Brain age estimates from multiple measures relate to Alzheimer’s markers and thinking skills in early stages

We formed a training sample of healthy controls spanning the adult lifespan by combining structural and FC-MRI data from three sources, as described previously (Millar et al., 2022): the Charles F. and Joanne Knight AD Research Center (ADRC) at Washington University in St. Louis (WUSTL), healthy controls from studies in the Ances lab at WUSTL (Thomas et al., 2013; Petersen et al., 2021), and mutation-negative controls from the Dominantly Inherited Alzheimer Network (DIAN) study of autosomal dominant AD at multiple international sites including WUSTL (McKay et al., 2022). To minimize the likelihood of undetected AD pathology in our training set, participants over the age of 50 were only included in the training set if they were cognitively normal, as assessed by the Clinical Dementia Rating (CDR 0; Morris, 1993), and had at least one biomarker indicating the absence of amyloid pathology (CN/A−, see below). We excluded 59 participants who did not have available CDR or biomarker measures (see Figure 1—figure supplement 1). As CDR and amyloid biomarkers were not available in the Ances lab controls, we included only participants at or below age 50 from this cohort in the training set. These healthy control participants were randomly divided into a training set (~80%; N=390) and a held-out test set (~20%; N=97), which did not significantly differ in age, sex, education, or race, see Table 1.

Finally, independent samples for hypothesis testing included three groups from the Knight ADRC: a randomly selected sample of 144 CN/A− controls who did not overlap with the training or testing sets, 154 CN/A+ participants, and 154 cognitively impaired (CI) participants (CDR > 0 with a biomarker measure consistent with amyloid pathology [see below] and/or a primary diagnosis of AD or uncertain dementia [McKhann et al., 2011]). See Table 1 for demographic details of each sample. All participants provided written informed consent in accordance with the Declaration of Helsinki and their local institutional review board. All procedures were approved by the Human Research Protection Office at WUSTL (IRB ID # 201204041).

Amyloid burden was imaged with PET using (11 C)-Pittsburgh Compound B (PIB; Klunk et al., 2004) or (18 F)-Florbetapir (AV45; Wong et al., 2010). Regional standard uptake ratios (SUVRs) were modeled from 30 to 60 min after injection for PIB and from 50 to 70 min for AV45, using cerebellar gray as the reference region (Su et al., 2013). Regions of interest were segmented automatically using FreeSurfer 5.3 (Fischl, 2012). Global amyloid burden was defined as the mean of partial-volume-corrected (PVC) SUVRs from bilateral precuneus, superior and rostral middle frontal, lateral and medial orbitofrontal, and superior and middle temporal regions (Su et al., 2013). Amyloid summary SUVRs were harmonized across tracers using a centiloid conversion (Su et al., 2018).

Tau deposition was imaged with PET using (18 F)-Flortaucipir (AV-1451; Chien et al., 2013). Regional SUVRs were modeled from 80 to 100 min after injection, using cerebellar gray as the reference region. A tau summary measure was defined in the mean PVC SUVRs from bilateral amygdala, entorhinal, inferior temporal, and lateral occipital regions (Mishra et al., 2017).

CSF was collected via lumbar puncture using methods described previously (Fagan et al., 2006). After overnight fasting, 20–30 mL samples of CSF were collected, centrifuged, then aliquoted (500 µL) in polypropylene tubes, and stored at –80°C. CSF amyloid β peptide 42 (Aβ42), Aβ40, and phosphorylated tau-181 (pTau) were measured with automated Lumipulse immunoassays (Fujirebio, Malvern, PA, USA) using a single lot of assays for each analyte. Aβ42 and pTau estimates were each normalized for individual differences in CSF production rates by forming a ratio with Aβ40 as the denominator (Hansson et al., 2019; Guo et al., 2020). As pTau/Aβ40 was highly skewed, we applied a log transformation to these estimates before statistical analysis.

Amyloid positivity was defined using previously published cutoffs for PIB (SUVR > 1.42; Vlassenko et al., 2016) or AV45 (SUVR > 1.19; Su et al., 2019). Additionally, the CSF Aβ42/Aβ40 ratio has been shown to be highly concordant with amyloid PET (positivity cutoff < 0.0673; Schindler et al., 2018; Volluz et al., 2021). Thus, participants were defined as amyloid-positive (for CN/A+ and CI groups) if they had either a PIB, AV45, or CSF Aβ42/Aβ40 ratio measure in the positive range. Participants with discordant positivity between PET and CSF estimates were defined as amyloid-positive.

