In the current study we conducted the first data-driven characterization of systematic heterogeneity in individual-level FDG-PET patterns among AD dementia patients, and provide evidence for the existence of three distinct hypometabolic subtypes of AD. These subtypes include a “typical” subtype of posterior temporo-parietal hypometabolism, as well as distinct “limbic-predominant” and “cortical-predominant” subtypes that resemble previously described MRI-based atrophy subtypes of AD and show corresponding differences in their clinical profiles. By stratifying an independent sample of longitudinally followed prodromal AD patients according to hypometabolic subtype we could further demonstrate that these subtypes can be detected at a prodromal disease stage and are characterized by differential courses of cognitive decline.
Distinct hypometabolic subtypes among patients with AD dementia
The “typical” subtype included the largest portion of AD dementia cases and was characterized by a typical posterior temporo-parietal pattern of hypometabolism that is commonly linked to AD (16, 42). The “limbic-predominant” subtype had most pronounced hypometabolism in the hippocampus and related medial temporal structures which showed similarities to the MRI-defined medial temporal-dominant atrophy subtypes (4, 33, 43, 44). Despite these similarities, the “limbic-predominant” subtype in the current study notably differs from previously described medial temporal-dominant atrophy subtypes by showing a more extensive hypometabolic pattern covering widespread limbic areas beyond the medial temporal lobe and including the frontal cortex. A previous study examining FDG-PET patterns in MRI-defined atrophy subtypes also reported pronounced frontal hypometabolism in the medial temporal-dominant AD atrophy subtype (44). This could be potentially attributed to differences in the specific pathologic substrate of hypometabolism on FDG-PET and grey matter reductions on MRI as neurodegeneration markers. Specifically, hypometabolism on FDG-PET has been reported to be sensitive to neurodegenerative dysfunction that does not regionally co-localize with neuronal loss as measured by grey matter atrophy on MRI, which may reflect early non-macroscopic neurodegenerative processes or functional changes caused by atrophy in remote but functionally interconnected brain areas (8, 17, 45-47).
Similarly to previous findings on the MRI-defined medial temporal-predominant atrophy subtype (4, 43), the “limbic-predominant” hypometabolic subtype in the current study was associated with older age and could possibly reflect the effects of comorbid age-related pathologies. For example, Zhang et al. (5) considered that the temporal factor described in their study could be linked to comorbid TDP-43 pathology. Indeed, the hypometabolic pattern of the “limbic-predominant” subtype identified in the present study shows a striking resemblance with a recently described FDG-PET pattern of pathologically confirmed patients with AD dementia with comorbid TDP-43 pathology and hippocampal sclerosis (HS) (48) (specifically, Fig. 3, page 1209 in that study). Corroborating this qualitative visual interpretation, in a complementary post-hoc analysis we found that the “limbic-predominant” subtype had a significantly higher inferior-to-medial temporal FDG-PET ratio compared to both the “typical” and “cortical-predominant” subtypes (Supplementary table 6). This ratio has been suggested to reflect the difference between the TDP-43/HS-related pattern and the AD-typical pattern within a simplified metric, and has been proposed as an imaging biomarker for comorbid TDP-43/HS in AD (48, 49). However, it is important to note that any possible involvement of comorbid pathologies in the observed hypometabolic subtypes remains entirely speculative in our in-vivo neuroimaging study, and we did not observe any notable differences in the proportion of AD-specific A/T biomarker profiles in the limbic-predominant subtype compared to the other subtypes (Supplementary table 1). Other studies suggested that the medial-temporal subtype could be additionally affected by small vessel disease (4, 50), which would coincide with the numerically highest WMH volume in the “limbic-predominant” subtype in our study. However, this difference did not reach statistical significance in our analysis. Additional neuropathologic examinations as well as mechanistic studies are needed to better understand the exact pathological substrates and neurobiological mechanisms that drive different neurodegeneration subtypes in AD.
