The effect of weight loss following 18 months of lifestyle intervention on brain age assessed with resting-state functional connectivity

Training, validation, and testing of the brain age model were conducted on data from two cohorts that included functional and structural brain MRI. The training was conducted on the enhanced Nathan Kline Institute (NKI)-Rockland Sample (Nooner et al., 2012) and testing on the Cam-CAN dataset (Shafto et al., 2014; Taylor et al., 2017). The NKI dataset is composed of 291 subjects (226 females, 65 males) recruited from Rockland County, USA. All participants provided informed consent and the study was approved by the Institutional Review Board at the Nathan Kline Institute (#226781 and #239708) and Montclair State University (#000983A and #000983B). The Cam-CAN dataset includes 358 (193 females, 165 males) subjects roughly uniformly distributed from Cambridge City, UK. All participants provided informed consent, and the study was approved by the local ethics committee, Cambridgeshire 2 Research Ethics Committee (reference: 10/H0308/50). In both datasets, we included only subjects within the DIRECT-PLUS age range (34–82 years). Exclusion criteria included unsuccessful completion of the preprocessing and quality control stages (see MRI preprocessing).

This work was based on a sub-study of the DIREC-PLUS trial (clinicaltrials.gov ID: NCT03020186↗). The primary aims of the DIRECT-PLUS trial were 18-month changes in VAT, intrahepatic fat, and adiposity across intervention groups. The results for the primary outcomes were presented in separate publications (Yaskolka Meir et al., 2021b). The DIRECT-PLUS was launched in May 2017 and conducted in an isolated workplace in Israel (Nuclear Research Center Negev, Dimona, Israel). Most clinical and medical measurements, including anthropometric measurements, blood drawing, and lifestyle intervention sessions, were performed on-site. Among 378 volunteers, 294 met age (30+ years of age) and abdominal obesity inclusion criteria waist circumference (WC): men >102 cm, women >88 cm (U.S. Department of Health and Human Services, 2013; Centers for Disease Control and Prevention, 2020) or dyslipidemia (TG >150 mg/dL and high-density-lipoprotein-cholesterol [HDL-C] ≤40 mg/dL for men, ≤50 mg/dL for women [Grundy et al., 2005]). Exclusion criteria were inability to perform physical activity (PA); serum creatinine ≥2 mg/dL; serum alanine aminotransferase or aspartate aminotransferase more than three times above the upper limit of normal; a major illness that might require hospitalization; pregnancy or lactation; active cancer, or chemotherapy treatment in the last 3 years; warfarin treatment; pacemaker or platinum implantation; and participation in a different trial. Among 294 eligible participants, 132 participants were randomly assigned to participate in the fMRI sub-study. The Soroka Medical Center Medical Ethics Board and Institutional Review Board provided ethics approval. All participants provided written consent and received no financial compensation.

All participants completed the baseline measurements and were randomized, using a computer-based program, in a 1:1:1 ratio, stratified by sex and work status (to ensure equal workplace-related lifestyle features between groups), into one of the three intervention groups: healthy dietary guidelines (HDG) as an active control group, Mediterranean diet (MED), green-MED, all combined with PA. Interventions lasted for 18 months and were contemporaneous, and participants were not blind to group assignment (open-label protocol). Each participant received complete dietary guidance (based on the specific intervention group) and a free and fully available clinical dietitians consult. Furthermore, all participants received free gym membership, including educational sessions encouraging moderate-intensity PA. Participants in both MED groups were assigned to a diet rich in vegetables, with poultry and fish replacing beef and lamb, with 1500–1800 kcal/day for men, 1200–1400 kcal/day for women. The diet additionally included 28 g/day of walnuts (+440 mg/day polyphenols provided). The green-MED group further consumed green tea (3–4 cups/day) and Wolffia globosa green shake (100 g/day frozen cubes, +1240 mg/day total polyphenols provided). A detailed description of the intervention outline is available in. Supplementary file 1