Knight ADRC participants completed a 2 hr battery of cognitive tests. We examined global cognition by forming a composite of tasks across cognitive domains, including processing speed (Trail Making A; Schindler et al., 2018), executive function (Trail Making B; Schindler et al., 2018), semantic fluency (Animal Naming; Armitage, 1946), and episodic memory (Free and Cued Selective Reminding Test free recall score; Goodglass and Kaplan, 1983; Grober et al., 1988). This composite has recently been used to study individual differences in cognition in relation the preclinical AD biomarkers and structural MRI (Aschenbrenner et al., 2018), as well as functional MRI measures (Millar et al., 2021).

All MRI data were obtained using a Siemens 3T scanner, although there was a variety of specific models within and across studies. As described previously (Millar et al., 2022), participants in the Knight ADRC and Ances lab studies completed one of two comparable structural MRI protocols, varying by scanner (sagittal T1-weighted magnetization-prepared rapid gradient echo sequence [MPRAGE] with repetition time [TR] = 2400 or 2300 ms, echo time [TE] = 3.16 or 2.95 ms, flip angle = 8 or 9°, frames = 176, field of view = sagittal 256×256 or 240×256 mm, 1 mm isotropic or 1×1×1.2 mm voxels; oblique T2-weighted fast spin echo sequence [FSE] with TR = 3200 ms, TE = 455 ms, 256×256 acquisition matrix, 1 mm isotropic voxels) and an identical resting-state fMRI protocol (interleaved whole-brain echo planar imaging sequence [EPI] with TR = 2200 ms, TE = 27 ms, flip angle = 90°, field of view = 256 mm, 4 mm isotropic voxels for two 6 min runs [164 volumes each] of eyes open fixation). DIAN participants completed a similar MPRAGE protocol (TR = 2300ms, TE = 2.95ms, flip angle = 9°, field of view = 270 mm, 1.1×1.1×1.2 mm voxels; McKay et al., 2022). Resting-state EPI sequence parameters for the DIAN participants differed across sites and scanners with the most notable difference being shorter resting-state runs (one 5 min run of 120 volumes; see Supplementary file 1 for summary of structural and functional MRI parameters; McKay et al., 2022).

All MRI data were processed using common pipelines. Initial fMRI preprocessing followed conventional methods, as described previously (Shulman et al., 2010; Millar et al., 2022), including frame alignment, debanding, rigid body transformation, bias field correction, and normalization of within-run intensity values to a whole-brain mode of 1000 (Power et al., 2012). Transformation to an age-appropriate in-house atlas template (based on independent samples of either younger adults or CN older adults) was performed using a composition of affine transforms connecting the functional volumes with the T2-weighted and MPRAGE images. Frame alignment was included in a single resampling that generated a volumetric time series of the concatenated runs in isotropic 3 mm atlas space.

As described previously (Fox et al., 2009; Millar et al., 2022), additional processing was performed to allow for nuisance variable regression. Data underwent framewise censoring based on motion estimates (framewise displacement [FD] > 0.3 mm and/or derivative of variance [DVARS] > 2.5 above participant’s mean). To further minimize the confounding influence of head motion on FC estimates (Power et al., 2012) in all samples, we only included scans with low head motion (mean FD < 0.30 mm and > 50% frames retained after motion censoring). BOLD data underwent a temporal band-pass filter (0.005 Hz < f < 0.1 Hz) and nuisance variable regression, including motion parameters, timeseries from FreeSurfer 5.3-defined (Fischl, 2012) whole brain (global signal), CSF, ventricle, and white matter masks, as well as the derivatives of these signals. Finally, BOLD data were spatially blurred (6 mm full width at half maximum).

Final BOLD time series data were averaged across voxels within a set of 300 spherical regions of interest (ROIs) in cortical, subcortical, and cerebellar areas (Seitzman et al., 2020). For each scan, we calculated the 300×300 Fisher-transformed Pearson correlation matrix of the final averaged BOLD time series between all ROIs. We then used the vectorized upper triangle of each correlation matrix (excluding auto-correlations; 44,850 total correlations) as input features for predicting age. Since site and/or scanner differences between samples might confound neuroimaging estimates, we harmonized FC matrices using an empirical Bayes modeling approach (ComBat; Johnson et al., 2007; Fortin et al., 2017), which has previously been applied to FC data (Yu et al., 2018).