The hypometabolic pattern of the “cortical-predominant” subtype was similar to that of the “typical” subtype, but with more extensive involvement of the frontal lobe and largely normal metabolism in the medial temporal lobe. This subtype showed particularly pronounced executive function impairment in addition to the memory deficit. Previously, studies by Collette et al. (11) and Mosconi et al. (12) have also described marked frontal hypometabolism in subsets of patients with AD. On the other side, Ossenkoppele et al. (51) described an autopsy/biomarker-confirmed dysexecutive AD variant, which shows markedly more pronounced impairment in executive function relative to the memory deficit, and is characterized by early onset of AD and a relatively low APOE ε4 frequency. Similarly, the “cortical-predominant” subtype in the current study also showed the youngest age and lowest percentage of APOE ε4 carriers among AD subtypes. However, due to the low number of patients in this group, current findings on this subtype require further corroboration.
In accordance with the subtype-defining hypometabolic patterns, we observed a difference between subtypes in the HV:CTV ratio. Specifically, it was the highest for the “cortical-predominant” subtype, intermediate for the “typical”, and numerically the lowest for the “limbic-predominant” subtype. The pattern of differences in HV:CTV ratio between FDG-PET subtypes in the current study is comparable to previous findings on AD subtypes based on neuropathological data or MRI-based atrophy patterns. The study by Whitwell et al. (2) examined AD subtypes based on neuropathological examination of distribution of neurofibrillary tangle counts. The ratio between hippocampal and cortical volumes measured on ante-mortem MRI allowed for the best discrimination between these subtypes. In their study, similarly to our results, the typical subtype showed a higher HV:CTV ratio than the limbic-predominant subtype, whereas the hippocampal sparing subtype had the highest value. Furthermore, in the study by Risacher et al. (28), three AD subtypes - hippocampal sparing, limbic predominant and typical AD - were defined based on the HV:CTV ratio. Across these subtypes, a higher HV:CTV ratio was also quantitatively associated with a more pronounced dysexecutive profile, similarly to the differences observed for the ADNI-EF and ADNI-DIFF variables between the subtypes in the current study. In a complementary analysis we could also reproduce this association between HV:CTV ratio and cognitive profile on a continuous scale (Supplementary table 7). Thus, across patients the HV:CTV ratio correlated positively with the ADNI-DIFF variable in both the AD dementia and prodromal AD groups, indicating a more pronounced executive function over memory deficit for higher values of this ratio. Therefore, the HV:CTV ratio measured in the current subtypes provides a link between our findings on hypometabolism subtypes and previously characterized AD subtypes based on neuropathological data or MRI-based atrophy patterns. However, there were also differences between the observed hypometabolism subtypes and previously reported atrophy patterns on MRI, such as the aforementioned more extensive involvement of the temporal and frontal areas in the “limbic-predominant” hypometabolism subtype, which might be attributed to the different structural and functional substrates of the respective imaging methods.
Stratification of prodromal AD patients according to hypometabolic subtype
Stratification of the prodromal AD cohort according to hypometabolic subtypes revealed a considerably sized subgroup of patients with no or only minimal hypometabolism. This “no hypometabolism” subtype also showed lower levels of Aβ biomarker burden, was less cognitively impaired, and had a lower risk of progressing to dementia compared to the other subtypes. While this may indicate that this subtype may be enriched for patients with only incidental amyloidosis, the comparably high proportion of concomitant tau biomarker positivity in this group (Supplementary table 1) would rather argue against this possibility.
Previous studies using visual classification of FDG-PET scans had also described subsets of patients with MCI without evidence of regional hypometabolism (12, 14). In the study by Cerami et al. (14), 31% of participants with MCI showed normal brain metabolism, although the large majority of these also had a negative amyloid biomarker finding. However, MRI-based subtyping studies have also consistently identified subsets of patients with AD dementia with no or only minimal atrophy, and this subtype was particularly prevalent among patients with prodromal AD (4, 43, 52). Interestingly, in our study we observed a similar “minimal” hypometabolism subtype in the AD dementia group when using a higher clustering solution (see Supplementary figure 2). However, since we established the best distinguishable AD subtypes using an objective hierarchical clustering cutoff as suggested by the Davies-Bouldin and the silhouette criteria, we did not further characterize this “minimal” subtype in our study. Nevertheless, our findings underline the importance of accounting for the considerably sized subgroup of patients with prodromal AD without evidence of regional hypometabolism when characterizing heterogeneity of hypometabolism patterns in this population.
Only three participants with prodromal AD were classified into the “cortical-predominant” subtype and could thus not be further analysed in our study. One potential explanation for this low prevalence could be that this hypometabolic subtype is characterized by more pronounced executive function deficits than memory deficits from its prodromal stage on, so that these patients would be underrepresented in an MCI cohort screened for memory deficits such as the ADNI cohort.
