All participants were recruited from the greater Omaha area as part of a multimodal imaging study of neural dynamics in AD (the dynamic mapping of Alzheimer’s disease pathology [DMAP] study; see Wiesman et al., 2021, 2022). This investigation was reviewed and approved by the Institutional Review Board at the University of Nebraska Medical Center. A detailed description of the study was provided to all participants, including informants for participants on the ADS. Individuals on the ADS whose capacity to consent was questionable provided written informed assent accompanied by written consent from a legally-authorized representative. All other participants provided written informed consent. Informants for each participant in the ADS group (regardless of capacity) were also consented for collection of additional data and completion of questionnaires.

Forty-four participants (n = 44) with amnestic MCI (aMCI) or mild probable AD, and a group of cognitively-normal healthy older adults (n = 20) were enrolled in this study. Participants with amnestic complaints were recruited from local Memory Disorders and Geriatrics Clinics and categorized as aMCI or probable AD by a fellowship-trained neurologist specializing in memory disorders. To confirm biomarker status, participants underwent amyloid PET imaging. All scans were read by a fellowship-trained neuroradiologist blinded to group assignment and assessed as being “amyloid-positive” or “amyloid-negative” using established clinical criteria. For all participants, exclusion criteria included neurological or psychiatric disorders (other than aMCI/AD), history of head trauma, moderate or severe depression (Geriatric Depression Scale ≥10), current substance abuse, or incomplete data.

Data from 18 participants were excluded due to the following: COVID-related withdrawal (ADS, n = 1), major incidental finding (ADS, n = 1), amyloid-negativity in cognitively impaired group (n = 4), MRI (T1 or TSE) scan quality (ADS, n = 9; Controls, n = 2; see below for more detail), no amyloid status (Controls, n = 1). The final sample for ADS included 29 amyloid-positive participants (aMCI = 12; AD = 17), and 17 amyloid-negative for the cognitively normal control group.

The neuropsychological data used in the present study is similar to previous reports (Wiesman et al., 2021, 2022). Briefly, participants completed the Functional Activities Questionnaire (FAQ), the Montreal Cognitive Assessment (MoCA) and the Mini-mental State Examination (MMSE; Folstein et al., 1975; Pfeffer et al., 1982; Nasreddine et al., 2005). All participants also underwent a battery of neuropsychological tests that assessed the following domains: attention and executive function (Wechsler Adult Intelligence Scale [WAIS-IV] Digit Span Forward, Backward, and Sequencing; Trail Making Test Part B), language (Boston Naming Test; Controlled Oral Word Association Test/Phonemic Verbal Fluency; Animals/Semantic Verbal Fluency), processing speed (WAIS-IV Digit Symbol Coding; Trail Making Test Part A; Heaton et al., 2004; Wechsler, 2008), immediate recollection (Wechsler Memory Scale Fourth Edition [WMS-IV] Logical Memory I Immediate Recall; Hopkins Verbal Learning Test-Revised [HVLT-R] Learning Trials 1–3), and delayed recollection (WMS-IV Logical Memory II Delayed Recall and Recognition; HVLT-R Delayed Recall and Recognition Discriminability Index; Brandt and Benedict, 2001; Wechsler, 2008).

Raw scores for each participant were converted to demographically-adjusted z-scores (e.g., based on age, education, etc.) using published normative data and following standard practice procedures (Brandt and Benedict, 2001; Heaton et al., 2004; Wechsler, 2008, 2009). Demographically corrected z-scores based on test-specific normative data were then averaged across tests to create composite cognitive domain z-scores by participant. From this battery, the neuropsychological tests which have been shown to measure processes commonly attributed to hippocampal functions (i.e., immediate and delayed recollection; Atienza et al., 2011; Fouquet et al., 2012; Carlesimo et al., 2015; Broadhouse et al., 2019; Ogawa et al., 2019; Zhao et al., 2019; Nurdal et al., 2020) and those which assess general cognitive status (i.e., MMSE and MoCA) were considered in statistical analyses. Note that immediate and delayed recollection were modeled as separate cognitive domains: while they share considerable variance, immediate vs. delayed learning and memory represent functions with important conceptual distinctions. Descriptive statistics for all neuropsychological measures are reported in Table 1.