All T1-weighted images underwent cortical reconstruction and structural segmentation through a common pipeline with FreeSurfer 5.3 (Fischl et al., 2002; Fischl, 2012). Structural processing included segmentation of subcortical white matter and deep gray matter, intensity normalization, registration to a spherical atlas, and parcellation of the cerebral cortex based on the Desikan atlas (Desikan et al., 2006). Inclusion and exclusion errors of parcellation and segmentation were identified and edited by a centralized team of trained research technicians according to standardized criteria (Su et al., 2013). We then used the FreeSurfer-defined thickness estimates from 68 cortical regions (Desikan et al., 2006), along with volume estimates from 33 subcortical regions (Fischl et al., 2002) as input features for predicting age. We harmonized structural features across sites and scanners using the same ComBat approach (Johnson et al., 2007; Fortin et al., 2017), which has also been applied to structural MRI data (Fortin et al., 2018).

As described previously (Millar et al., 2022), machine-learning analyses were conducted using the Regression Learner application in Matlab (MathWorks, 2021). We trained two Gaussian process regression (GPR; Rasmussen et al., 2004) models, each with a rational quadratic kernel function to predict chronological age using fully-processed, harmonized MRI features (FC or structural) in the training set. The σ hyperparameter was tuned within each model by searching a range of values from 10^–4 to 10*SD_age using Bayesian optimization across 100 training evaluations. The optimal value of σ for each model was found (see Figure 1—figure supplement 2) and was applied for all subsequent applications of that model. All other hyperparameters were set to default values (basis function = constant and standardize = true).

Model performance in the training set was assessed using 10-fold cross validation via the Pearson correlation coefficient (r), the proportion of variance explained (R²), the mean absolute error (MAE), and root-mean-square error (RMSE) between true chronological age and the cross-validated age predictions merged across the 10 folds. We then evaluated generalizability of the models to predict age in unseen data by applying the trained models to the held-out test set of healthy controls. Finally, we applied the same fully-trained GPR models to separate analysis sets of 154 CI, 154 CN/A+, and 144 CN/A− controls to test our hypotheses regarding AD-related group effects and individual difference relationships. Unimodal models were each constructed with a single GPR model. The multimodal model was constructed by taking the ‘stacked’ predictions from each first-level unimodal model as features for training a second-level GPR model (Liem et al., 2017; Engemann et al., 2020; Dunås et al., 2021).

For each participant, we calculated model-specific BAG estimates as the difference between chronological age and age predictions from the unimodal FC model (FC-BAG), structural model (S-BAG), and multimodal model (S+FC BAG). To correct for regression dilution commonly observed in similar models (Le et al., 2018; Smith et al., 2019; Liang et al., 2019), we included chronological age as a covariate in all statistical tests of BAG (Cole et al., 2017a; Le et al., 2018). However, to avoid inflating estimates of prediction accuracy (Butler et al., 2021), only uncorrected age prediction values were used for evaluating model performance in the training and test sets.

All statistical analyses were conducted in R 4.0.2 (R Development Core Team, 2020). Demographic differences in the AD samples were tested with independent-samples t tests for continuous variables and χ² tests for categorical variables, using CN/A− controls as a reference group. Differences in brain age model performance were tested using Williams’s test of difference between dependent correlations sharing one variable, i.e., Pearson’s r between age and each model prediction of age. To correct for age-related bias in BAG (Le et al., 2018; as previously mentioned), we controlled for age as a covariate during all statistical tests. Group differences in each BAG estimate were tested using an omnibus ANOVA test with follow-up pairwise t tests on age-residualized BAG estimates, using a false discovery rate (FDR) correction for multiple comparisons. Assumptions of normality were tested by visual inspection of quantile-quantile plots. Assumptions of equality of variance were tested with Levene’s test. Linear regression models tested the effects of cognitive impairment (CDR > 0 vs. CDR 0) and amyloid positivity (A− vs. A+) on BAG estimates from each model, controlling for true age (as noted above), sex, and years of education. Given the potential confounding influence of head motion on FC-derived measures (Power et al., 2012; Van Dijk et al., 2012; Satterthwaite et al., 2012), we also included mean FD as an additional covariate of non-interest in the FC and S+FC models. We tested continuous relationships with AD biomarkers and cognitive estimates using linear regression models, including the same demographic and motion covariates. Since the range of amyloid biomarkers was drastically reduced in the CN/A− sample, we excluded these participants from models testing continuous amyloid relationships. Effect sizes were computed as partial η² (η_p²).

