Peripheral insulin resistance attenuates cerebral glucose metabolism and impairs working memory in healthy adults
This study design, hypotheses and analyses were preregistered at OSF registrations (https://osf.io/93mnd).
Ethical considerations
The study protocol was reviewed and approved by the Monash University Human Research Ethics Committee in accordance with Australian Code for the Responsible Conduct of Research (2007) and the Australian National Statement on Ethical Conduct in Human Research (2007). Administration of ionizing radiation was approved by the Monash Health Principal Medical Physicist, following the Australian Radiation Protection and Nuclear Safety Agency Code of Practice (2005). For participants older than 18 years, the annual radiation exposure limit of 5 mSv applies. The effective dose in this study was 4.9 mSv. Participants provided informed consent to participate in the study.
Participants
Ninety participants were recruited from the general community via local advertising. An initial screening interview ensured that participants had the capacity to provide informed consent, did not have a diagnosis of diabetes, neurological or psychiatric illness. Participants were also screened for claustrophobia, non-MR compatible implants, and clinical or research PET scan in the past 12 months. Women were screened for current or suspected pregnancy. Participants received a $100 voucher for participating in the study. Eleven participants were excluded from further analyses due to blood haemolysis or well counter issues preventing insulin measurement or kinetic modelling (N = 7), excessive head motion (N = 2) or incomplete PET scan or image reconstruction (N = 2). The final sample included 79 individuals, 36 younger (mean 27.8; SD 6.2; range 20–42 years) and 43 older (mean 75.5; SD 5.8; range 66–86 years) adults (see Table 1). Exclusion criteria included a known diagnosis or history of diabetes reported by participants at the time of recruitment to the study.
Data acquisition
Cognitive battery
Participants completed an online demographic and lifestyle questionnaire and a cognitive test battery. The following cognitive measure were used:
Wechsler abbreviated scale of intelligence (WASI-IQ)
An assessment of intelligence suitable for ages 6–90 years53. There are 4 subtests: block design, vocabulary, matrix reasoning and similarities. WASI-IQ was scored by converting raw scores into a scale score, which were transformed into a composite score reflecting verbal comprehension and perceptual reasoning abilities (FSIQ2). This score was converted to an age-based T scores established in a normal population.
Hopkins verbal learning test (HVLT)
A three-trial list learning and free recall task comprising 12 words, four words from each of three semantic categories54. Approximately 20–25 minutes later, a delayed recall trial and a recognition trial was completed. The delayed recall required free recall of any words remembered. The recognition trial comprised 24 words, including the 12 target words and 12 false-positives, six semantically related, and six semantically unrelated. Delayed recall (total words recalled) and a recognition discrimination index (number of correct minus number of false positives in the recognition task) were calculated.
Digit span
A measure of verbal short term and working memory used in two formats: Forward and backward digit span55. Participants were presented with a series of digits, and are asked to repeat them in either the order presented (forward span) or in reverse order (backwards span). After two consecutive failures of the same length, the test was stopped. Scores were derived as the length of longest correct series for both forward and backward recall.
Task switching
A computer-based test in which participants were given a word and had to perform one of two simple categorisation tasks, depending on the cue that appeared with the word: (1) ‘living’ task. If the cue was a heart, participants were asked to categorise the word via a key press based on whether it represents a LIVING versus a NON-LIVING object; and (2) ‘size’ task. If the cue was an arrow-cross, participants were asked to categorise the word via a key press based on whether it represents an object that is BIGGER or SMALLER than a basketball. The cue selection for each new trial was randomised. Half the test trials were switch trials; half non-switch trials. Half the switch and non-switch trials was congruent in the key presses for either task, half was incongruent. The measures used included the percentage of correct switch trials and mean latency of correctly responding to a switch trial56.
Stop signal
A computer-based test in which participants were presented an arrow that pointed either right or left57. The task was to press the left response key if the arrow pointed to the left and press the right response key if the arrow pointed to the right, unless a signal beep was played after the presentation of the arrow. In this case the response should be stopped before execution. The delay between presentation of arrow and signal beep (starting at 250 ms) was adjusted up or down (by 50 ms) depending on performance. The delay increased if the previous signal stop was successful (up to 1150 ms) and decreased if the previous signal stop was not successful (down to 50 ms). The stimulus onset asynchrony between the start of each trial (onset of fixation circles) was 2000 ms. Variables were the mean reaction time in stop signal trials and stop signal reaction time. Stop signal reaction time is an estimate of inhibition ability, that is, the time required to stop the initiated go-process. The slower the stop signal reaction time, the more difficult to stop the go-process.
