Reduced expression of fMRI subsequent memory effects with increasing severity across the Alzheimer’s disease risk spectrum

Abstract

In functional magnetic resonance imaging (fMRI) studies, episodic memory is commonly investigated with the subsequent memory paradigm in which brain activity is recorded during encoding and analyzed as a function of subsequent remembering and forgetting. Impaired episodic memory is common in individuals with or at risk for Alzheimer’s disease (AD), but only few studies have reported subsequent memory effects in AD or its risk states like mild cognitive impairment (MCI). One reason for this might be that subsequent memory responses may be blunted in AD or MCI and thus less likely to manifest in fMRI signal differences. Here, we used Bayesian model selection of single-subject fMRI general linear models (GLMs) for a visual novelty and memory encoding experiment to compare the model performance of categorical and parametric subsequent memory models as well as memory-invariant models in a clinical cohort (N = 468) comprising healthy controls (HC) as well as individuals with subjective cognitive decline (SCD), MCI, and AD, plus healthy relatives of AD patients (AD-rel). We could replicate the previously reported superiority of parametric subsequent memory models over categorical models (Soch et al., 2021a) in the HC and also in the SCD and AD-rel groups. However, memory-invariant models outperformed any model assuming subsequent memory effects in the MCI and AD groups. In the AD group, we additionally found substantially lower model preference for models assuming novelty compared to models not differentiating between novel and familiar stimuli. Our results suggest that voxel-wise memory-related fMRI activity patterns in AD and also MCI should be interpreted with caution and point to the need for additional or alternative approaches to investigate memory function.

Single-value fMRI scores reflect both stage and risk across the Alzheimer’s disease continuum

“…Unlike previous studies of neuropsychological performance in healthy relatives of patients with AD (for a review, see 67 ), we additionally found no performance difference between healthy relatives and control participants (see Fig. 1 in 38 ). However, unlike HC and similar to individuals with SCD, healthy relatives exhibited a significant effect of amyloid positivity on the novelty-SAME score (Figure 5).…”

“…To control for potential confounding variables, we computed a three-way between-subjects ANOVA to assess effects of (i) acquisition site (8 sites; cf. Table 1 in 38 ), (ii) gender (male vs. female), and (iii) diagnostic group (HC, SCD, MCI, AD, AD-rel). Because the factor site had eight levels, we did not include interactions with site in this model.…”

“…In voxel-wise fMRI data analysis, we found novelty processing and subsequent memory to engage overlapping temporo-parieto-occipital networks, accompanied by DMN deactivations (Figure 1C/D and Supplementary Figure S1), also replicating earlier studies with other stimulus types and encoding tasks 25,26 . (Please note that, despite their anatomical overlap, novelty processing and successful encoding both contributed to fMRI signal variance explanation within the memory network in the HC, SCD, and AD-rel groups, albeit not in the MCI and AD groups 38 .) Compatibly, we found that single-value scores from the HC group were statistically indistinguishable from those of healthy older adults in the preceding study (Figure 2 and Supplementary Figure S6).…”

“…This model ("GLM_1t-a", cf. Table 3 in 38 ) had emerged as the winning theoretical parametric GLM from Bayesian model selection between fMRI models in an independent cohort of healthy young and older adults 28 , as well as in the HC, SCD and AD-rel groups from the DELCODE study 38 .…”

“…One reason for this may be that poor episodic memory in AD, and to some extent in MCI, reduces the signal-to-noise ratio of encoding-specific fMRI responses, making it difficult to differentiate between subsequently remembered and forgotten items. Compatibly, we have recently shown that, when comparing first-level fMRI models using Bayesian model selection, memory-invariant fMRI models provide a better fit than subsequent memory models in individuals with MCI or mild AD dementia 38 .…”