Among a larger pool of clinically and cognitively normal (NL) individuals participating in longitudinal brain imaging studies at New York University (NYU) Langone School of Medicine, this study was based on a sub-set of 62 NL participants who participated in a lifestyle survey between 2013-2014. Subjects were derived from multiple community sources, including individuals interested in research participation, family members and caregivers of impaired patients. All subjects provided written informed consent to participate in this IRB approved study. Individuals with medical conditions or history of conditions that may affect brain structure or function, i.e. stroke, diabetes, head trauma, any neurodegenerative diseases, depression, hydrocephalus, intracranial mass, and infarcts on MRI, and those taking psychoactive medications were excluded. Subjects were 25 - 71 year old, with education ≥ 12 y, Clinical Dementia Rating (CDR) = 0, Global Deterioration Scale (GDS) ≤ 2, Mini Mental State Examination (MMSE) ≥ 28, Hamilton depression scale < 16, Modified Hachinski Ischemia Scale < 4 and normal cognitive test performance for age and education [19] . Study analyses focused on 45 participants who fulfilled our inclusion criteria and completed all clinical, MRI, PiB-and FDG-PET exams, physical activity and dietary questionnaires within 6 months of each other. While all subjects were normoglycemic young adults, insulin sensitivity was calculated using the Homeostasis Model Assessment (HOMA) [20] . Presence of hypertension was determined based on current antihypertensive treatment or blood pressure assessments (systolic blood pressure ≥ 140 mmHg or diastolic blood pressure ≥ 90 mmHg) [21] . A family history of LOAD that included at least one 1^st degree relative whose AD onset was after age 60 was elicited using standardized questionnaires [22] . Apolipoprotein E (APOE) genotypes were determined using standard qPCR procedures [23] .

Physical activity was estimated by means of the Minnesota leisure time activity (LTA) questionnaire [24] . The questionnaire consists of a list of 62 leisure-time physical activities separated into three sections, including walking and miscellaneous (e.g., walking for pleasure, using stairs when an elevator is available, etc.), conditioning exercise (e.g., running, health club activities, etc.), and other exercise (e.g., playing golf, shoveling snow, etc.). Participants were asked if they participated in any of the given activities. For each activity, information was collected on the frequency (how often per month and how many months per year) and duration of engagement (time spent per session). Frequency and duration information were multiplied with an activity-specific intensity code indicating calorie expenditure [24] . The activity-dependent scores were summed to obtain the overall intensity of physical activity per person during the last 12 months and converted to metabolic equivalents (MET), which are independent of body weight [24] . A cut-off of 7.5 MET hours/week (i.e., the energy cost of engaging in 30 min of moderate activity 5 days per week) was used to divide subjects into more (LTA+) and less physically active (LTA−) groups, according to American Heart Association guidelines [25] [26] .

Dietary data regarding average food consumption over the prior year were obtained using the 153-item version of Harvard/Willett’s semi-quantitative food frequency questionnaire (SFFQ) [27] . The SFFQ has been used and validated for the determination of nutrient intake in the elderly as well as in young adults, yielding high reliability [27] . Food items were categorized into 30 food groups based on similarities in food and nutrient composition, and intake (g/day) of each food group was calculated by summing the intakes of member food items. The MeDi is characterized by high intake of plant foods; moderate consumption of dairy products, fish, poultry; olive or vegetable oil as the primary dressing; low to moderate intake of wine; low intake of red meat; very low intake of processed foods [7] . Published methods were followed for the construction of MeDi scores [4] -[6] [9] . Briefly, we first regressed caloric intake (in kilocalories) and calculated the derived residuals of daily gram intake for each of the following 7 categories: dairy, meat, fruits, vegetables, legumes, cereals, and fish. Individuals were assigned a value of 1 for each beneficial component (fruits, vegetables, legumes, cereals, and fish) whose consumption was at or above the sex-specific median; a value of 1 for each detrimental component (meat and dairy products) whose consumption was below the median; a value of 1 for a ratio of monounsaturated fats to saturated fats above the median; and a value of 1 for mild to moderate alcohol consumption (1 - 2 drinks per day) [4] -[6] [9] . The MeDi score was the sum of the scores in the food categories, with a greater score indicating greater similarity to the MeDi pattern. Participants were dichotomized into higher and lower MeDi adherence groups [9] .

