Hyperglycemia selectively increases cerebral non-oxidative glucose consumption without affecting blood flow

Multiple studies have shown that hyperglycemia increases the cerebral metabolic rate of glucose (CMRglc) in subcortical white matter. This observation remains unexplained. Using positron emission tomography (PET) and euinsulinaemic glucose clamps, we found, for the first time, that acute hyperglycemia increases non-oxidative CMRglc (i.e., aerobic glycolysis (AG)) in subcortical white mater as well as in medial temporal lobe structures, cerebellum and brainstem, all areas with low euglycemic CMRglc. Surprisingly, hyperglycemia did not change regional cerebral blood flow (CBF), the cerebral metabolic rate of oxygen (CMRO2), or the blood-oxygen-level-dependent (BOLD) response. Regional gene expression data reveal that brain regions where CMRglc increased have greater expression of hexokinase 2 (HK2). Simulations of glucose transport revealed that, unlike hexokinase 1, HK2 is not saturated at euglycemia, thus accommodating increased AG during hyperglycemia.


Supplemental Table 2:
The concentrations of plasma glucose and insulin during our study were not entirely steady after the desired plasma glucose level was reached (Figure 1).To quantify these observations, we performed a piecewise linear regression where the slope of the regression line was allowed to differ after 55 minutes into the glucose clamp (see Methods).Slope estimates from the piecewise regression are shown above.Values are means and symmetric 95% confidence intervals.All slopes are significantly different from zero at the 0.05 level without correction for multiple comparisons.
For the first 55 minutes, plasma glucose and insulin increased with time in both conditions (p < 0.0001).However, after 55 minutes plasma glucose decreased with time during both clamps (p < 0.0001), and plasma insulin decreased in the euglycemic condition (p < 0.0001).Plasma insulin continued to increase slightly with time after 55 minutes during the hyperglycemic clamp (p < 0.0001).
Supplemental Table 3 Supplemental Figure 1: Quantification of CMRglc with PET requires a corrected factor to account for differences in transport between [ 18 F]FDG and glucose.This correction factor is historically called the lumped constant.There is evidence that the lumped constant decreases slightly during hyperglycemia ).Koffn and Konn are the dissociation and association constants for each membrane.So, Se, Si, and Sc are the concentration of glucose in the blood, endothelium, interstice, and intracellular comparents respectively.Fhex is the irreversible breakdown of glucose.See Methods for an explanation of the simplifications that were performed to facilitate solving the system of differential equations.
to compute a rigid body transformation between the average label and average control images.A rigid body transformation was then computed between the realigned pCASL timeseries and each subject's average T1-weighted image using FSL's flirt.For subjects where field maps were acquired, the field map magnitude image was used as an intermediary between the pCASL and T1-weighted images.
Nonlinear registration between T1-weighted image and MNI152 space was performed using FNIRT in FSL (6).All transformations were then composed, and the pCASL time series was resampled to MNI152 2mm atlas space in a single step.
When computing the mean CBF for a single pCASL run, a weighting scheme was applied to down weight frames where subject motion produced large global shifts in image intensity (9)

Task MRI preprocessing
Runs whose average absolute or relative movement was greater than two standard deviations from the mean (absolute = 1.5 mm, relative = 0.36 mm) were excluded from all analyses.When field map images were available, geometric distortions in the phase encoding direction were removed using the FUGUE and PRELUDE tools (10).Rigid-body registration with a gray/white boundary-based cost function (11) was used to align the EPI images to each participant's T1-weighted anatomical image.
FSL's FNIRT was used to compute a nonlinear transformation between each participant's anatomical image and MNI152 space.
Each task run was modeled as events with a 2-second duration convolved with a double gamma hemodynamic response function intermixed with jittered periods of rest.The temporal derivative of this regressor was also included in the model.Runs within a subject and task (e.g., both encoding runs from one day for one subject) were combined using a fixed-effects model in FILM (12).To mitigate the effect of motion, fsl_motion_outliers was used to identity high motion frames.A frame was flagged as an outlier if the root-mean-square (RMS) difference between it and first frame in the run was greater than third quartile plus 1.5 times the interquartile range.A separate regressor for each outlier frame was added to the fixed effect model.After removing outlier frames there was no difference in average frame-to-frame mean squared error between hyperglycemia and euglycemia during either encoding (p=0.12) or retrieval (p=0.07).