Digit symbol substitution
A computer-based task in which participant were presented with an 18 column × 16 row matrix58. The task was to translate symbols shown above the matrix (key) into digits in the matrix within a two minute period. Total count of correct responses and seconds per correct response were recorded.
MR-PET data acquisition
Participants underwent a 90-minute simultaneous MR-PET scan in a Siemens (Erlangen) Biograph 3-Tesla molecular MR scanner. Participants were directed to consume a high-protein/low-sugar diet for the 24 hours prior to the scan. They were also instructed to fast for six hours and to drink 2–6 glasses of water. Prior to FDG infusion, participants were cannulated in the vein in each forearm and a 10 ml baseline blood sample taken. At the beginning of the scan, half of the 260 MBq FDG tracer was administered via the left forearm as a bolus, providing a strong PET signal from the beginning of the scan. The remaining 130 MBq of the FDG tracer dose was infused at a rate of 36 ml/hour over 50 minutes, minimising the amount of signal decay over the course of the data acquisition. We have previously demonstrated that this protocol provides a good balance between a fast increase in signal-to-noise ratio at the start of the scan, and maintenance of signal-to-noise ratio over the duration of the scan59.
Participants were positioned supine in the scanner bore with their head in a 32-channel radiofrequency head coil and were instructed to lie as still as possible. The scan sequence was as follows. Non-functional MRI scans were acquired during the first 12 minutes, including a T1 3DMPRAGE (TA = 3.49 min, TR = 1640 ms, TE = 234 ms, flip angle = 8°, field of view = 256 × 256 mm2, voxel size = 1.0 × 1.0 × 1.0 mm3, 176 slices, sagittal acquisition) and T2 FLAIR (TA = 5.52 min, TR = 5000 ms, TE = 396 ms, field of view = 250 × 250 mm2, voxel size = 0.5 × 0.5 × 1 mm3, 160 slices) to image the anatomical grey and white matter structures, respectively. Thirteen minutes into the scan, list-mode PET (voxel size = 1.39 × 1.39 × 5.0 mm3) and T2* EPI BOLD-fMRI (TA = 40 minutes; TR = 1000 ms, TE = 39 ms, FOV = 210 mm2, 2.4 × 2.4 × 2.4 mm3 voxels, 64 slices, ascending axial acquisition) sequences were initiated. A 40-minute resting-state scan was undertaken in naturalistic viewing conditions watching a movie of a drone flying over the Hawaii Islands. At 53 minutes, pseudo-continuous arterial spin labelling (pc-ASL) began, and at 58 minutes, diffusion-weighted imaging (DWI) was acquired with 71 directions to index white matter connectivity. pcASL, DWI and fMRI results are not reported here.
Plasma radioactivity levels were measured throughout the scan. Beginning at 10-minutes post infusion onset, 5 ml blood samples were taken from the right forearm using a vacutainer at 10-minute intervals for a total of nine samples. The blood sample were immediately placed in a Heraeus Megafuge 16 centrifuge (ThermoFisher Scientific, Osterode, Germany) and spun at 2000 rpm (RCF ~ 515 g) for 5 minutes. 1000-μL plasma was pipetted, transferred to a counting tube, and placed in a well counter for four minutes. The count start time, total number of counts, and counts per minute were recorded for each sample.
MRI pre-processing and cortical thickness
For the T1 images, the brain was extracted in Freesurfer; quality of the pial/white matter surface was manually checked, corrected and registered to MNI152 space using Advanced Normalization Tools (ANTs). Cortical thickness for the Schaefer 100 regions was obtained from the Freesurfer reconstruction statistics for each participant.
PET image reconstruction and pre-processing
The list-mode PET data for each subject were binned into 344 3D sinogram frames of 16 s intervals. Attenuation was corrected via the pseudo-CT method for hybrid PET-MR scanners60. Ordinary Poisson-Ordered Subset Expectation Maximization algorithm (3 iterations, 21 subsets) with point spread function correction was used to reconstruct 3D volumes from the sinogram frames. The reconstructed DICOM slices were converted to NIFTI format with size 344 × 344 × 127 (voxel size: 1.39 × 1.39 × 2.03 mm3) for each volume. All 3D volumes were temporally concatenated to form a single 4D NIFTI volume. After concatenation, the PET volumes were motion corrected using FSL MCFLIRT61, with the mean PET image used to mask the 4D data.
Correction for partial volume effects
PET images were corrected for partial volume effects using the modified Müller-Gartner method implemented in PetSurf (https://surfer.nmr.mgh.harvard.edu/fswiki/PetSurfer)62,63. A grey matter threshold of 20–30% is recommended in ageing because atrophy can influence results62. For our analyses, we chose a 25% grey matter threshold and surface-based spatial smoothing63. We used a Gaussian kernel with a full width at half maximum of 12 mm to increase the signal-to-noise ratio. Subcortical structures were partial volume corrected and spatially smoothed in volume space and merged with the cortical data.
