scale one in being able to model longitudinal atrophy. While the continuous-scale approach characterizes the tau-PET patterns in terms of typicality and severity, two continuous dimensions of biological AD subtypes proposed by the recent conceptual framework [
17], the conventional discrete-scale approach categorizes individuals into four discrete patterns based on the contribution of the entorhinal cortex and neocortex [
26]. Typicality in the continuous-scale approach in fact factors in contributions from both the entorhinal cortex and neocortex used in the discrete-scale approach and further provides information on disease stage in terms of severity. The continuous-scale approach avoids arbitrary cutpoints, making it suitable for populations where the prevalence of different patterns is not well-known (e.g., beyond AD dementia including the AD continuum) and to small cohorts. The discrete-scale approach defines patterns based on a cutpoint (e.g.,
Z-score>1 relative to healthy Aβ− individuals in our study) [
26,
35], influencing the prevalence of the identified patterns. Comparing across the four discrete-scale tau-PET patterns by the continuous-scale typicality and severity, we observed that each pattern was significantly different from the others in typicality as well as severity. It is thus, important to bear in mind that the discrete-scale tau-PET patterns representing heterogeneity are at different disease stages. Nevertheless, both approaches share some correspondence (Fig.
1): examining typicality, higher E:N may reflect a limbic predominant pattern while lower E:N may reflect a cortical predominant pattern; examining severity, higher global tau-PET SUVR may reflect a typical AD pattern while lower global tau-PET SUVR may reflect a minimal tau pattern. All previous subtyping methods in AD characterized heterogeneity on a discrete scale [
14‐
16,
18], which is critical to delineate pattern-specific characteristics. However, discrete-scale characterizations often lack individual-level agreement [
15]. A continuous-scale characterization of heterogeneity may be more useful as it is free from the assumption of pre-defined prevalence in a population. Hence, we encourage future studies to explore and validate new operationalizations of typicality and severity representing disease heterogeneity. Compared to the discrete-scale characterization of the tau-PET patterns which force-classifies each case into one of four categories (typical AD, limbic predominant, cortical predominant, minimal tau), the continuous-scale characterization additionally provides information on the extent of typicality and severity of each individual relative to others, thus, disentangling subtypes from disease stage, which could better inform the design of future clinical trials.
Furthermore, we noted differential profiles of the A/T/longitudinal-N biomarker scheme across tau-PET patterns (Fig.
5). Per definition, while the limbic predominant pattern demonstrated T+ in the entorhinal cortex and T− in the neocortex, the cortical predominant pattern demonstrated the opposite profile. This contrast may suggest a non-uniform sequence of tau accumulation across the tau-PET patterns. This aligns with the proposed hypothesis of alternative possible pathways for initiation/spread of tau pathology in the cortical predominant pattern [
50]. All patterns showed some longitudinal neurodegeneration (adjusted for age), but only typical AD and limbic predominant patterns showed ≥50% prevalence of longitudinal N+. Combined with reports suggesting a preferential association of atrophy to tau pathology over Aβ [
4,
51], this result may imply that atrophy may not entirely be tau-related and could be partly tau-independent, extending beyond the effect of normal aging. The minimal tau pattern presented a greater prevalence of T− both in the entorhinal and the neocortical regions. Relatively small proportion of the minimal tau cases show longitudinal N+. This may indicate the minimal tau group, while mostly reflecting Alzheimer’s pathologic change (A+/T−/longitudinal N−), could also contain cases with Alzheimer’s and concomitant suspected non-Alzheimer’s pathologic change (A+/T−/longitudinal N+). Whether the minimal tau pattern will remain as such or is a precedent manifestation of one of the other three tau-PET patterns will require analysis of longitudinal tau-PET. One caveat, however, is that the prevalence of A/T/longitudinal-N profiles varied widely depending on the cutpoint used (Supplementary Tables
2-
3), an issue that is known in the field [
1] which should be taken into account in future studies.
Our study has some limitations. Although the overall goal of our study was to understand the heterogeneity in tau-PET patterns across the AD continuum, cognitively normal individuals (Aβ+) were overrepresented. This dominance of the early stages of AD likely translated to the relatively less pronounced tau-PET patterns. Moreover, the ability of [
18F] AV-1451 tracer in detecting tau pathology may be limited at these early disease stages [
52]. Quantification of tau-PET patterns was based on tau-PET SUVR in the entorhinal and neocortex, regions with different availability of binding sites for this tracer [
53]. Thus, alternative operationalizations of typicality, keeping in mind the relationship to severity, should be assessed in future work. Hippocampus, a key region in most neuropathological and MRI studies investigating heterogeneity in AD [
12,
13,
25,
26,
54,
55], was not evaluated as its signal is confounded by off-target binding in tau-PET [
27,
28,
56]. Thus, the limbic predominant pattern observed in our study may not directly be comparable to a limbic predominant subtype reported in postmortem investigation [
12]. However, we have previously shown that tau-PET patterns based on the entorhinal cortex are similar to those based on hippocampus [
15]. While we used cerebellum gray matter as a reference region for tau-PET, future studies would benefit from exploring alternative reference regions to minimize the spill-in effects [
57]. Tau-PET patterns on the discrete scale may be influenced by the relatively lenient cutpoints used in our study. Given the large range of tau-PET cutpoints reported in the literature, future studies should focus on continuous characterization where possible or apply standardized thresholds [
58]. Although we tracked the longitudinal atrophy changes relative to baseline tau pathology, we could not assess longitudinal tau-PET changes due to the limited samples of longitudinal tau-PET in the ADNI. Finally, considering the strict inclusion criteria in ADNI, the generalizability of our findings in a clinical setting or more heterogeneous population including non-amnestic clinical phenotypes remains to be validated.
In conclusion, we demonstrated that the associations are not the same between different tau-PET patterns and longitudinal atrophy in the AD continuum. Methodologically, we posit treating heterogeneity as a continuous phenomenon over the conventional discrete categorization. Together, our findings can have practical implications towards the design of clinical trials, development of targeted therapeutics, and ultimately, realization of precision medicine.