All subjects received MRI, PiB- and FDG-PET scans following standardized protocols [28] [29] . The diagnostic MRI study was performed using contiguous 3 mm axial T2-weighted images. These scans were used to rule out MRI evidence of hydrocephalus, intracranial mass, strokes and subcortical gray matter lacunes. The research MRI scan was a volumetric 124 slice T1-weighted Fast-Gradient-Echo acquired in a sagittal orientation as 1.2 mm thick sections. For PET, subjects were positioned in the scanner 60 min after injection of 15 mCi of ¹¹C-PiB, and scanned for 30 min in 3D-mode on an LS Discovery or BioGraph PET/CT scanner. The FDG scan was performed 30 min after completion of the PiB scan or on a separate day. After an overnight fast, subjects were injected with 5 mCi of ¹⁸F-FDG, positioned in the scanner 35 min after injection, and scanned for 20 min. All images were corrected for photon attenuation, scatter, and radioactive decay and smoothed for uniform resolution [30] . Image analysis was done blind to clinical data. For each subject, summed PET images corresponding to 40 - 60 min of FDG data and 60 - 90 min of PiB data were coregistered to MRI using the Normalized Mutual Information (NMI) routine of Statistical Parametric Mapping (SPM8) [31] . Parametric standardized uptake value ratio (SUVR) images were generated by normalizing PiB uptake by cerebellar grey matter uptake and FDG by whole brain activity. MRIs were segmented into grey (GM), white matter (WM) and cerebrospinal fluid (CSF) and normalized to Montreal Neurological Institute (MNI) space by high-dimensional warping (DARTEL) using voxel-based morphometry, VBM8 [31] [32] . MRI-coregistered PET scans were spatially normalized using subject-specific transformation matrixes obtained from MRI, and smoothed with a 10 mm FWHM filter. MRIs were processed using VBM8. A custom template was created using MRI from all subjects by normalizing and segmenting the MRIs using the unified segmentation model with the MNI template and tissue probability maps (TPMs), and averaging the normalized subject TPMs [31] . Individual scans were then processed using the custom TPMs. Jacobian modulation was applied to restore absolute GM volumes (GMV) in the GM images, which were smoothed with an 8-mm FWHM kernel and normalized to total intracranial volumes.

SPSS v.21 (SPPS Inc., 2013) and PLS Tool v1.0 implemented using MATLAB v7.2.0 (StatSoft Inc.) were used for analysis. The General Linear Model (GLM) and χ² tests were used to compare clinical and demographical measures across LTA and MeDi groups (p < 0.05). Multivariate Partial Least Squares (PLS) regression analysis as implemented in PLS v1.0 was used for image analysis to compare LTA (LTA+ vs. LTA−), MeDi (MeDi+ vs. MeDi−) and LTA × MeDi combinations [33] -[35] . PLS regression is a multivariate extension of the multiple linear regression model that is used to construct predictive models when the number of variables (e.g., voxels) are far larger than the number of observations (e.g. subjects), and multi-collinear, as previously described [33] -[35] . Briefly, while univariate analysis is used to identify reliable signal changes at the level of individual image elements (i.e., voxels), multivariate analysis focuses on the examination of distributed patterns by taking advantage of the spatial and/or temporal dependencies (i.e., covariance) among image elements, thus enabling inferences about differences across space and/or time. The PLS procedure generates a latent variable (LV) brain map that includes all brain regions that are most correlated with the predicting variable (e.g., condition, or group), thus yielding spatial patterns of brain biomarkers that represent the optimal association between the images and the grouping factors [33] . The strength of the relationship is measured by the correlation between conditions and LV scores in the data. The final LV is a numeric score or value that corresponds to the expression of a voxel-pattern or, if only discrete variables are used, a mathematical combination of those discrete variables, for each subject [33] .

