Alzheimer resemblance atrophy index, BrainAGE, and normal pressure hydrocephalus score in the prediction of subtle cognitive decline: added value compared to existing MR imaging markers

Using magnetic resonance (MR) neuroimaging to predict subsequent cognitive decline is a main issue in Alzheimer’s disease (AD) research. Several neuroimaging markers have been proposed in clinical routine, including hippocampal atrophy assessed using the semiquantitative visual reading mesio-temporal atrophy (MTA) score [1], T2w hyperintense white matter lesions (WML) using the Fazekas score [2], and number of cortical microbleeds (CMB) in community- and hospital-based cohorts. MTL atrophy allowed for identifying patients at high risk for Alzheimer type dementia among those with minor cognitive impairment (for review, see [3]). In the same line, longitudinal studies suggested a negative effect of WMH presence and progression on general intelligence, attention, and executive functions in non-demented elders [4‐6]. Three population-based studies (Rotterdam [7], Framingham Heart [8], and AGES-Reyjkavik [9]) also supported the deleterious effect of CMB on cognition in non-demented elders.

While the diagnostic performance of these classic MR markers is well-established in patients with already present cognitive decline, their value in predicting subtle cognitive decline in healthy elderly individuals remains a matter of debate [10]. MTA, cortical microbleeds, and WML have all been thought to impact cognitive performance at the pre-mild cognitive impairment (MCI) state [11‐16]. However, negative or ambiguous data were also frequently reported [17‐19]. This raises the question of whether more recently proposed artificial intelligence (AI) MR imaging markers might have an added value in predicting subsequent cognitive decline in healthy elderly controls. The Alzheimer Disease Resemblance Atrophy Index (AD-RAI) is an AI-based metric derived from 3DT1 volumetric brain MRI [20‐23]. In the meantime, the concept has also been adopted by commercial vendors. This operator-independent AD-RAI encapsulates hippocampal atrophy as well as atrophy of the established AD signature regions [24]. In a totally different perspective, the Brain Age Gap Estimate (BrainAGE) score is based on a machine learning regression task and has been proposed to capture global deviations from normal brain aging [25‐27]. In a first step, associations between brain MRI and patient age are learned by an algorithm and then applied to new individuals in order to predict their age based on his or her brain MRI (usually 3D T1-weighted scans serve as input). Contrasting predicted and chronological brain age can either result in expected (predicted age ~ chronological age), advanced (predicted age > chronological age), or delayed (predicted age < chronological age) brain aging [25]. Both parameters are defined based on structural brain MRI atrophy patterns. The normal pressure hydrocephalus NPH index is another AI calculated marker derived from volumetric 3D T1 images, which indicates the degree of an NPH configuration in elderly cohorts [28]. Definite full (NPH) fulfilling diagnostic and clinical criteria remains rare in cognitively preserved elderly individuals. In the context of healthy elderly individuals, we consider NPH configuration as a spectrum, rather than a binary variable. Some participants may display beginning imaging features of NPH configuration, without fulfilling the clinical diagnosis of NPH. Considering currently emerging theories, impairment of CSF flow (resulting in diminished clearance) may be another factor contributing to cognitive decline, which might confound clinical and biomarker interpretation in AD [29, 30].

The main purpose of the present study is to assess the added value of the AI-based parameters AD-RAI, BrainAGE, and NPH index, as compared to the established visual imaging markers (MTA, Fazekas, and number of CMB), in the prediction of early cognitive decline in healthy elderly individuals. We had the opportunity to investigate these imaging parameters at baseline in a community-dwelling sample of 80 cognitively intact elderly individuals who were followed up longitudinally during a 4.5-year period.

The present data reveal that two among the newly proposed AI MRI markers have a limited but still significant added value in predicting the cognitive trajectory in elderly controls. Although they are not related to global measures of cognition such as the MMSE score, AD-RAI index, and BrainAGE score are associated with the decrement of visuospatial and executive function measures in cognitively preserved elders. Of importance, both markers have a significant added value when compared to widely used MRI markers of brain aging such as the Fazekas score, MTA, and number of CMB. For the NPH-score, no significant association with neuropsychological performances in old age was found.

Three recent studies explored the relevance of AD-RAI as a diagnostic marker for AD [23], for predicting the MCI conversion to AD [44], and for the detection of preclinical and prodromal AD [45]. In a series of 50 AD patients and 50 controls, the AD-RAI use led to a diagnostic accuracy of 0.91 outperforming the volume of any single brain structure measured [23]. An AD-RAI cut-off of 0.5 had a better performance to differentiate MCI converters from non-converters [44]. Moreover, this parameter achieved the best metrics (compared to conventional MRI markers) in detecting elderly controls positive for amyloid and tau (A+T+) [45]. Our results go beyond these observations by showing that AD-RAI is an independent predictor of visuospatial memory decline in the present cohort. MTA, Fazekas score, and CMB were all unable to predict such subtle changes in cognitive performances in these elderly controls. Both likelihood-ratio test and receiver operating characteristic (ROC) curves demonstrated that the addition of AD-RAI significantly increases the ability to predict the decline of visuospatial memory. Of importance, even in these cases with low amyloid burden, the combination of amyloid positivity and MRI markers including AD-RAI provides the best performance in predicting the cognitive outcome.