Demographic characteristics of the training sets, test sets, and analysis sets are reported in Table 1. CN/A+ participants were older (t = 6.15, p < 0.001) and more likely to be APOE ε4 carriers (χ² = 34.73, p < 0.001) than amyloid-negative controls. Furthermore, CI participants were older (t = 9.71, p < 0.001), more likely male (χ² = 8.60, p = 0.003), more likely to be APOE ε4 carriers (χ² = 56.67, p < 0.001), and had fewer years of education (t = 2.03, p < 0.043), and lower MMSE scores (t = 12.46, p < 0.001) than amyloid-negative controls.

All models accurately predicted chronological age in the training sets, as assessed using 10-fold cross validation, as well as in the held-out test sets. Overall, prediction accuracy was lowest in the FC model (MAE_FC/Train = 8.67 years, R²_FC/Train = 0.68, MAE_FC/Test = 8.25 years, R²_FC/Test = 0.73; see Figure 1A & B). The structural MRI model (MAE_S/Train = 5.97 years, R²_S/Train = 0.81, MAE_S/Test = 6.26 years, R²_S/Test = 0.82; see Figure 1C & D) significantly outperformed the FC model in age prediction accuracy, Williams’s t_{S vs. FC} = 5.39, p < 0.001. There was a significant, but modestly sized, positive correlation between FC-BAG and S-BAG in the adult lifespan CN/A− training and testing sets (r = 0.095, p = 0.036; see Figure 1—figure supplement 3A), as well as the AD analysis sets (r = 0.134, p = 0.004; see Figure 1—figure supplement 3B).

Finally, the multimodal model (MAE_S+FC/Train = 5.34 years, R²_S+FC/Train = 0.86, MAE_S+FC/Test = 5.25 years, R²_S+FC/Test = 0.87; see Figure 1E & F) significantly outperformed both the FC model (Williams’s t_{S+FC vs. FC} = 11.20, p < 0.001) and the structural MRI model (Williams’s t_{S+FC vs. S} = 5.67, p < 0.001). It is possible that the modest increase in the multimodal model was due to capitalizing on noise, simply by adding more features to the structural model. Hence, we also compared the observed R²_S+FC to a bootstrapped distribution of R² performance estimates from 1000 resamples using a model in which the original structural MRI model was stacked with a model trained on randomly reshuffled FC features. Thus, this distribution represents the expected improvements in model performance from simply adding new features to the structural MRI model with the stacked approach. The observed R²_S+FC outperformed all R² estimates from this bootstrapped distribution (p < 0.001; see Figure 1—figure supplement 4), suggesting that the modest increase in model performance observed in the stacked multimodal (S+FC) model over the unimodal structural model is due to meaningful age-related FC signal, rather than capitalizing on noise in a larger feature set.

Residual FC-BAG was normally distributed (see Figure 2—figure supplement 1), and variance in FC-BAG did not significantly differ between the analysis sets, Levene’s statistic = 0.01, p = 0.988. An omnibus ANOVA revealed significant differences in residual FC-BAG across the three groups, F(2,449) = 9.80, p < 0.001. FC-BAG was 2.17 years older in CI participants compared to CN controls (β = 2.17, p = 0.030, η_p² = 0.01; see Figure 2A&B, Table 2A). Follow-up t tests revealed that residual FC-BAG was significantly elevated in CI relative to CN/A+participants (p_FDR < 0.001). FC-BAG was also 1.64 years lower in A+ participants compared to A− (β = –1.64, p = 0.035, η_p² = 0.01), controlling for global CDR and the other covariates. Follow-up t tests revealed that residual FC-BAG was significantly lower in CN/A+ participants compared to CN/A− controls (p_FDR = 0.002).