For each modality (PiB, FDG, MRI), PLS analysis was performed to identify spatial patterns of brain biomarkers that represent the optimal association between brain images and grouping factors (LTA, MeDi, LTA × MeDi). LV scores were extracted from each significant pattern. Age, gender, education, ethnicity, BMI, APOE status, family history, presence of hypertension and HOMA scores were examined as covariates. The General linear model with post-hoc t-tests was used to compare LV scores across LTA × MeDi groups in post-hoc testing of interaction effects. All results were considered significant at p < 0.05. Multivariate analysis controls for the multiple chances to find group differences, and it does so without assuming independence of the dependent variables, yielding corrected p values. A gray segment mask derived from the MNI template was applied to the PiB and FDG scans during PLS analysis to focus the analysis on gray matter voxels. Split-half resampling was used to obtain unbiased measures of brain-LV reproducibility and prediction of the PLS model for independent data [35] , whereby for each of the 2-group LTA and MeDi, and 4-group LTA × MeDi evaluations, five additional PLS results were generated using five different randomized combinations of 50% of the subjects from each group. These combinations included completely distinct sub-populations [35] . The resulting patterns were then compared for commonality across the six combinations of subjects, before and after resampling, to increase stability of results and provide indication of generalizability [35] . Anatomical location of brain regions showing significant effects was described using Talairach and Tournoux coordinates, after conversion from MNI space.

Subjects’ characteristics are found in Table 1. None of the participants were diabetics, regular smokers, or met criteria for obesity as defined by a Body-Mass index (BMI) > 30 kg/m². Thirty subjects (67%) were taking no medications. The remaining 15 subjects reported taking either or a combination of the following medications: high blood pressure medications (beta-, angiotensin receptor-, calcium channel-blockers, ACE inhibitors) and/or statins (16%), anti-depressants/SSRI (4%; note: all subjects washed off anti-depressants for at least 2 weeks before FDG-PET), prostate medications (4%), hormone replacement therapy (9%).

Abbreviations: MeDi = adherence to a Mediterranean-type diet (higher vs. lower, MeDi+ vs. MeDi−); LTA = engagement in leisure time activity (higher vs. lower, LTA+ vs. LTA−).

Of the 45 participants, 21 (47%) were more physically active (LTA+) and 24 (53%) were less physically active (LTA−). Eighteen (40%) showed higher adherence to the MeDi (MeDi+) and 27 (60%) showed lower adherence (MeDi−). The combination of physical activity and diet resulted in 4 sub-groups: physically active with higher MeDi adherence (LTA+/MeDi+; n = 7, 16%), physically active with lower MeDi adherence (LTA+/ MeDi−; n = 14, 30%), less physically active with higher MeDi adherence (LTA−/MeDi+; n = 12, 27%) and less physically active with lower MeDi adherence (LTA−/MeDi−; n = 12, 27%).

Groups were comparable for clinical, demographical and neuropsychological measures (Table 1). There was a non-significant trend towards a higher frequency of women vs. men in LTA− vs. LTA+ (83% vs. 57%, p = 0.09), and in MeDi+ vs. MeDi− groups (83% vs. 63%, n = 0.14, n.s.). The MeDi− group included slightly more individuals with hypertension than the MeDi+ group (33% vs. 12%, p = 0.11).

Physical activity (Figure 1 (A)). Significant differences in PiB retention were observed between LTA− and LTA+ groups (p < 0.001). The spatial pattern that provided the best separation between LTA groups included the precuneus, posterior cingulate, parietal, frontal, temporal and occipital cortex of both hemispheres (R² = 0.33, p < 0.001). This pattern was characterized by greater PiB retention, reflecting increased Aβ load, in LTA− as compared to LTA+ subjects. Age and HOMA scores were the only covariates that showed significant associations with the identified pattern (R² = 0.09, p < 0.05). Including these variables in the model as covariates did not attenuate the strength of the association (R² = 0.39, p < 0.001).