MRI scans were conducted at the Soroka University Medical Center (SUMC), Beer Sheva. Participants were scanned in a 3T Philips Ingenia scanner (Amsterdam, The Netherlands) equipped with a standard head coil. Subjects were instructed to refrain from food and non-water beverages 2 hr before the MRI sessions. Each of the two sessions at T0 and T18 included 2 RS-fMRI runs of 7 min each and a 3D T1-weighted anatomical scan to allow registration of the functional data. Before each RS session, participants were instructed to remain awake with their eyes open and lie still. fMRI BOLD contrast was acquired using the gradient-echo echo-planner imaging sequence with parallel acquisition (SENSE: factor 2.2). Scanning parameters were as follows: whole-brain coverage 41 slices (3 × 3 × 3 mm³), transverse orientation, 3 mm thickness, no gap, TR = 2200 ms, TE = 30 ms, flip angle = 90°, FOV = 200 × 222 (RL × AP) and matrix size 68 × 71 (RL × AP). High-resolution anatomical volumes were acquired with a T1-weighted 3D pulse sequence (1 × 1 × 1 mm³, 150 slices).

The preprocessing pipelines used in this work were extensively described in a previous publication (Levakov et al., 2021). T1w scans were preprocessed through FreeSurfer’s (Fischl et al., 1999) (version 6.0) recon-all processing. FreeSurfer’s cortical segmentation and spherical warp were used to transfer the Schaefer 100-node cortical parcellation (Schaefer et al., 2018) to each subject’s volumetric anatomical space. Functional images of the NKI dataset were preprocessed with fMRIPrep (version 1.1.8; Esteban et al., 2019) and images of the DIRECT-PLUS and Cam-CAN datasets were preprocessed with the Configurable Pipeline for the Analysis of Connectomes (C-PAC [Cameron et al., 2013] version 1.6.2). Briefly, both pipelines included the following steps: slice-timing correction, motion correction, skull stripping, estimation of motion parameters, and other nuisance signal time series. For the NKI dataset, functional scans were bandpass filtered (0.008–0.08 Hz) and confound regressed in a manner orthogonal to the temporal filters. Confounds included six motion estimates, the mean time series derived in CSF, WM, and whole-brain masks, the derivatives of these nine regressors, and the squares of these 18 terms. Spike regressors were added for each frame with framewise displacement above 0.5 mm. Data were linearly detrended and standardized. Nuisance regression in the DIRECT-PLUS and Cam-CAN fMRI dataset included the first five principal components of the signal from white matter and CSF (Behzadi et al., 2007), six motion parameters, and linear and quadratic trends, global signal regression, followed by temporal filtering between 0.1 and 0.01 Hz. Finally, a scrubbing threshold of 0.5 mm frame-wise displacement was applied (Power et al., 2014) (removal of 1 TR before and 2 TR after excessive movement). The time series of the two functional scans in the DIRECT-PLUS were concatenated to a single T0 and T18 scans. The exclusion criterion for excessive movements was determined a priori to less than 70% (9 min and 48 s) of the resting-state session after the scrubbing procedure (23% omitted; 102 subjects left). In all datasets, functional connectivity was defined as the Pearson’s correlation among pairs of ROIs’ time series followed by Fisher’s r-to-z transformation.

All parameters were measured at baseline and after 18 months of intervention. All clinical measures in the current study were selected a priori from a large pool of variables taken in the DIRECT-PLUS trial (Yaskolka Meir et al., 2021b). These measures were taken from five pre-selected categories: (1) Anthropometry that includes body mass index (BMI) and WC; (2) liver biomarkers that included AST, alanine transaminase (ALT), gamma-glutamyl transferase (GGT), ALKP, FGF 21, and chemerin; (3) glycemic markers, including glucose HOMA-IR and HbA1c; (4) lipids including cholesterol, HDL-C, LDL-C, and triglycerides; (5) imaging measures included liver fat, VAT, DSC, SSC, and the hippocampal occupancy score (HOC). WC was measured to the nearest millimeter halfway between the last rib and the iliac crest using an anthropometric measuring tape. Blood and urine samples were collected at 8:00 AM after a 12 hr fast. Blood samples were centrifuged and stored at –80°C. HOC was calculated as the hippocampal volume divided by (hippocampal volume + inferior lateral ventricle volume) in each hemisphere, then averaged across hemispheres (Kaplan et al., 2022; Heister et al., 2011).