CT imaging
For attenuation correction, a Siemens Biograph 40 PET/CT was used to acquire a CT image of the head (120 keV, 25 effective mAs, voxel size = 0.59 x 0.59 x 3.0 mm, acquisition matrix = 512 x 512 x 74 mm voxels).From the CT image, a µ-map was created by converting the CT Hounsfield values into attenuation coefficients (13).Subject-specific µ-maps were combined with PET/MR hardware µ-maps provided by the manufacturer (Siemens).

PET preprocessing
Dynamic image frames were generated using axial compression, four iterations of ordered sets with expectation maximization, Gaussian filtering to 4. In the second stage, we used a modified version of a previously published strategy to correct for between-frame motion (34).Briefly, for each scan each frame was registered to every other frame.
From this set of pairwise registrations, a linear system of transformation cycles was created from which it was possible to compute the least squares estimate of rigid body transformations between any two frames.These transformations were used to align the previously computed µ-map with the timeresolved PET reconstructions.The aligned µ-maps were used in the second stage of reconstruction to create time-sliced PET images with attenuation, decay, and scatter and additional motion correction.
The use of attenuation and scatter correction in the second stage allowed us to use shorter time bins for PET reconstruction.The first frame for the final reconstructed 15 O image began up three seconds prior to tracer inflow.Subsequent frames had progressively increasing durations which were inversely proportional to rate of radionucleotide decay (λ):∆ ~2! " .Because of variable inflow times, the number and duration of frames was variable across scans.The frame durations for [ 18 F]FDG increased by shorter proportional increments, beginning with a 10-second frame and ending with a 108-second frame.
After reconstruction, the motion corrected time series for each tracer was summed across time to create a single 3D PET emission image.Within individual participants, the sum images for each tracer were aligned to each other using rigid body registration (14).After alignment, the sum images were averaged to create a mean image for each tracer.The mean [ 18 F]FDG image was then brought into alignment with the T1-weighted image using rigid body registration and a contrast-preserving gradient vector-field algorithm (15).The final linear transformation was computed by minimizing the error between the forward ([ 18 F]FDG → T1) and backward (T1→ [ 18 F]FDG) transformations ( 16).