Diet (Figure 1 (B)). Significant differences in PiB retention were observed between MeDi− and MeDi+ groups (p < 0.001). The spatial pattern that provided the best separation between MeDi groups included frontal and temporal cortex, insula, and putamen of both hemispheres (R² = 0.39, p < 0.001). This pattern was characte-

rized by greater PiB retention in MeDi− as compared to MeDi+ subjects. Age was borderline associated with the identified pattern (R² = 0.06, p ≤ 0.09). Correcting for age left results unchanged (R² = 0.39, p < 0.001).

Physical activity × Diet (Figure 1 (C)). There were significant associations between increasing lifestyle risk and PiB retention (p < 0.001). The spatial pattern that provided the best separation between LTA × MeDi subgroups included the precuneus, posterior cingulate, frontal and parieto-temporal cortex of both hemispheres, and left occipital gyrus, and was characterized by increasingly higher PiB retention as: LTA−/MeDi− > LTA−/Me- Di+ > LTA+/MeDi− > LTA+/MeDi+ (R² = 0.40, p < 0.001). On post-hoc analysis, the LTA−/MeDi− group had higher PiB LV scores than LTA+/MeDi+ (p < 0.001) and LTA+/MeDi− groups (p = 0.005); and the LTA−/ MeDi+ group had higher PiB LV scores than the LTA+/MeDi+ group (p = 0.01) and marginally higher LV scores than the LTA+/MeDi− group (p = 0.07).Among covariates, age and HOMA scores showed trends towards associations with the identified pattern (R² ≤ 0.08, p ≤ 0.10). Including these variables as covariates did not attenuate the strength of the association (R² = 0.47, p < 0.001). Despite the significant correlations, the interaction of LTA × MeDi factors did not reach statistical significance, with and without correcting for covariates.

Resampling. Split-half resampling confirmed results obtained with the entire data set, with consistently higher PiB retention in AD-vulnerable regions for the higher-risk vs. lower-risk groups (p < 0.001). The greatest consistency across split half combinations was found for the LTA− vs. LTA+, and LTA−/MeDi− vs. LTA+/MeDi+ comparisons. Anatomical location and coordinates of significant clusters are found in Table 2.

Physical activity. Significant CMRglc differences were observed between LTA− and LTA+ groups (p < 0.001). The spatial pattern that provided the best separation between LTA groups included lateral and medial temporal cortex (parahippocampal gyrus and uncus), fusiform gyrus and inferior frontal cortex of both hemispheres (R² = 0.37, p < 0.001, Figure 2(A)). This pattern was characterized by reduced CMRglc in LTA− as compared to LTA+ subjects (Figure 2(A)). Age and HOMA scores showed borderline associations with the identified pattern (R² ≤ 0.11, p ≤ 0.09). Including these variables as covariates in the model did not attenuate the strength of the association (covariates: age, R² = 0.38, p < 0.001; HOMA scores, R² = 0.40, p < 0.001).

Diet. Significant CMRglc differences were observed between MeDi− and MeDi+ groups (p < 0.001). The spatial pattern that provided the best separation between MeDi groups included broadly the same regions as with the LTA groups, but was primarily restricted to the lefthemisphere (R² = 0.24, p < 0.001, Figure 2(B)). This pattern was characterized by reduced CMRglc in MeDi− as compared to MeDi+ subjects. Age was the only covariate associated with the identified pattern (R² = 0.18, p = 0.003). Correcting for age left results unchanged (R² = 0.23, p < 0.001, Figure 2(B)).

Physical activity × Diet. There were significant associations between increasing risk and reduced CMRglc (p < 0.001), as well as significant interaction effects of LTA × MeDi on CMRglc (p = 0.03). The spatial pattern that provided the best separation between LTA × MeDi subgroups included the same brain regions as above (e.g., lateral and medial temporal cortex, inferior frontal cortex, bilaterally) and was characterized by increasingly lower CMRglc, as: LTA−/MeDi− < LTA−/MeDi+ < LTA+/MeDi− < LTA+/MeDi+ (R² = 0.40, p < 0.001, Figure 2 (C)). Interaction effects were driven by the LTA+/MeDi+ group showing higher FDG LV scores than all other groups (p < 0.05); and the LTA+/MeDi− group showing higher FDG LV scores than LTA−/MeDi− and LTA−/MeDi+ groups (p ≤ 0.04). Results remained significant correcting for age (p < 0.05, Figure 2 (C)). None

of the other covariates showed significant associations with the identified pattern.