Assessment of nutritional intake and lifestyle habits was self-reported online using validated food frequency questionnaires (FFQ) (Shai et al., 2005). The questionnaires were administered at baseline, after 6 months, and at the end of the trial. We selected a priori the questionnaire variables that were associated with brain age attenuation. These variables included the change in the following categories: sweets and beverages, weekly Wolffia globose intake, nut and seeds, eggs and milk, beef, processed food, green tea, and walnuts. The closed workplace enabled monitoring of the freely provided lunch and the intense dietary and PA sessions, which were provided simultaneously to all three groups.

To quantify and follow IHF% changes, we used H-MRS, a reliable tool for liver fat quantification (Kukuk et al., 2015). Localized, single-voxel proton spectra were acquired using a 3.0 T magnetic resonance scanner (Philips Ingenia, Best, The Netherlands). The measurements were taken from the right frontal lobe of the liver, with a location determined individually for each subject using a surface, receive-only phased-array coil. Spectra with and without water suppression were acquired using the single-voxel stimulated echo acquisition mode with the following parameters: TR = 4000 ms, TE = 9.0 ms, and TM = 16.0 ms. The receiver bandwidth was 2000 Hz and the number of data points was 1024. Second-order shimming was used. Four averages were taken in a single breath hold for an acquisition time of 16 s. The total image hepatic fat fraction was determined as the ratio between the sum of the area under all fat divided by the sum of area under all fat and water peaks (Hu et al., 2010).

Abdominal fat depots were assessed at baseline and 18 months thereafter using 3 T MRI scans (Ingenia 3.0T, Philips Healthcare, Best, the Netherlands). The scanner utilized a 3D modified DIXON imaging technique without gaps (2 mm thickness and 2 mm of spacing), fast-low-angle shot sequence with a multi-echo two excitation pulse sequence for phase-sensitive encoding of fat and water signals (TR, 3.6 ms; TE1, 1.19 ms; TE2, 2.3 ms; FOV 520 × 440 × 80 mm³; 2 × 1.4 × 1 mm³ voxel size). Four images of phantoms were generated: in-phase, out-phase, fat, and water phase (Thomas et al., 2013). Participants were instructed to hold their breath to avoid motion artifacts when their abdomen was scanned. A continuous line over the fascia superficialis was drawn to differentiate deep-SAT from superficial-SAT and calculated mean VAT, deep-SAT, and superficial-SAT along two axial slices: L5-S1 and L4-L5. We quantified fat mass regions as area and relative proportion of each fat subtype (percentage).

Subjects’ chronological age was predicted from the lower triangle of the functional connectivity matrix depicting all unique edges (4950 edges). We used a support vector regression model (Smola et al., 2000) implemented using Scikit-learn (Fabian et al., 2011) with a linear kernel and all the default parameters. Model accuracy was quantified as the Pearson’s correlation between the observed and predicted age. We additionally report the mean absolute error (MAE) in years, along with a p-value based on a non-parametric permutation test created by shuffling the data labels 1000 times (Hilger et al., 2020).