The florbetapir-PET data acquisition and preprocessing in the present study was similar to previous reports (Wiesman et al., 2021, 2022). All scans were read by a fellowship-trained neuroradiologist blinded to group assignment and assessed as being “amyloid-positive” or “amyloid-negative” using established clinical criteria. Briefly, 18F-florbetapir (Amyvid^™, Eli Lilly) PET data were collected on a GE Discovery MI digital scanner (Waukesha, WI), reconstructed, body-weight normalized, and MNI-normalized in MIMNeuro (Joshi et al., 2012; Minoshima et al., 2016). Data were then normalized to the crus of the cerebellum (SUIT template), back-transformed into each patient’s native structural MRI space (mri_vol2vol), smoothed to a common resolution (8 mm FWHM), projected onto native surfaces (mri_vol2surf; maximum value; projection fraction = 1; steps of 2), and then normalized to FSAverage template surface using FreeSurfer (Diedrichsen, 2006; Fischl, 2012; Landau et al., 2012; Guo et al., 2020). Surface masks from seven target regions of interest of the Desikan-Killiany Atlas (parahippocampal, entorhinal, inferior and superior parietal, precuneus, lateral occipital, and superior frontal; Joshi et al., 2012; Landau et al., 2012) were used to compute mean standardized uptake value ratios (SUVr) of amyloid deposition per patient on the ADS.

During data acquisition, participants were in constant contact with research personnel through real-time audiovisual monitoring. Structural MRI data were collected on a 3 T Siemens Prisma System with a 64-channel head coil, and included a medial temporal lobe focused T2-weighted turbo spin echo (TSE) image [TR = 7.79 s, TE = 66 ms, flip angle = 145°, FOV = 170 mm, in plane resolution = 0.4 × 0.4 mm, slice thickness = 2 mm, slices = 32] and a whole-head T1-weighted three-dimensional 1 mm isotropic MPRAGE sequence [TR = 2.3 s, TE = 2.98 ms, flip angle = 9°, FOV = 256 mm], per best modern practices (Yushkevich et al., 2015; Mueller et al., 2018; Olsen et al., 2019; Wisse et al., 2021).

MRI Quality Control (MRIQC) v0.16.1 (Esteban et al., 2017) was used as an initial check of MRI data quality, and these preliminary automated ratings were supplemented with systematic quality assessment by a trained rater. Assessment included a review of hippocampal coverage/completeness, contrast/noise, and motion artifacts. Data from eight participants were excluded due to motion artifacts in the T2 TSE images. MRI scans that survived quality assessment of the raw data were then processed through the segmentation workflow. High resolution hippocampal subfield volumes were quantified with the Automatic Segmentation of Hippocampal Subfields (ASHS) software and the UPENN atlas consisting of scans of MCI individuals and older adults (Yushkevich et al., 2015). Briefly, the dedicated T2-weighted image of the hippocampi was coregistered to the routine T1-weighted image, segmented, and bias-corrected. Next, hippocampal subfields were parceled from the medial temporal lobe via multi-atlas joint-label fusion with multiple pairwise registrations, consensus segmentation, learning-based error correction, and bootstrapping. This method was chosen as it has been shown to perform with similar or better accuracy as compared to other methods employed for work with similar clinical populations (de Flores et al., 2015; Yushkevich et al., 2015; Giuliano et al., 2017; Mueller et al., 2018; Olsen et al., 2019). From the resulting parcellation, we used the following volumes for our analyses: CA1, CA2, CA3, dentate gyrus, and subiculum (Figure 1A). Data from 3 participants were excluded at this stage due to segmentation errors. All hippocampal ROI’s were reviewed by a trained rater for coverage/completeness. As a result, 46 (ADS = 29, Controls = 17) images survived quality control and/or segmentation procedures, and were used for all subsequent analyses, and 11 (ADS = 9, Controls = 2) images were excluded. Importantly, participants included in our final sample after imaging-based exclusion did not significantly differ in terms of demographics (i.e., Age, Education, Sex) or cognition (i.e., MoCA, MMSE) from participants who were excluded at this MRI processing stage (Supplementary Table S1). T1-weighted images for all subjects that passed quality control were also processed through the recon-all workflow of FreeSurfer (v7.1.1; Fischl, 2012) which produced segmentations of whole-hippocampal volumes and derivations of global cortical thickness. These whole-hippocampal segmentations were reviewed by a trained rater following the same procedures as described above for subfields.