Residual S-BAG was also normally distributed (see Figure 2—figure supplement 1), and variance in S-BAG did not significantly differ between the analysis sets, Levene’s statistic = 0.10, p = 0.902. An omnibus ANOVA revealed significant differences in residual S-BAG across the three groups, F(2,449) = 20.64, p < 0.001. S-BAG was 5.10 years older in CI participants compared to CN controls (β = 5.10, p < 0.001, η_p² = 0.04; see Figure 2C&D, Table 2B). Follow-up t tests revealed that residual S-BAG was significantly elevated in CI participants relative to CN/A− and CN/A+ participants (p_FDR’s < 0.001). S-BAG did not significantly differ as a function of amyloid positivity, controlling for CDR and the other covariates.

Residual S+FC-BAG was also normally distributed (see Figure 2—figure supplement 1), and variance in S+FC-BAG did not significantly differ between the analysis sets, Levene’s statistic = 0.89, p = 0.412. An omnibus ANOVA revealed significant differences in residual S+FC-BAG across the three groups, F(2,449) = 21.84, p < 0.001. S+FC-BAG was 4.31 years older in CI participants compared to CN controls (β = 4.1, p < 0.001, η_p² = 0.04; see Figure 2E, F, Table 2C). Follow-up t tests revealed that residual FC-BAG was significantly elevated in CI participants relative to CN/A− and CN/A+ participants (p_FDR’s < 0.001). S+FC-BAG did not significantly differ as a function of amyloid positivity, controlling for CDR and the other covariates.

355 participants (144 CN/A−, 154 CN/A+, 57 CI) had an available amyloid PET scan, and 300 (120 CN/A−, 137 CN/A+, 43 CI) had an available CSF estimate of Aβ42/40. In the FC model, FC-BAG was not significantly related with amyloid PET nor was there an interactive relationship with amyloid PET between groups (see Figure 3A). There were also no significant main effects or interactions between FC-BAG, S-BAG, or S+FC BAG and CSF Aβ42/40 (See Figure 3B, D and F).

In the structural and multimodal models, there were significant main effects, such that greater S-BAG (β = 0.79, p = 0.004, η_p² = 0.041; see Figure 3C) and greater S+FC BAG (β = 0.81, p = 0.015, η_p² = 0.029; see Figure 3E) were both associated with greater amyloid PET. In the multimodal model only, this relationship was further characterized by a non-significant interaction (β = 1.16, p = 0.087, η_p² = 0.014), such that the association was significantly positive in CI participants interaction (β = 1.53, p = 0.029, η_p² = 0.092) but not in CN/A+ (β = –0.05, p = 0.881, η_p² = 0.001).

99 participants (42 CN/A–, 40 CN/A+, 17 CI) had an available tau PET scan, and 300 (120 CN/A–, 137 CN/A+, 43 CI) had an available CSF estimate of pTau-181/Aβ40. In the FC model, FC-BAG was not significantly related with tau PET or CSF pTau-181/Aβ40 (see Figure 4A and B). However, there was a non-significant interaction, suggesting a more positive association between CSF pTau-181/Aβ40 and FC-BAG in CI participants but not in CN controls (β = 0.02, p = 0.059, η_p² = 0.016).

In the structural and multimodal models, there were significant main effects, such that greater S-BAG (β = 0.02, p < 0.001, η_p² = 0.141; see Figure 4C) and greater S+FC BAG (β = 0.02, p = 0.001, η_p² = 0.110; see Figure 4E) were both associated with greater tau PET. These main effects were further characterized by significant interactions (S-BAG: β = 0.04, p < 0.001, η_p² = 0.176; S+FC-BAG: β = 0.07, p < 0.001, η_p² = 0.250), such that the positive association was only observed in CI participants, but not in the other groups.

Consistent with tau PET, CSF pTau/Aβ40 demonstrated similar interactive relationships, such that greater S-BAG (β = 0.02, p < 0.001, η_p² = 0.052; see Figure 4D) and greater S+FC BAG (β = 0.04, p < 0.001, η_p² = 0.075; see Figure 4F) were both associated with greater CSF pTau/Aβ40 in the CI participants, but not in the other groups.

445 participants (144 CN/A−, 153 CN/A+, 148 CI) had available performance measures from the cognitive composite tasks. In the FC model, there was a significant main effect, such that across all groups, greater FC-BAG was associated with lower cognitive composite score (β = –0.01, p = 0.006, η_p² = 0.017; see Figure 5A). However, this effect was driven by group differences in both variables, as there were neither relationships between FC-BAG and cognition within any of the groups nor were there any significant interactions.