In the last decade, BrainAGE was established as an image-based biomarker to capture the general aging processes of the brain [40]. It was shown to be altered in various neuropsychiatric conditions, most prominently in age-associated disorders such as Alzheimer’s disease (AD) [27]. Moreover, BrainAGE was able to predict the risk of incident dementia in a cohort with mild cognitive impairment [46] and in the general population [47]. In this latter study, the incidence of dementia was related to the prediction of brain age based on MRI-derived gray matter in 3688 dementia-free cases from the Rotterdam study. This association persisted in logistic regression models adjusted for white matter hyperintensities and hippocampal volume. Our data are less impressive since the BrainAGE score was not related to the evolution of global cognition upon 4.5-year follow-up. However, and unlike conventional MRI markers, BrainAGE was an independent predictor of declining performance in executive functions (Trail Making Test B). This imaging parameter was an independent predictor of executive function performance decrease after adjusting for Fazekas score, CMB, and MTA at baseline. In particular, likelihood ratio test revealed that the addition of Brain Age scores significantly improves the percentage of cognitive variability explained by the regression models. This was also the case for amyloid positivity. However, only the model that included all conventional MRI markers and Brain Age was statistically significant in terms of area under the ROC indicating that the contribution of BrainAge as a single imaging marker in the prediction of executive function decline in our cohort remains modest.

Several reasons may explain the discrepancy between the present data and the recent observations by Wang et al [47]. First, none of our cases evolved to MCI during the follow-up period pointing to the presence of cognitive resilience in our sample without significant vascular burden. The cognitive outcome was defined as the binary change (decrement versus improvement or stability) in neuropsychological scores without reference to incident dementia. Furthermore, since the BrainAGE score used in our study—like most other approaches—is based on T1-weighted imaging data only, it can be assumed that age-related changes are predominantly found in other sequences (such as white matter lesions as assessed on FLAIR imaging) might be neglected. In addition, BrainAGE is also influenced substantially by parameters such as lifestyle factors (e.g. smoking [48]), genotypic variations (e.g. APOE status) [46], or early neurodevelopmental influences such as preterm birth [49] that cannot be taken into account given the limited sample of this study. Despite these differences, the present findings imply that the BrainAGE score remains a useful predictor of longitudinal changes in executive functions during the very early stages of the aging process.

Recent studies suggest that imaging markers of NPH, which imply cerebrospinal fluid disorders, might be an independent imaging markers of non-AD pathophysiology [29]. The NPH index is an operator-independent imaging marker of possible cerebrospinal fluid disorder that has been proposed as a possible candidate for the prediction of cognitively asymptomatic cases at risk for AD [28]. In our cohort, we could not demonstrate significant associations between NPH index at baseline and both global cognition and neuropsychological test changes—taking into account as a limitation that NPH index values were very low in our cohort. The NPH index might provide more valuable results in other cohorts with a higher proportion of cases with elevated NPH index.

Among the strengths of the present study, one should note the detailed cognitive analysis and use of multivariable models that allow for defining the added value of the newly proposed AI imaging parameters compared to traditional imaging markers of brain aging. Several limitations should be considered. The small sample size limits the number of imaging predictors that can be simultaneously included in regression models, we excluded all of the cases with significant vascular burden in order to focus on the AD trajectory in brain aging. Thus, our cases are not representative of the whole spectrum of brain aging. Despite their recruitment in the community, they display no or very mild vascular pathology and a relatively high level of education. The assessment of white matter lesions was made using T2w sequences and not the traditional FLAIR images. However, it is quite rare that e.g. Fazekas 1 in FLAIR is considered Fazekas 2 in T2w (or the inverse) by a well-trained radiologist. Most importantly, the low occurrence of amyloid positivity and brain hypometabolism in this sample indicated that the number of incipient AD cases may be lower than expected in a community-based sample of elderly persons. It is thus likely that the added value of AI measures may increase when cases with a higher amyloid burden are considered. Based on our observations, future studies in mixed samples are clearly warranted to define the relevance of the AI-MRI measures in the field of brain aging and AD.

The AI-based AD-RAI index and BrainAGE scores have limited but significant added value in predicting the subsequent cognitive decline in elderly controls when compared to the established visual MRI markers of brain aging, notably MTA, Fazekas score, and number of CMBs.