Cohort
701 participants were recruited in the PREVENT-Dementia study from five study sites: West London, Edinburgh, Cambridge, Oxford and Dublin. The main entry criteria were age between 40 and 59 and absence of dementia or other neurological disorders. The primary risk stratification approach was dementia family history defined by having one or both parents with dementia (50–50 recruitment target for those with and without FHD). Participant APOE genotype analysis was carried out on QuantStudio12K Flex to establish APOE variants. APOE information was not collected for five participants. The CAIDE score was calculated based on published thresholds on age, education, blood pressure, activity, cholesterol, body mass index and did not incorporate APOE4 genotype; in particular, we have used model 1 from Kivipelto et al. [9]. Overall, 23 participants had missing information for CAIDE calculation.
A detailed description of all the data acquired as part of the study can be found in [29, 30]. From the 701 recruited participants, 647 had an MRI scanning session at baseline (Fig. 1). 27 scans were excluded from analysis after visual inspection either due to poor quality, artifacts (e.g. excessive motion, poor contrast) or incidental findings (e.g. meningiomas). An estimated years to dementia onset (EYO) variable was calculated for participants with dementia family history, based on the difference between the age of the participant and the age of parental dementia diagnosis (if both parents had dementia, the younger onset was used) and had a mean of 22.8 years.
MRI protocol
As part of the PREVENT-Dementia MRI protocol, a T1-weighted magnetization prepared rapid gradient echo (MPRAGE) (repetition time = 2.3 s, echo time = 2.98 ms, 160 slices, flip angle = 9°, voxel size = 1 mm3 isotropic) and a T2-weighted hippocampal (repetition time = 6.42 s, echo time = 11 ms, 20 slices, flip angle = 160°, voxel size = 0.4 × 0.4 × 2.0 mm) scan were acquired. All study sites used 3T Siemens scanners and in particular, the following models: Prisma (Oxford, Edinburgh), Prisma fit (Cambridge), Verio (West London, Edinburgh) and Skyra (Dublin; Edinburgh). All scans were corrected for field inhomogeneities using the Advanced Normalisation Toolbox (ANTs) N4 algorithm [31].
Surface-based analysis
Freesurfer version 7.1.0 was used for data processing [32]. The recon-all pipeline was run for every subject with standard settings. The brain masks and surfaces were inspected following recon-all and manual corrections were applied: (a) in the form of erosion of non-brain voxels from the brain mask or non-WM voxels from the WM mask, (b) in the form of filling of areas where the brain was not correctly identified or (c) with the addition of control points in cases where white matter was not successfully identified. Manual corrections were applied for the majority of subjects (87%). We quantified cortical thickness in a vertex wise level and in 68 regions based on the Desikan-Killiany atlas [33]. Furthermore, the volume of bilateral hippocampi, thalami, amygdala, putamen, caudate, accumbens and pallidum were quantified. Values from the left and the right hemisphere were averaged for cortical thickness and added up for volumetric analysis.
Hippocampal subfield segmentation
A dedicated Freesurfer hippocampal subfields segmentation module was applied using information from both the T1 and T2-weighted images. Hippocampal subfield segmentation relies on an atlas constructed based on in-vivo and high-resolution post-mortem scans and on manual subfield delineation from experienced radiographers [34]. In particular, the Freesurfer algorithm utilizes the constructed probabilistic atlas and the individual voxel intensities to proceed to subfield segmentation using Bayesian inference [34].
Following previous methodology, we concatenated subfields in the following: hippocampal fissure, hippocampal tail, subiculum (subiculum + presubiculum + parasubiculum), CA4/DG (CA4 + GC_ML_DG), CA1, molecular layer (ML) and CA3 [28]. We have also combined the left and the right hemispheres.
The molecular layer is a structure of particular interest since it covers the CA1-SRLM area, a region rich in synapses from CA3 and the entorhinal cortex to CA1 [35]. CA1-SRLM is known to be susceptible to tau accumulation in early disease stages [17, 36]. In typically acquired T1-weighted image, it is not discernible, however, when tailored T2-weighted acquisitions are applied, it appears as a dark band due to its myelin content [37]. It has been shown that its clarity reduces in Alzheimer’s disease [38], hence we proceeded in the evaluation of its clarity in the cohort. Toward that end, the T2w images were quality checked and a visual rating on the clarity of the molecular layer was recorded ranging from 1 to 3. The dataset was split in three and was rated by three independent raters (MED, CN, NJ). Agreement was evaluated using Cohen’s κ for 60 scans. For each hemisphere starting from the slice where the body of the hippocampus was dominant and for another two slices posteriorly, we rated clarity per slice and also overall clarity per hemisphere, as follows: 1—non-clear delineation less than 20% can be seen; 2—more than 20% can be seen but not perfectly discernible and 3—clearly seen. The starting slice per hemisphere, was defined by registration of the head-body-tail segmentations generated by the Freesurfer pipeline to the T2 space using the linear transform generated as part of the main pipeline. The pipeline and example ratings are shown in Fig. 2. For this part of the analysis, a further 60 subjects were excluded due to absence of a high-resolution T2 acquisition, poor T2 image quality or due to issues with between-modality registration.
Statistical methodology
Statistical analysis was conducted in Matlab 2021b (R2021b, The MathWorks Inc., Natick, MA, USA). Wilcoxon rank sum test and χ2 test were used for between-group comparisons of demographic factors.
The ComBat harmonization algorithm was used for parametric adjustment of the structural measures to account for inter-site differences [39]. This method has been shown to remove site-differences when preserving the relevance of covariates of interest [39, 40]. For the present implementation, we explicitly included as modulators in ComBat: APOE4, FHD, years of education, age and sex. Data from 6 participants were excluded due to missing information (5 APOE; 1 education).
Linear regression models were used to investigate independently the association of structural measures with FHD and APOE4 genotype. Age, sex and years of education were added as predictors to the models. When volumetric measures were examined, eTIV was also included as a predictor. We additionally investigated interactions of APOE4 and FHD with age. Further sub-analysis was conducted to investigate the impact of carrying one or two copies of the APOE4 gene by creating a three-group variable (no_APOE4, APOE4_1, APOE4_2).
CAIDE score is a discrete variable ranging from 0 to 15 and was non-normally distributed in our cohort. It incorporates all risk factors which would be used normally as analysis covariates (e.g. age, sex, education), hence, to investigate the association of CAIDE with cortical thickness, Spearman correlations were run between the CAIDE and regional thickness. To examine associations with volumetric measures partial Spearman correlations controlling for eTIV were used.
For the volumetric comparisons as well as the cortical thickness analysis within the 34 regions of the Desikan-Killiany atlas, the false discovery rate (FDR) method was used to correct for multiple comparisons [41]. Normality, autocorrelation and homoskedasticity of standardized residuals were checked with the Kolmogorov–Smirnov, Ljung-Box Q-test and Engle test, respectively.
The molecular layer clarity was investigated in relation to age, sex, education, APOE4 genotype and FHD in a single linear regression model.