The heterogeneity of Alzheimer’s disease

Several years ago, when being interviewed for a grant, I found myself defending the claim that Alzheimer’s disease (AD) is a heterogeneous condition, that it consists of subtypes, and that it does not always follow a stereotyped trajectory. One expert on the grant assessment panel, who wasn’t a clinician, was having none of this. We know that the pathology starts in the entorhinal cortex and spreads from there in a predictable fashion to the hippocampus and beyond, they asserted. Pathological studies had shown this definitively to be the case and, accordingly, episodic memory deficits were therefore the earliest manifestation of AD. So why was I disputing the issue?

Keeping calm in the face of what seems self-evidently to be wrong, especially when so much depends on the outcome of one’s responses, is not always easy. I pointed out that we see AD patients in the clinic who simply do not conform to the canonical example that the panelist had portrayed. Some of our cases present first with a language disorder (logopenic primary progressive aphasia), I explained. Others start with a visual agnosia and visuospatial impairments (posterior cortical atrophy); and still others begin with early executive function deficits (frontal variant). Although presenting first with the amnestic syndrome is indeed the common variant of AD, there are others. Even within the amnestic subgroup, cognitive presentations and trajectories of progression vary considerably. My argument was not gaining traction, however, because to the expert it was seemingly based on anecdotal evidence. Of course, there are always outliers, but what we are talking about is the common or garden variant of AD, reasserted the panelist.

Since that interview, the evidence for AD heterogeneity has become far stronger. The data confirming that different phenotypes of AD are associated with various patterns of brain atrophy have slowly accumulated,^1-3 and it is now recognized that factoring in such heterogeneity may be crucial to the design of clinical trials.⁴ These and other considerations have led researchers to develop and test data-driven methods to distinguish different subtypes of AD. A major complication with such approaches is that patients might not only belong to different phenotypical groups but also be at different points in their disease course. Disentangling subtype and disease-stage heterogeneity is not straightforward. However, a computational technique, Subtype and Stage Inference (SuStaIn), is claimed to do just that. SuStaIn is a machine learning approach that combines both clustering and disease progression modelling to identify disease subtypes.⁵ Cross-sectional ‘snapshots’ can be used as the input, and the validity of the approach can be assessed using longitudinal data. An initial study revealed the potential power of this technique for AD as well as genetic subtypes of frontotemporal dementia.

Now, in this issue of Brain, Baumeister and colleagues⁶ present the results from a much larger dataset from the German Center for Neurodegenerative Diseases (DZNE), which was analysed using the SuStaIn algorithm and replicated in the Swedish BioFINDER-2 cohort. In addition to AD cases, the study included cognitively intact individuals as well as patients with mild cognitive impairment and subjective cognitive decline. Two AD subtypes were identified. The ‘limbic predominant’ group had atrophy that first involved the medial temporal lobes (MTL), followed by further temporal regions and then remaining cortical areas. In contrast, the ‘hippocampal-sparing’ group showed evidence first of atrophy outside the temporal lobe, with the MTL remaining relatively intact, even with advanced progression. The latter group made up nearly 30% of the DZNE AD cohort and 44% of the Swedish validation cohort. Importantly, the pattern of future atrophy was predicted by the baseline subtype. Thus, group allocation at the first scan held true following subsequent imaging, reassuringly indicating that the methodology is robust.

Limbic predominant AD cases had amnestic cognitive impairment as their clinical presentation and typically worse overall cognitive scores, with higher pathological AD biomarker levels in CSF. They were like the canonical example that my interview panelist had in mind, whereas the hippocampal-sparing patients were very different: they had more generalized cognitive deficits. Because longitudinal data were available for some individuals, it was also possible to gain insights into the trajectory of cognitive decline. This occurred more precipitously in the limbic predominant subtype of AD, and the same pattern was observed in individuals who were cognitively intact at baseline but had this pattern of brain volume loss.

The findings of this study make it very clear that, although there might be a canonical exemplar of AD, there is also significant heterogeneity, a conclusion that has also been reached through pathological studies.⁷ Intriguingly, Baumeister et al.⁶ noted that, among the CSF amyloid-β positive cases, the levels of AD CSF biomarkers were not significantly different between the limbic predominant and hippocampal-sparing groups and hence cannot explain the differences in cognitive decline between these groups. Instead, there is a possibility that individuals with limbic predominant AD might actually have additional non-AD co-pathology which accounts for the differences in progression observed between subtypes. If this were indeed to be the case, it would be quite a paradoxical finding, given the supposedly ‘typical’ nature of AD, which clearly had been imprinted in the mind of my questioner on the interview panel.

Fortunately for me, I did manage to obtain that grant thanks to the open-mindedness of the other panel members. The new data from Baumeister and colleagues⁶ now provide strong evidence—for clinicians and non-clinicians alike—that we should not focus on early hippocampal volume loss as an MRI biomarker of AD, which clearly is a far more heterogenous condition than some researchers once thought.

References