Cerebral metabolic rates of glucose
Calculations of regional CMRGLC were undertaken in PMOD 4.4 ( using the FDG time activity curves for the Schaefer 100 atlas parcellation and AAL subcortical structures. The FDG in the plasma samples was decay-corrected for the time between sampling and counting, and used as the input function to Patlak models. A lumped constant of 0.89 was used64, and equilibrium (t) set at 10 mins, the time corresponding to the peak of the bolus and onset of a stable signal65. The fractional blood space (vB) was set at 0.0566. Participant’s plasma glucose (mmol) was entered in the model from their baseline blood sample.
CMRGLC in the 17 networks was calculated from the regional CMRGLC values for each participant. Because the regions within a network differ in cortical volume, the regional CMRGLC values could not simply be averaged. Rather, each regional CMRGLC value was weighted by the percentage its volume represented within the total network cortical volume. An overall subcortical CMRGLC value was calculated by weighting each structure by the percentage its volume represented from the total volume of the subcortical structures.
HOMA-IR
A blood sample taken prior to FDG infusion was used to collect 2 ml of plasma for insulin and glucose measurement, which was undertaken by a commercial laboratory. HOMA-IR was calculated as fasting glucose (mmol/L) x fasting insulin (µU/ml) / 22.567. The constant of 22.5 is a normalising factor for normal fasting plasma insulin and glucose (i.e., 4.5 mmol/L x 5 μU/ml = 22.5). Higher HOMA-IR values indicate greater insulin resistance. We also calculated HOMA-IR2 ( We compared the relationship between HOMA-IR and HOMA-IR2 with CMRGLC and found minimal to no differences (see Supplementary Tables 5 and S11). This was expected as HOMA-IR2 models increases in the insulin secretion curve for plasma glucose concentrations above 10 mmol/L68; a threshold that than none of our participants reached. Hence, the results reported here are based on HOMA-IR.
Data analysis
The CMRGLC and HOMA-IR data was inspected for and found to be satisfactory for assumptions of normality and potential impact of any outliers (see Supplementary Fig. 1).
Age, cortical thickness, HOMA-IR and CMRGLC
Hypothesis 1: Independent sample T-tests were run to test hypothesis 1 that older people would have greater insulin resistance and lower cortical thickness than younger people.
Hypothesis 2 and 3: A series of general linear models (GLMs) was run in which regional CMRGLC was the dependent variable and age group, HOMA-IR and age group x HOMA-IR were the predictors. Cortical thickness in the same region was included as a covariate. For the subcortical structures, whole brain average cortical thickness was used as the covariate. The age group main effect was used to assess hypothesis 2 that older people would have lower regional cerebral metabolic rates of glucose than younger people, even after adjusting for lower cortical thickness in older people. The age group x HOMA-IR effects were used to test hypothesis 3 that greater insulin resistance would be associated with lower cerebral metabolic rate of glucose and that this association would be moderated by age, with the effect being stronger in older adults. For significant age group x HOMA-IR effects, post-hoc GLMs were run separately for younger and older adults. A series of GLMs was also run for CMRGLC at the 17 network level with the same design.
Partial eta squared (η2p) was used to quantify the effect sizes in the GLMs. Each series of analyses was also FDR-corrected at p < 0.05 for the overall GLM. We also calculated the percentage change in regional CMRGCL from a 10% change in HOMA-IR from the slope of the regression lines for younger and older adults separately.
CMRGLC and cognition
Hypothesis 4: We applied data reduction techniques to reduce the dimensions in the cognitive test data. The cognitive scores were converted to Z-scores and entered in a Principal Component Analysis (PCA). Principal components (PCs) with eigenvalues greater than one were retained and subject to varimax rotation to optimally reduce dimensionality69. Participant component scores were saved for further analyses.
A series of GLMs was run to test hypothesis 4 that greater cerebral metabolic rates of glucose and lower insulin resistance would be associated with better cognitive test performance. The five cognition PCs were entered as the dependent variable, with CMRGLC in the 17 networks entered as independent variables, together with age group, whole brain cortical thickness and HOMA-IR. To test for a moderating effect of HOMA-IR, where a network CMRGLC predicted a principal component of cognition, a product term was created between CMRGLC and HOMA-IR. Stepwise regression was used with the significant CMRGLC network(s) entered in block 1, and the product term(s) with HOMA-IR in block 2. A significant increase in variance explained from block 1 to block 2 was indicative of moderation.
link