The primary outcome of the current work was brain age attenuation quantified as the difference between the expected and observed brain age at T18 (Yaskolka Meir et al., 2021a). The expected brain age at T18 was calculated by first producing brain age prediction for all participants at T0. Then, a linear regression was used to estimate brain age from the chronological age at T0. The fitted regression formula, representing the expected aging trajectory in the absence of intervention, was used to estimate the expected brain age at T18 given each participant’s T0 brain age and the time passed between the T0 and T18 MRI scans. The observed brain age was produced by applying the brain age model to the T18 scans. At baseline, brain age gap was computed as the difference between the predicted and observed age after regressing out the effect of the chronological age on the this gap (brain age bias correction [Smith et al., 2019]). We note that the result of computing the difference between the bias corrected brain age gap at both times was nearly identical to the brain age attenuation measure (r=0.99, p<0.001; MAE = 0.45). The difference between the two is because the brain age attenuation model takes into account the difference in the exact time that passed between the two scans for each participant (mean = 21.36 m, std = 1.68 m). Association between brain age attenuation and change in clinical measures following the intervention were reported using Pearson’s correlation. Correction for multiple comparisons was conducted within each biomarker category using the Benjamini–Hochberg false discovery rate (FDR; Benjamini and Hochberg, 1995) with an alpha of 0.05. Associations to food consumption reports were reported using Kendall’s tau correlation for ordinal data. Processed food at T18 had only two levels, ‘same consumption’ and ‘less consumption’, thus relation to brain age attenuation was tested with independent t-test. Change in clinical measurements were computed as a delta (Δ), the value at T0 minus the value at T18. We quantified change in reported food consumption as the change between the T0 and T18 questionnaires for food groups (i.e. processed food, sweets, and beverages) and as total consumption for polyphenols-provided foods (i.e. Mankai, green tea, walnuts). To control for the possible effect of age or gender, we used partial regression by regressing out the linear effect of age and gender from both brain age attenuation and the clinical measures. This was done by predicting each clinical measure, with the covariate as a predictor, keeping only the residual.

In a recent study (Zeighami et al., 2022), the authors reported a significant decrease in delta age 12 months following bariatric surgery. We converted the reported t statistic (t=3.66, p<0.001, df = 85) to an effect size (Cohen’s d=0.79, r=0.369) using an effect size calculator (https://lbecker.uccs.edu/↗). Using a sample size calculator (Python statsmodels.stats.power), we found that with an alpha of 0.05 and beta >0.05, a sample size >90 is required. With the given sample size (n=102), the probability of failing to reject the null hypothesis under the alternative hypothesis (β, Type II error rate) is 3%. Alternatively, for a Type II error rate lower than 0.1 with the given sample size, an effect size of 0.664 (Cohen’s d) is required.

To estimate chronological age from RSFC, we utilized data from 649 participants from two separate cohorts for the brain age model training, validation, and testing. We predicted chronological age from functional connectivity among the 100 nodes of the Schaefer brain atlas (Schaefer et al., 2018) (4950 edges) using a linear support vector regression model. The model was first trained and validated on 291 participants from the NKI dataset (Nooner et al., 2012; n=291) using fivefold cross-validation. As expected, a positive correlation was found between the predicted and observed age (r=0.439, p<0.001; MAE = 8.544, p<0.001). Next, we retrained the model on the entire sample and tested it in an independent sample from the Cam-CAN dataset (Shafto et al., 2014; n=358) again, yielding a positive correlation between the predicted and observed age (r=0.290, p<0.001; MAE = 11.402, p=0.005). Finally, we used the fitted model to estimate the brain age within the DIRECT-PLUS cohort. Of the 132 subjects that participated in the fMRI sub-study, 102 were included in all analyses after exclusions due to excessive in-scanner motion (23% omitted; MRI preprocessing). The predicted brain age and observed chronological age were correlated (r=0.244, p=0.013; MAE = 8.337, p<0.001; Figure 2), reproducing the results found within the two other datasets. Despite being significant and reproducible, we note that the correlations between the observed and predicted age were relatively moderate.

