Alzheimer's disease (AD) is characterized by an extended pre-dementia phase, during which patients progress from an absence of observable cognitive impairment to a subjective perception of decline (referred to as subjective cognitive decline; SCD), followed by mild cognitive deficits (referred to as mild cognitive impairment; MCI; Albert et al., 2011, 2013; Sperling et al., 2011). Many structural and functional brain changes occur during the pre-dementia phase of AD (Clément & Belleville, 2010, 2012; Corriveau-Lecavalier, Duchesne, et al., 2021; Clément et al., 2013; Corriveau-Lecavalier et al., 2019; Pasquini et al., 2019; Sperling et al., 2010, 2011). Understanding these changes is crucial for a comprehensive understanding of the disease.

Previous studies have reported hyperactivation in individuals with SCD or MCI, who exhibited higher task-related activation compared to cognitively healthy controls (HCs). Hyperactivation has been reported during working memory and divided attention tasks for MCI (Clément et al., 2013) and during episodic memory tasks for SCD and MCI (Clément & Belleville, 2010, 2012; Corriveau-Lecavalier, Duchesne, et al., 2021). Conversely, older adults with late MCI and dementia show hypoactivation, that is lower levels of activation than HCs (Celone et al., 2006; Clément & Belleville, 2010, 2012; Clément et al., 2013). When considering the combined activation results in persons with SCD, MCI and dementia, the trajectory of brain activation clearly follows an inverse U-shape function. This trajectory begins with a very early increase, possibly during the SCD and early MCI phases, and subsequently transitions to a decrease as the disease progresses, possibly in the late MCI and dementia stages. Thus, the relationship between activation and markers of disease severity across the entire preclinical and clinical continuum of AD should be modeled using an inverse U-shape function (see Fig. 1a). Linear functions that reflect either the ascending or descending portion of the inverse U-shape function may also be observed when examining a narrower part of the continuum. In MCI, hyperactivation has been reported in AD signature regions that is, regions that are susceptible to early neuronal changes associated with AD (Billette et al., 2022; Dickerson et al., 2008; Maass et al., 2019; Savva et al., 2009). Both inverse U-shape (observed in left superior parietal lobe activation) and linear positive (observed in right inferior temporal activation) functions have been observed in individuals with MCI and in those with SCD who possess an increased risk of AD (reduced hippocampal volumes and/or the presence of an apolipoprotein E (APOE) ε4 allele or SCD+; Corriveau-Lecavalier, Duchesne, et al., 2021).

While the existing literature offers some evidence for the presence of a linear or inverse U-shape relationship between activation and disease severity in the pre-dementia stage of AD, this pattern has predominantly been examined using standard regression models. Standard regression models assume a normal distribution and provide limited insights into a broader distribution of the variable of interest. This limitation may be critical in the context of pre-dementia. Examining activation patterns in this stage of AD is inherently complex due to the substantial interindividual variability that characterizes this early stage. Participants may be at different stages of disease progression, their symptoms may reflect a range of different biological processes, and not all participants with MCI and SCD+ will progress to dementia. Since the literature reports both hyper- and hypoactivation, it is likely that the expected inverse U-shape or linear relationship between activation and disease severity will be observed for higher and lower levels of activation. For these reasons, it is particularly interesting to test whether the relationship between activation and disease severity proxies vary when examined at different levels of the activation distribution.

This can be examined using quantile regression analysis, which is an extension of standard linear regression. Quantile regression is a compelling approach to examine potentially variable populations, particularly when the effect of a predictor is expected to vary as a function of the distribution of scores. Quantile regression (Koenker & Bassett, 1978) has been increasingly used in the field of health and neuroimaging. For example, quantile regression has been used to determine that the status of AD medication and medication consumed daily predict medication adherence in participants who are poorly adherent (lower quantile of the distribution), but not in those who adhere well to their medication protocols (higher quantile of the distribution) (Borah & Basu, 2013). This illustrates that examining relationships between variables in extreme portions of the distribution may identify a more meaningful pattern of association than traditional regression approaches.

In our study, quantile regression was used to assess how the estimates of the conditional quantiles of activation vary as a function of disease severity predictors (for instance, hippocampal volume, cortical thickness, and associative memory). The relationship between activation and proxies of disease severity was expected to be stronger in participants with abnormal levels of activation, either in the higher or lower quantile range. Thus, this approach can be used to assess cerebral activation, while considering the distribution of individual activation levels in the pre-dementia stage of AD. Finally, quantiles have the advantage of providing threshold values to identify higher and lower activation levels that are based on the distribution of the parameter rather than arbitrarily determined.

The goal of this study was to use quantile regression modeling to determine the activation pattern as a function of overall activation level in a group of individuals with either SCD+ or MCI. We assessed whether the relationship between activation and proxies of disease severity varied with activation levels and whether it differed in the higher (>75th) and lower (<25th) quantile of the activation distribution relative to the median level. We used the same proxies as those identified by Corriveau-Lecavalier, Duchesne, et al. (2021), which included hippocampal volume, cortical thickness, and associative memory performance. We also examined whether participants in different quantiles differed in terms of their cognitive profile. We expected that participants in the lower quantile would be more impaired than those in the higher quantile and include more participants with MCI than SCD+. Activation was assessed during an associative object-location task, which is known to be impaired early in the disease (Corriveau-Lecavalier, Duchesne, et al., 2021; Corriveau-Lecavalier, Rajah, et al., 2021; Corriveau-Lecavalier et al., 2022; Rabipour et al., 2020, 2021; Rajah et al., 2010; Subramaniapillai et al., 2021). Initial analyses focused on the left hippocampus, right inferior temporal lobe, and left superior parietal lobule regions. These regions were selected because they had previously exhibited hyperactivation associated with proxies of disease severity in a prior study (Corriveau-Lecavalier, Duchesne, et al., 2021; Corriveau-Lecavalier, Rajah, et al., 2021). These analyses focused on regions linked with markers of disease progression as these are more hyperactivation relevant to the disease rather than to other factors such as reserve differences, task difficulty, etc. Subsequently, we conducted exploratory quantile regression within all AD signature regions exhibiting hyperactivation. This approach was adopted due to the anticipation that using quantile regression might reveal associations between activation and proxies of disease severity that had not been previously observed using standard regression. Should activation differences with disease severity exclusively manifest in the extreme ends of the distribution, quantile regression holds the potential to unveil regions that have not been previously observed. Consequently, the present study builds upon its predecessors by employing quantile regression to examine the relationship between brain activation and proxies of disease severity and by exploring the complete AD signature regions using quantile regression.

Activation variation was modeled against three proxies of disease severity: left hippocampal volume, cortical thickness, and associative memory performance. As the mean of activation is not homogeneous, the estimated function was expected to vary based on the quantiles under consideration. Consequently, the heterogeneity of mean values might result in distinct effects of proxies of disease severity depending on the activation levels. We therefore expected that the relationship between activation and proxies of disease severity would vary as a function of the activation quantile (see Fig. 1b for a summary of the expected function between proxies of disease severity and activation as a function of quantile). In particular, we hypothesized that participants in the higher and lower quantiles of activation would show a larger linear or inverse U-shape relationship between activation and indicators of disease severity. We also expected that the quantile regression would reveal hyperactivation in other regions that are part of the AD signature.