To estimate the extent of disease-related atrophy and to control for inter-individual variation in brain size, individualized intracranial volume estimates (ICV) were regressed from the hippocampal subfield volume estimates resulting in ICV-adjusted hippocampal subfield volumes (Buckner et al., 2004; Raz et al., 2004; Voevodskaya et al., 2014; Pintzka et al., 2015). Moreover, all analyses considered age, sex, and education as potential covariates to minimize any potential confounding effects on outcome measures (Mueller and Weiner, 2009; La Joie et al., 2010; Piras et al., 2011; Apostolova et al., 2012; Noble et al., 2012; Farfel et al., 2013; Malykhin et al., 2017; Tang et al., 2017; O’Shea et al., 2018; van Eijk et al., 2020; Veldsman et al., 2021). Inclusion of these potential additional covariates (i.e., age, sex, education) was determined per each model by comparing models with successive inclusion of potential covariates (i.e., model 1: age, model 2: age + sex, model 3: age + sex + education) and selecting the model that was best at explaining outcome variance, using the anova function in R. Models including covariates that significantly improved prediction vs. the no-covariate base model (p < 0.05) were selected. Models that included any of these additional potential covariates are listed in Supplementary Table S2. All models were also inspected for normality with plots of the residuals against fitted values and using the Shapiro–Wilk test. When the assumption of normality of residuals was violated, instead of parametric linear models we employed a nonparametric Wilcoxon test or nonparametric permutation testing to derive inferential statistics (wilcox.test and lmPerm in R). The models that required nonparametric tests are indicated in Supplementary Table S2.

We first computed a series of general linear models to examine any group differences in whole hippocampal and hippocampal subfield volumes. Next, we sought to determine how degeneration of the hippocampus relates to impairments in immediate and delayed recollection, and general cognitive status, in participants on the ADS. To quantify hippocampal degeneration in participants on the ADS relative to controls, we first created hippocampal volume profiles by vectorizing the left and right hemisphere volumes for each subfield for each participant. We then computed a Euclidean distance matrix between all combinations of patient (N) and control (M; Figure 1B), resulting in an NxM matrix of distance metrics that represents the dissimilarity between each ADS participant’s subfield profile and every control’s subfield profile (Figure 1C). This multidimensional approach is preferable to simpler metrics in that it allows for the simultaneous consideration of several subfields from right and left hemispheric volumes to ensure asymmetric hemispheric variability is also captured (Maruszak and Thuret, 2014; Sarica et al., 2018; Ogawa et al., 2019; Zhao et al., 2019; Huang et al., 2020). For the analyses in this study, these Euclidean distance values were averaged across all controls per each participant on the ADS, thereby representing the mean dissimilarity of each ADS participant hippocampal subfield profile from controls across hemispheres, termed hippocampal degeneration score (HiDs; Figure 1D). We then used these HiDs to test how decrements across hemispheres were associated with cognitive domain scores using partial correlations. We focused these analyses on the MMSE and MoCA, and the composite domain scores representing learning and memory as prior studies have found that these cognitive measures are most robustly linked to hippocampal volumes (Atienza et al., 2011; Fouquet et al., 2012; Carlesimo et al., 2015; Broadhouse et al., 2019; Ogawa et al., 2019; Zhao et al., 2019; Nurdal et al., 2020).

Extant research suggests there is more to be gained than lost by demarcating the hippocampus into subregions. For example, Wolk et al. (2017) have argued that global hippocampal measures obscure critical information at early stages of disease progression, and others have shown measures of subfields outperform whole hippocampus measures to discriminate controls from individuals on the ADS (Mueller et al., 2010; Maruszak and Thuret, 2014). To understand whether the degeneration of individual subfields was associated with cognitive impairments, we followed the same HiD procedure as above to quantify degeneration, but separately for each respective subfield using the vectorized left and right hemisphere values. Moreover, to determine whether any of the reported associations between degeneration of subfields and cognition could be explained by non-specific cortical tissue degeneration, we also computed models substituting global cortical thickness estimates from FreeSurfer (averaged over all cortical locations per participant) in the place of HiDs. Finally, we examined whether cortical amyloid deposition was associated with degeneration of hippocampal subfields, or moderated the relationships between subfield HiDs and cognition.