In the structural model and multimodal models, there were significant main effects, such that greater S-BAG (β = –0.03, p < 0.001, η_p² = 0.104; see Figure 5B) and greater S+FC BAG (β = –0.03, p < 0.001, η_p² = 0.096; see Figure 5C) were both associated with lower cognitive composite scores. Both effects were further characterized by significant interactions such that the negative associations were observed in the CI participants, but not in the other groups (S-BAG: β = –0.03, p < 0.001, η_p² = 0.045; S+FC-BAG: β = –0.04, p < 0.001, η_p² = 0.047).

We found that a GPR model trained on structural MRI features predicted chronological age in a cognitively normal, amyloid-negative adult sample with an R² of 0.81. This level of performance is comparable to other structural models, which have reported R²s ranging from 0.80 to 0.95 (Cole and Franke, 2017b; Liem et al., 2017; Eavani et al., 2018; Wang et al., 2019; Bashyam et al., 2020; Ly et al., 2020; Gong et al., 2021; Lee et al., 2022). As previously reported (Millar et al., 2022), the FC-trained model predicted age with an R² of 0.68, again consistent with previous FC models, which have achieved R²s from 0.53 to 0.80 (Liem et al., 2017; Eavani et al., 2018; Gonneaud et al., 2021). Our observation that structural MRI outperformed FC in age prediction is also consistent with previous direct comparisons between modalities (Liem et al., 2017; Eavani et al., 2018; Dunås et al., 2021).

Importantly, however, there was only a modest positive correlation between FC and structural BAG estimates, after correcting for age-related biases, suggesting that functional and structural MRI capture distinct age-related signals. Indeed, the multimodal model outperformed both unimodal models by integrating these complementary signals. These observations, again, are consistent with other recent reports of multimodal age prediction models (Liem et al., 2017; Eavani et al., 2018; Engemann et al., 2020; Dunås et al., 2021). Future models may improve age prediction accuracy by combining data from structural, FC, and/or other neuroimaging modalities, several of which may be available in typical MRI sessions of multiple sequences.

Structural BAG was elevated by 5.10 years in CI participants compared to CN controls. This effect is comparable to previous structural age prediction models, demonstrating elevations in AD and MCI samples between 5 and 10 years (Cole and Franke, 2017b; Franke and Gaser, 2019). As previously reported, FC BAG was also elevated in CI participants, but to a relatively smaller extent, i.e., 2.17 years (Millar et al., 2022). The multimodal BAG was similarly elevated in CI participants by 5.10 years. Thus, each model is clearly sensitive to group differences in AD status at the symptomatic stage.

Consistent with one previous report (Lee et al., 2022), we demonstrated that within the CI participants, BAG estimates were related to individual differences in AD biomarkers and cognitive function. These effects were most pronounced in the structural model, which showed relationships with tau biomarkers and cognition in the CI participants, and the multimodal model, which showed relationships with tau, cognition, and amyloid PET. Thus, age prediction models that include structural MRI (including unimodal and multimodal approaches) may be useful in tracking AD pathological progression and cognitive decline within the symptomatic stage of the disease.

We found that structural and multimodal BAG did not differ between cognitively normal participants with and without amyloid pathology. In cognitively normal participants, structural BAG estimates did not significantly associate with individual differences in any AD biomarkers. Overall, although structural and multimodal BAG estimates track well with some biomarkers of AD pathophysiology, as previously reported (Lee et al., 2022), our novel results suggest that these relationships are not observed until the symptomatic stage of the disease, at which point structural changes become more apparent.

As we have previously reported (Millar et al., 2022), FC-BAG was lower in presymptomatic AD participants compared to amyloid-negative controls. Extending beyond this group difference, we now also note that FC-BAG was negatively associated with amyloid PET in CN/A+ participants. The combined reduction of FC-BAG in the presymptomatic stage and increase in the symptomatic stage suggest a biphasic functional response to AD progression, which is partially consistent with some prior suggestions (Jagust and Mormino, 2011; Jones et al., 2016; Jones et al., 2017; Schultz et al., 2017; Wales and Leung, 2021; see Millar et al., 2022 for a more detailed discussion).