The “limbic-predominant” and “typical” subtypes classified in the prodromal AD sample demonstrated similar subtype characteristics as in the AD dementia sample. Thus, the “limbic-predominant” subtype also had older age, numerically higher WMH volume, and the most severe degree of hippocampus atrophy. Hence, current results confirm previous findings that the heterogeneity evident in patients with AD dementia can also be observed at the prodromal stage of the disease (4, 5). Interestingly, although the “limbic-predominant” and “typical” subtypes had a comparable risk of progressing to dementia, the “limbic-predominant” subtype showed a more memory-selective cognitive decline compared to the “typical” subtype, paralleling the cross-sectional subtype differences in cognitive profiles observed in the AD dementia cohort. This finding is notable, because these subsets of patients with prodromal AD did not show significant differences in the respective cognitive functions at baseline. This indicates that subtype classification of FDG-PET patterns may provide additional information for predicting future cognitive decline that is not contained in neuropsychological assessments.
Our current findings on subtype-specific trajectories of cognitive decline are largely consistent with previous findings on differences in dementia risk and domain-specific cognitive decline between atrophy subtypes based on MRI data. For example, the studies by Ten Kate et al. (4) and Dong et al. (5) both only found a significantly lower risk for progression to dementia in the no/minimal atrophy subtype, and while the other atrophy subtypes showed similar overall risk for progression to dementia they differed significantly in the relative decline in specific cognitive domains.
A previous study by Morbelli et al. (53) used classical voxel-wise analyses of FDG-PET images in a group of prodromal AD patients to determine a “prognostic pattern” of regional brain hypometabolism that best correlated with time to conversion to dementia. Interestingly, this “prognostic pattern” was found to be considerably different from the AD-typical “diagnostic pattern” of hypometabolism as determined by contrasting AD patients with healthy controls. In our current study we used a data-driven approach to demonstrate that this typical group-averaged “diagnostic pattern” can be decomposed into different regional subtypes corresponding to distinct subgroups of AD patients. It is likely that these different FDG-PET subtypes also correspond to different “prognostic patterns” that best predict time to dementia conversion in the respective subgroups of prodromal AD patients.
Strengths and Limitations
One conceptual strength of the current study is that we only included patients with biomarker evidence of Aβ pathology. Moreover, we used the identified hypometabolism subtypes in patients with AD dementia to classify an independent dataset of Aβ-positive patients with MCI and assess clinical and biomarker characteristics of these subtypes at a prodromal disease stage.
As with other unsupervised subtyping studies, a principal limitation of the current study is that the employed clustering methodology cannot naturally distinguish between subtypes and different disease stages. We aimed to mitigate the effect of differing disease stages by normalizing the individual FDG-PET profiles to their global signal before clustering, so that the cluster assignations were primarily driven by relative regional metabolic differences instead of global differences accompanying disease progression. Global signal scaling is a commonly used method to control neuroimaging clustering analyses for individual differences in disease severity, and analogous approaches have been used in previous FDG-PET subtyping studies in other neurodegenerative dementias (32, 34), as well as in MRI-based subtyping studies of neurodegeneration heterogeneity in AD (see e.g. (3) for a recent review of this literature). We also note that the hypometabolic characteristics of the identified subtypes would not be consistent with the notion of merely reflecting variability in disease severity. As an example, the interpretation that the “limbic-predominant” subtype would reflect an earlier stage of the “typical” subtype cannot be easily reconciled with the observation of more severe medial temporal hypometabolism in the “limbic-predominant” subtype. A similar argument also applies to the comparison between the “typical” and “cortical-predominant” subtype (also see Supplementary figure 3 for a direct voxel-wise comparison of regional hypometabolic differences between the subtypes)
The contribution of disease stage to observed subtype phenotypes has been addressed in a recent MRI-based subtyping study by Young et al. (54), which proposed an analytical approach combining clustering with event-based modelling to assess subtypes and their respective stage progressions at the same time. However, this approach still relies on extrapolations from cross-sectional data. Future research on neurodegeneration subtypes in AD will benefit from longitudinal imaging assessments allowing to directly characterize disease progression within subtypes and to determine the possibility of conversion between them.