Euclidean distance matrices were computed using the scikit-learn package in Python (Pedregosa et al., 2011). We report all p-values adjusted for multiple comparisons correction. Multiple comparisons were corrected by controlling the false discovery rate (FDR) using the Benjamini-Hochberg method with the p.adjust function from the stats package as implemented in R (Benjamini and Hochberg, 1995).

Forty-six participants were included in all statistical analyses (ADS, 29; Controls, 17). Demographic information for the final sample is provided in Table 1. Also included in Table 1 are inferential statistics for group comparisons of baseline demographics, however strong interpretations of these inferential significance tests are ill-advised (De Boer et al., 2015).

The groups were matched on age, sex, and education. Clinical assessments (e.g., MMSE, MOCA), cortical thickness, and whole hippocampal volume data of our ADS group were normalized to the mean and standard deviation of comparable data from the cognitively normal control group and were paired with the neuropsychological data. By computing z-scored deviations from healthy participants of these measures for all participants in the ADS group, we can visualize changes in different neuropsychological and neuroanatomical measures together in the same plot in parallel with amyloid deposition (Figure 2).

Here, subplots shifted further to the right represent measures which deviate more from controls. Participants on the ADS performed significantly worse than controls across all cognitive domains (Wiesman et al., 2021, 2022). Statistical results for all neuropsychological measures are reported in Table 1, and comparative sample distributions of controls and ADS groups for all neuropsychological data are shown in Supplementary Figure S1.

We tested for differences in whole hippocampal volumes between individuals on the ADS and healthy control participants. These models revealed significant group differences in left (F(1,44) = 34.98, 95% CI [−1015.58, −499.13], p_adj < 0.001) and right (F(1,44) = 17.49, 95% CI [−836.02, −292.14], p_adj < 0.001) hippocampus, such that individuals on the ADS exhibited reduced whole hippocampal volumes relative to controls. We then conducted a post-hoc test to examine whether the observed effects in whole hippocampal volume differed between ADS subgroups (i.e., AD vs. MCI). No significant differences were observed between ADS subgroups in either right or left whole hippocampus volumes (Supplementary Table S3).

To evaluate the degeneration of hippocampal subfields among individuals on the ADS, we computed ANCOVAs for group differences in each of the five subfields bilaterally (e.g., left and right CA1, CA2, CA3, dentate gyrus, and subiculum). As shown in Figure 3, these models revealed significant group differences in left (F(1,44) = 33.29, 95% CI [−378.89, −182.64], p_adj < 0.001) and right (F(1,44) = 7.40, 95% CI [−273.99, −40.82], p_adj = 0.018) CA1, left (F(1,43) = 26.27, 95% CI [−256.09, −111.47], p_adj < 0.001) and right (F(1,44) = 6.56, 95% CI [−182.81, −21.79], p_adj = 0.023) dentate gyrus, and left (F(1,43) = 14.90, 95% CI [−109.34, −34.30], p_adj = 0.001) and right (F(1,43) = 10.34, 95% CI [−94.15, −21.58], p_adj = 0.006) subiculum. In bilateral CA1, dentate gyrus, and subiculum, these differences were such that individuals on the ADS exhibited reduced hippocampal volumes relative to controls. No group differences were observed in either left (F(1,44) = 0.33, 95% CI [−2.49, 4.49], p_adj = 0.630) or right (F(1,44) = 0.07, 95% CI [−3.75, 2.85], p_adj = 0.782) CA2, nor in left (F(1,44) = 3.84, 95% CI [−0.27, 19.46], p_adj = 0.080) or right (F(1,44) = 0.85, 95% CI [−8.40, 22.57], p_adj = 0.327) CA3. We then conducted a series of post-hoc tests to examine whether the observed effects in the CA1, dentate gyrus, and subiculum subfields differed between ADS subgroups (i.e., AD vs. MCI). No significant differences were observed between ADS subgroups (Supplementary Table S4). We would like to note that these exploratory subgroup analyses may have been limited by reduced statistical power as a result of dividing the ADS sample into small subgroups, and should be interpreted with caution. Subsequent subfield analyses were limited to subfields which showed significant group differences in the initial group difference ANCOVA models (i.e., CA1, dentate gyrus, and subiculum).