( 1 )Supplemental Figure 2 :Supplemental Figure 8 :Supplemental Figure 9 : 7 .Supplemental Figure 10 : 7 .Supplemental Figure 11 :Supplemental Figure 12 :
. To determine what impact changes in the lumped constant could have on our results, we computed the values shown in Figure2after adjusting for potential differences in the lumped attributable to glycemic level (see Methods).All figures follow the same conventions as Figure2.The focal increases in CMRglc were reported in Figure2were not affected by the lumped constant adjustment.However, the decreases in CMRglc that were found throughout gray matter in Figure2were greatly diminished after this correction.Identification of a reference region for relative PET SUVR measurements.A) Group-average CMRglc for 48 non-overlapping brain regions defined by FreeSurfer.Error bars are 95% confidence intervals.The black dot is the whole brain average, and the green dots are the three regions with the smallest absolute change in CMRglc during hyperglycemia (middle temporal, lateral orbitofrontal, and postcentral gyrus).For PET SUVR measurements these three regions were combined into a single reference region.B) CMRglc within the combined reference region ROI.There was no significant difference between conditions (0.06 ± 2.62 μMol•hg -1 •min -1 ; p = 0.96).Group average images of the difference in glucose consumption between hyperglycemia and euglycemia using C) quantitative CMRglc (n = 27 scans), D) whole-brain SUVR (n=46 scans), and E) SUVR using the combined reference region (n=46 scan).The maximum colormap value was matched to the 98% percentile value for each image to account for the difference in scale between the CMRglc and SUVR images.Because hyperglycemia increased the whole-brain CMRglc slightly (see Results), using a whole-brain reference region produced larger decreases in relative gray matter glucose consumption (D) than was seen with quantitative CMRglc (C).Using the empirically defined combined reference region (E) produced results much more consistent with quantitative CMRglc.Note that the SUVR data in D) and E) was not spatially smoothed in order to match the CMRglc data in C).Supplemental Figure3: A) Group average (n=26) image of the relative oxygen extraction (rOEF) measured [ 15 O]O2 and [ 15 O]H2O SUVR.Values are normalized to an empirically derived reference region (see Supplemental Figure 2).B) Group average (n=21) image rOEF during the hyperglycemic clamp.C) Group average difference in rOEF between the hyperglycemic and euglycemic clamp.Only voxels that are significantly different from zero after correction for multiple comparisons (False Discovery Rate 0.05) are shown in color.Voxels where rOEF decreased during hyperglycemia are in blue, whereas increases are shown in orange/yellow.Hyperglycemia does not affect activity evoked by a face-name retrieval task.A) Group average image of brain regions activated (red/yellow) and deactivated (blue/cyan) by the task.B) Average difference in BOLD activity between hyperglycemia (n=20) and euglycemia (n=21).C) None of the differences in B) were significant after correction for multiple comparisons using threshold-free cluster enhancement.Four ROIs were defined by placing 10 mm spheres on the task-activated regions in Supplemental Figure Average responses within the ROIs are plotted for the encoding task (second row) and the retrieval task (third row).No significant differences were found in any of the ROIs in either condition.However, there was a trend-level increase in the inferior frontal gyrus during retrieval C).Four ROIs were defined by placing 10 mm spheres on the task-deactivated regions in Supplemental Figure Average responses within the ROIs are plotted for the encoding task (second row) and the retrieval task (third row).No significant differences were found in any of the ROIs in either condition.However, there was a trend-level increase (i.e., decrease in deactivation) in the precuneus during encoding A).No effect of hyperglycemia on reaction time.A) During the encoding condition participants were shown a face-name pair and asked whether the name fit the face.There was no difference in reaction time between euglycemia and hyperglycemia for either response (p > 0.05).B) In the retrieval condition participants were shown an image of a face and asked if they remembered the face.Again, hyperglycemia did not change reaction time (p > 0.05).Model of glucose transport introduced by Barros et al. (2).Filled boxes are carriers with bound glucose and empty boxes are unbound carriers.Dashed lines represent membrane boundaries.Cout and Cin are the unbound carrier concentrations on either side of the membrane, while CoutS and CinS are the corresponding values for carriers bound with glucose.Rates constants for movement of the carrier across the membrane are denoted with an fxn where x=1 denotes a unbound carrier, x=2 a bound carrier, and n indicates the membrane (1, 2, or 3 The same procedure was used to align the [15 O] sum images to the [ 18 F]FDG sum image.The computed transformations were combined and used to resample each PET time series in MNI152 2mm atlas space.Nonlinear registration between each T1-weighted image and MNI152 space was performed using the Symmetric Normalization transformation model(17) in ANTS(18).Relative PET Standardized uptake values ratios (SUVRs) were computed using specific time windows for each tracer.For [ 18 F]FDG, a time window from 40 to 60 minutes post injection was chosen.A sixty second window, starting approximately after the bolus reached the brain, was used for the [ 15 O]H2O and [ 15 O]O2 data.A sixty-second window starting 180 seconds after the bolus reached the brain was used for the [ 15 O]CO scans.All SUVR images were computed in native space and then resampled to MNI-152 2mm space.To minimize the impact of vascular artifact on our SUVR measurements of oxygen metabolism, a voxelwise spatial regression was run using the resampled [ 15 O]O2 SUVR as a dependent variable and [ 15 O]H2O and [ 15 O]CO SUVR as independent variables (19).The [ 15 O]O2 SUVR was adjusted by subtracting from it the product of the [ 15 O]CO SUVR and its regression coefficient.An SUVR approximation of OEF (rOEF) was calculated by dividing the adjusted [ 15 O]O2 SUVR by the product of the [ 15 O]H2O SUVR and its regression coefficient.Finally, a SUVR estimate of the relative oxygen-to-glucose index (rOGI) was computed by dividing the [ 15 O]O2 SUVR by the [ 18 F]FDG SUVR.

.15•10 -2
3 mm FWHM, randoms corrections by delayed timing, and a novel voxel-based single scatter model implementing 3D scatter sinograms (33).To ensure the highest possible data quality, our PET reconstruction strategy consisted of two stages.In the first stage, the listmode data were scatter corrected and reconstructed without attenuation or scatter