Resampling. Split-half resampling confirmed results obtained with the entire data set, with consistently lower CMRglc, especially in medial temporal cortex, for the higher-risk vs. lower-risk groups (p < 0.001 for main effects, p < 0.05 for interaction effects). The greatest consistency across split half combinations was observed for the LTA− vs. LTA+, and LTA−/MeDi− vs. LTA+/MeDi+ comparisons (p < 0.001).Anatomical location and coordinates of significant clusters are found in Table 3.

Physical activity. Significant differences in GMV were observed between LTA− and LTA+ groups (p < 0.001). The spatial pattern that provided the best group separation included the superior and orbital frontal cortex, and cerebellum (R² = 0.397, p < 0.001, Figure 3(A)). This pattern was characterized by reduced GMV, reflecting increased atrophy, in frontal cortex, and relative GMV increases in cerebellum, of LTA− as compared to LTA+ subjects. None of the covariates showed significant associations with the identified pattern.

Diet. Significant differences in GMV were observed between MeDi− and MeDi+ groups (p < 0.001). The spatial pattern that provided the best separation between MeDi groups included posterior cingulate cortex and precuneus, frontal and temporal cortex, and cerebellum (R² = 0.49, p < 0.001, Figure 3(B)). This pattern was characterized by reduced cortical GMV, and relative cerebellar GMV increases, in MeDi− as compared to MeDi+ subjects. Gender was associated with the identified pattern (R² = 0.11, p = 0.02), and age showed borderline effects (R² = 0.05, p = 0.11). Including these variables as covariates in the model did not attenuate the strength of the association (covariates: age, R² = 0.48, p < 0.001; gender, R² = 0.40, p < 0.001).

See legend to Table 2.

Physical activity × Diet. There were significant associations between increasing risk and reduced GMV (p < 0.001). The spatial pattern that provided the best separation between LTA × MeDi subgroups included the superior, medial frontal and to a lesser extent, posterior cingulate cortex, and was characterized by increasingly lower GMV, as: LTA−/MeDi− < LTA−/MeDi+ < LTA+/MeDi− < LTA+/MeDi+ (R² = 0.42, p < 0.001, Figure 3 (C)). On post-hoc analysis, the LTA−/MeDi− group had lower GMV LV scores than LTA+/MeDi+ and LTA+/Me- Di− groups (p < 0.07). None of the covariates showed significant associations with the identified pattern. Despite the significant correlations, the interaction of LTA × MeDi factors did not reach statistical significance, with and without correcting for covariates.

Resampling. Split-half resampling confirmed the relative GMV decreases in cerebellum found in MeDi and LTA × MeDi comparisons for the higher-risk vs. lower-risk groups (p < 0.01), and lack of interaction effects overall. However, split half comparisons showed greater variability and lower concordance of other regional effects across combinations than with PiB and FDG data. It was noted that the clusters found to be significant in the group comparisons as a whole were less in extent than those observed with PiB or FDG. Anatomical location and coordinates of significant clusters are found in Table 4.

PiB LV scores from the LTA pattern were positively associated with HOMA scores and plasma triglycerides (r = 0.31 and 0.32, respectively, p < 0.05) and negatively, though weakly associated with HDL/LDL ratios (r = −0.26, p = 0.09). CMRglc LV scores from the LTA pattern were negatively associated with HOMA scores (r = −0.33), plasma triglycerides (r = −0.36) and HDL/LDL ratios (r = −0.36, p < 0.04). CMRglcLV scores from the LTA × MeDi pattern were also negatively associated with HDL/LDL ratios (r = −0.33, p < 0.03). GMV LV scores were

See legend to Table 2.

not associated with any clinical measures. There were no associations between LV scores and neuropsychological measures for any biomarkers.