Study population and design
UK Biobank (UKB) is a population-based biomedical initiative that aims to collect data from 500 thousand United Kingdom adults between 38 and 73 years of age26. In November 2021, raw and tabular data were accessed using an approved UK Biobank project. Figure 1 summarizes aspects of the current experimental design, namely how individuals were screened for eligibility to the current study. Participants were included if the following were available: T1-weighted MRI, resting-state functional MRI (rs-fMRI), ICD-10 medical history report, and blood biochemistry test reports. Individuals were excluded if required data were incomplete, previous medical reports indicating diabetes, head injury, or ‘unspecified brain disease,’ MRI scans were not available/collected, or failed MRI image quality as identified by the UK Biobank.
The following describes the methodology to create three groups based on obesity screening to match for age, sex, medical history, HbA1c, grey matter volume, and white matter volume, as well as nearly equal and large sample sizes (see Fig. 1). A total of N = 988 adults met study eligibility. Participants were initially sorted based on BMI > 30 and BMI < 30 (controls). The BMI > 30 individuals (n = 393) were split into two subgroups based on the sorted BMI ranking, and the first n = 200 with higher BMI ranks were assigned to the Group OHigh, while the remaining n = 193 were assigned to the Group OLow. The third group had a normal BMI, and n = 200 was randomly chosen from a list of 595 control participants. Some participants were later removed when the MRI quality check was deemed poor quality (e.g., raw MR image data were corrupted, incomplete, missing). We collected demographic information for each participant who passed screening, including age, sex, BMI, WHR, frequency of alcohol consumption, and medical history. We collected lipid-related data, including HbA1c, HDL-C, LDL, and TG, through each participant's blood biochemistry report. Finally, brain volume was estimated by adding total grey and white matter volumes from T1-weighted MRI.
Blood biochemistry data acquisition and lipid health factor creation
Peripheral HDL-C, LDL, TG, and HbA1c were measured from participant blood samples during the 2012–2013 assessment visit to the UK Biobank. TG level was measured using a standard clinical biochemistry assay (glycerol-3-phosphate (GPO)-peroxidase (POD) chromogenic method). In contrast, HDL-C was measured using the enzymatic selective protection method, HbA1c was measured using high-performance liquid chromatographic, and the LDL level was measured using the enzyme immunoinhibition method. A detailed UK Biobank biochemistry protocol is available to download at https://www.ukbiobank.ac.uk/
media/oiudpjqa/bcm023_ukb_biomarker_panel_website_v1-0-aug-2015-edit-2018.pdf.
We used blood biochemistry records (UK Biobank Data field 30760, 30780, and 30870 for HDL-C, LDL, and TG, respectively) from the first-repeat assessment visit data (estimated year of collection: 2012–2013). An omnibus principal component analysis (PCA) was performed using HDL-C, LDL, and TG data from all three groups. A singular principal component was used to create a lipid health score. The eigenvalue and the percentage-variance explanation were noted. The TG/HDL ratio was calculated for each participant as well.
Structural and functional MRI acquisition
We accessed rs-fMRI and the T1-weighted (T1w) data from the first imaging visit to UK Biobank (the estimated year of the collection starts from 2014). T1w images were acquired during a 5-minute scan that used a 3D MPRAGE gradient echo-planar imaging pulse sequence (repetition time/echo time/inversion time = 2000/2.01/880 msec, iPAT = 2, flip angle = 8°). The T1w images had a spatial resolution of 1x1x1 mm and a field-of-view of 208x256x256 mm. The rs-fMRI images were acquired during a 6-minute scan that used a gradient-echo echo-planar imaging pulse sequence (TR/TE = 0.735/39 msec, 8x multi-slice acceleration, no iPAT, and flip angle = 52°) with a spatial resolution of 2.4x2.4x2.4 mm and field-of-view of 211.2x211.2x153.6 mm. A total of 490-time point volumes were collected. DICOM data were converted into NIFTI format using a dcm2niix tool. Additional UK Biobank MR sequence parameters are available at the UK Biobank documentation https://biobank.ndph.ox.ac.uk/showcase/ukb/docs/bmri_V4_23092014.pdf
Image processing and fALFF calculations
In addition to the preprocessing protocol provided by UK Biobank (https://biobank.ndph.ox.ac.uk/showcase/ukb/docs/brain_mri.pdf), extra steps were done to calculate fALFF maps. Software from the Analysis of Functional NeuroImages (AFNI) and the FMRIB Software Library (FSL) packages were used for rs-fMRI preprocessing and fALFF calculations. Initial rs-fMRI volumes were removed, and images were de-obliqued to facilitate analysis. As per data cleaning recommendations27, BOLD timeseries were corrected for head motion, skull removal, de-spiking, detrending, spatially smoothing (FWHM = 6mm), and global intensity normalization.
For fALFF calculation, the square root of the power spectrum of each rs-fMRI timeseries voxel was computed to calculate the amplitude value for each voxel. Then, we calculated fALFF by summing the amplitude data in each voxel, which falls in the 0.01Hz – 0.1 Hz low-frequency range, and dividing by the sum of amplitude in the entire frequency spectrum. (Eq. 1)
\(fALFF=\frac{Sum. Amplitude (0.01 \tilde 0.1 Hz)}{Sum. Amplitude \left(whole frequency spectrum\right)}\) [1]
Brain structure segmentation
T1w and fALFF images were registered to standard space using the MNI-152 template. The FMRIB’s Integrated Registration and Segmentation Tool (FIRST) from FSL was used to segment the hippocampus on the T1w images28. The number of voxels in the segmented hippocampi masks was counted for volume calculation. The binary left/right hippocampus masks were used for the fALFF region of interest as the mean value in the mask. Estimated brain volumes were downloaded from tabular data (UK Biobank data field 25010).
Statistical analysis
T-statistics were used to test for group differences in demographics and blood lipid variables. A multivariable regression model tested for an association between hippocampus fALFF and lipid score (Eq. 2) along with the following covariates: age, biological sex, HbA1c, and total brain volume. Note that we also replaced the lipid health factor in Eq. 2 with the triglyceride-to-HDL ratio; this was a pre-defined sensitivity analysis. Participant age was reported by the corresponding assessment center. Biological sex was a categorical variable accessed from the UK Biobank registry at recruitment.
Mean hippocampus fALFF = Age + Sex + HbA1c + Brain volume + Lipid health factor [2]
The beta weights for each regressor were reported, with left and right hippocampus fALFF tested separately. A critical P-value of 0.05 was set to determine the statistical significance. The sensitivity analysis used the same regression model but replaced the lipid health factor by the triglyceride-to-HDL ratio.
The second voxel-wise analysis consisted of non-parametric permutation-based testing for each group separately. This model had the same parameters as the fALFF regional analysis (i.e., as in Eq. 2); the dependent variables were the fALFF voxel intensities. HbA1c was chosen to account for the effect of insulin resistance. Voxel-wise t-statistics and corresponding non-parametric p-value maps were calculated using the ‘randomise’ program in FSL, based on 10,000 permutations of the data29. Threshold-free cluster enhancement (TFCE) (H = 2, E = 5, C = 6) settings were used to produce a p-value corrected for multiple comparisons, and a critical corrected P-value < 0.01 was chosen. Significant clusters were displayed using the t-statistic map, thresholded by TFCE non-parametric P-value smaller than 0.01. Clusters inside the brain stem, cerebellum, and ventricles were excluded from the current analysis. A ‘cluster’ command from FSL was used to generate a summary of cluster size and location (the latter was deduced using the Harvard-Oxford atlas).