We examined the association between cognitive impairments and the observed degeneration of hippocampal subfields in participants on the ADS. Individual differences in hippocampal degeneration were first computed across all subfields per each member of the ADS group, by deriving the mean Euclidean distance between each patient’s vectorized subfield volumes and those of all healthy controls, resulting in hippocampal degeneration scores (HiDs). These HiDs were then included in separate partial correlations with MMSE, MoCA, and immediate and delayed recollection scores. Significant associations were found between reduced MMSE scores and increased HiDs (r = −0.55, 95% CI [−0.89, −0.20], p_adj = 0.013), but not between MoCA scores and HiDs (r = −0.24, 95% CI [−0.62, 0.14], p_adj = 0.210). Further, significant associations were found between reduced immediate (r = −0.42, 95% CI [−0.78, −0.05], p_adj = 0.047) and delayed recollection (r = −0.36, 95% CI [−0.73, 0.00], p_adj = 0.047) scores and increased HiDs.

To test whether these associations between HiDs and cognitive measures were unique to degeneration of specific subfields, we recomputed HiD scores per each subfield separately using left and right hemispheric values for each member of the ADS group. Here, we focused on those cognitive measures that showed significant associations with the whole-hippocampus HiDs (e.g., MMSE, and immediate and delayed recollection). We found significant associations between reduced immediate recollection performance and increased dentate gyrus HiDs (r = −0.45, 95% CI [−0.80, −0.10], p_adj = 0.042), and between reduced delayed recollection performance and increased CA1 HiDs (r = −0.39, 95% CI [−0.76, −0.03], p_adj = 0.045; Figure 4). Further, significant associations were found between reduced MMSE scores and increased HiDs from CA1 (r = −0.41, 95% CI [−0.77, −0.05], p_adj = 0.045), the dentate gyrus (r = −0.48, 95% CI [−0.83, −0.14], p_adj = 0.036), and the subiculum (r = −0.54, 95% CI [−0.91, −0.16], p_adj = 0.036; Figure 4). There was a trending association observed between increased CA1 HiDs and reduced immediate recollection performance (r = −0.39, 95% CI [−0.75, 0.02], p_adj = 0.053). No significant associations were found between dentate gyrus HiDs and delayed recollection performance (r = −0.38, 95% CI [−0.74, −0.01], p_adj = 0.114), nor between the subiculum and immediate (r = −0.34, 95% CI [−0.74, 0.06], p_adj = 0.115) or delayed (r = −0.218, 95% CI [−0.57, 0.20], p_adj = 0.347) recollection performance.

Next, we conducted a series of exploratory regression models to determine whether the reported effects of HiDs on cognitive measures significantly differed by ADS subgroup level (i.e., AD vs. MCI). We found no evidence for any such moderation effect (Supplementary Table S5). Note, exploratory analyses may have been limited by reduced statistical power as a result of dividing the ADS sample into subgroups. To test the possibility that these associations between degeneration of subfields and cognition could be due to non-hippocampus-specific cortical atrophy, we computed these subfield HiDs-cognition models substituting global cortical thickness estimates from FreeSurfer for subfield HiDs. None of the reported associations with cognitive scores were found when substituting global cortical thickness for HiDs (Supplementary Table S6), indicating a degree of hippocampal specificity for these effects.

We also examined effects of cortical amyloid deposition on the degeneration of hippocampal subfields and their relationships to cognition. We found that the amount of global cortical amyloid-β, measured across a set of amyloid-accumulating cortical areas, did not significantly relate to any subfield HiDs (CA1: r = 0.14, 95% CI [−0.25, 0.53], p = 0.882; dentate gyrus: r = 0.24, 95% CI [−0.14, 0.62], p = 0.272; subiculum: r = 0.16, 95% CI [−0.24, 0.55], p = 0.706; Supplementary Figure S2), nor did it moderate any of the previously-reported effects of hippocampal degeneration on cognitive measures (MMSE~CA1: β = 0.16, 95% CI [−0.19, 0.52], p = 0.292; MMSE~dentate gyrus: β = 0.14, 95% CI [−0.23, 0.51], p = 0.418; MMSE~subiculum: β = 0.12, 95% CI [−0.23, 0.44], p = 0.507; immediate recollection~dentate gyrus: β = 0.11, 95% CI [−0.24, 0.46], p = 0.517; delayed recollection~CA1: β = 0.27, 95% CI [−0.08, 0.61], p = 0.173; Supplementary Figure S3).