Baseline characteristics among the 102 participants with valid RSFC MRI scans are presented in Table 1 (see Supplementary file 2 for additional measures). The mean participant age was 51.5±10.5 years (median = 50.6, range 33.9–81.9), and 91.2% were men. The mean BMI and WC were 30.1±2.5 kg/m² and 107.1±6.6 cm, respectively. The mean baseline predicted brain age by RSFC was 52.8±4 years. In Table 1 we report the correlation between baseline characteristics and age, predicted age, and the brain age gap. The brain age gap was computed as the difference between the predicted and observed age after regressing out the effect of age on the this gap (brain age bias correction [Smith et al., 2019]). At baseline, brain age gap was correlated with chemerin (r=0.22, p=0.029), and with obesity-associated measurements obtained by MRI including visceral abdominal tissue (VAT): r=0.23, p=0.02 and superficial subcutaneous fat (SSC): r=−0.25, p=0.014.

Our primary hypothesis was that success in lifestyle intervention, as assessed by anthropometric measurements, will attenuate functional brain aging. Brain aging attenuation was quantified as the difference between the expected and observed brain age at T18 (Yaskolka Meir et al., 2021a; Figure 1e). Following 18 months of lifestyle intervention, participants showed a reduction of 0.76 (±1.86) units in BMI on average, 2.31 (±5.61) kg reduction in weight, and 5.39 (±5.89) cm reduction in WC. These constitute a –6.45% ± (5.60%) and –4.35% ± (5.86%) reduction from baseline for WC and BMI and weight, respectively. Additionally, at T18, the observed age was lower than expected in 56.8% of the subjects, while the opposite was found in 43.1% of the subjects (X²=1.922, p=0.166; see Figure 3, top). Importantly, we found a correlation between ∆BMI and brain age attenuation such that participants that showed a decrease in BMI also exhibited attenuated brain aging (r=0.319, p<0.001; Figure 3, bottom). Specifically, 1% of BMI or weight loss resulted in an 8.9 months’ attenuation of brain age (Figure 3—figure supplement 1). Similar results were found with ∆body weight (r=0.319, p<0.001) and ∆WC (r=0.198, p=0.046; Figure 4). The correlations to ∆BMI and ∆weight were significant after correcting for age and baseline brain age (p<0.05 for all), while the correlation to WC did not show a significant association (r>0.171, p’s <0.079 for all).

To examine the clinical outcomes associated with attenuated brain aging, we further tested the correlation of brain age attenuation with liver, glycemic, lipids, and MRI-assessed fat deposition biomarkers (Figure 4). Except of deep subcutaneous changes, all fat deposition measurements, superficial subcutaneous, visceral, and liver fat changes (e.g. loss) were significantly and directly associated with brain age attenuation (p<0.05, FDR corrected), i.e., the more the individual succeeded in diet-induced fat depots loss, the more brain age attenuation has been achieved. We then tested the association between brain age attenuation and liver and glycemic biomarkers. Out of all examined liver biomarkers, a decrease in ALT, GGT, alkaline phosphatase, and serum chemerin were significantly associated with attenuation in brain age (p<0.05 for all, FDR corrected). Of all examined lipid profile markers, only an increase in ∆HDL-C was significantly correlated with brain age attenuation (r=-0.273, p=0.005). Finally, a decrease in HOC was significantly correlated with brain age attenuation (r=-0.296, p=0.003). All results were reproduced after controlling for baseline age and predicted age at T0. However, after further correction for changes in BMI, only ∆alkaline phosphatase, ∆chemerin, and ∆HOC were associated with brain age attenuation (all p’s<0.018), with no significant associations with all other biomarkers (p>0.05, for all; see Supplementary file 3).

We examined whether food consumption, as reported using an FFQ, could be associated with functional brain aging attenuation. Associations were tested using Kendall’s rank correlation. We began with categories that could negatively affect brain aging attenuation. In line with our hypothesis, we found that decreased consumption of processed food (t=3.131, p=0.002) and sweets and beverages (τ=−0.231, p=0.002) was associated with more attenuation in brain age. An increase in green tea and walnut consumption, for which we hypothesized an attenuation effect on brain aging due to their high polyphenol content, did not result in a significant correlation (all τ’s<0.081, p’s>0.121; see Supplementary file 4 for all measures).