Interpretation of this biphasic pattern is still unclear, although the present results provide at least one novel insight. Specifically, one potential interpretation is that the ‘younger’ appearing FC pattern in the presymptomatic stage may reflect a compensatory response to early AD pathology (Cabeza et al., 2018). This interpretation leads to the prediction that reduced FC-BAG should be associated with better cognitive performance in the preclinical stage. However, this interpretation is not supported by the current results, as FC-BAG did not correlate with cognition in any of the analysis samples.

Alternatively, pathological AD-related FC disruptions may be orthogonal to healthy age-related FC differences, as supported by our previous observation that age and AD are predicted by mostly non-overlapping FC networks (Millar et al., 2022). For instance, the ‘younger’ FC pattern in CN/A+ participants may be driven by hyper-excitability in the preclinical stage (Harris et al., 2020; Ranasinghe et al., 2022). It is also worth considering that patterns of younger FC-BAG in CN/A+ participants may somehow correspond to a recent observation that patterns of youthful-appearing aerobic glycolysis are relatively preserved in the presymptomatic stage of AD (Goyal et al., 2022). Finally, this effect may simply be spuriously driven by poor performance of the FC brain age model, sample-specific noise, and/or statistical artifacts related to regression dilution and its correction (Butler et al., 2021). Hence, future studies should attempt to replicate these results in independent samples and further test potential theoretical interpretations.

Although FC-BAG was not associated with individual differences in a global cognitive composite within any of our analysis samples, greater structural and multimodal BAG estimates were associated with lower cognitive performance within the CI participants. Hence, these estimates may be sensitive markers of cognitive decline in the symptomatic stage. This finding is consistent with previous reports that other structural brain age estimates are associated with cognitive performance in AD (Eavani et al., 2018), Down syndrome (Cole et al., 2017a), HIV (Petersen et al., 2021; Petersen et al., 2022), as well as cognitively normal controls (Richard et al., 2018).

The training sets included MRI scans from a range of sites, scanners, and acquisition sequence parameters, which may introduce noise and/or confounding variance into MRI features. We attempted to mitigate this problem by: (1) including only data from Siemens 3T scanners with similar protocols; (2) processing all MRI data through common pipelines and quality assessments; and (3) harmonizing across sites and scanners with ComBat (Fortin et al., 2017).

Additionally, the training set (N = 390) was relatively small compared to prior models, which have included training samples over 1,000 (e.g., Cole et al., 2015; Bashyam et al., 2020). Future studies may further improve model performance by including larger samples of well-characterized participants in the training set.

Although we took appropriate steps to detect and control for AD-related pathology in the CN/A− training sets, we were unable to control for other non-AD pathologies, e.g., Lewy body disease, TDP-43, etc., which may be present.

Structural MRI was quantified using the Desikan atlas (Desikan et al., 2006), which, although widely used, provides a relatively coarse parcellation of structural anatomy and, moreover, does not align with the parcellation used to define FC regions (Seitzman et al., 2020). Although the structural MRI data still outperformed FC in predicting age, future brain age models may further improve performance by using more refined and harmonized anatomical parcellations to define brain regions.

The sample size of continuous biomarker and cognitive analyses differed across the measures, depending on the availability, and was particularly low for analyses of tau PET. Future studies might improve upon this approach by a larger and more complete biomarker sample.

Moreover, estimates of BAG likely capture variance in early-life factors, which may obscure associations with AD and cognition, especially in cross-sectional designs (Vidal-Piñeiro et al., 2021). Future studies may improve the sensitivity of BAG estimates to disease-related markers by testing associations with longitudinal change.

Finally, although the Ances lab controls were relatively diverse, participants in other samples were mostly white and highly educated. Hence, these models may not be generalizable to broader samples. Future models would benefit by using more representative training samples.

We compared three MRI-based machine-learning models in their ability to predict age, as well as their sensitivity to early-stage AD, AD biomarkers, and cognition. Although FC and structural MRI models were both successful in detecting differences related to healthy aging and cognitive impairment, we note clear evidence that these modalities capture complementary signals. Specifically, FC-BAG was uniquely reduced in cognitively normal participants with elevated amyloid, although the interpretation of this finding still warrants further investigation. In contrast, structural BAG was uniquely associated with biomarkers of AD pathology and cognitive function within the CI participants. Finally, the multimodal age prediction model, which combined FC and structural MRI, further improved the prediction of healthy age differences and also was related to biomarkers and cognition in CI participants. Thus, multimodal brain age models may be useful maximizing sensitivity to AD across the spectrum of disease progression.