Exploratory Multisite Magnetic Resonance Spectroscopic Imaging Shows White Matter Neuroaxonal Loss Associated with Complications of Type 1 Diabetes in Children

MRSI Image Processing in Children with T1D

We found 2D MRSI to be an effective tool for identifying localized tissue-weighted NAA/Cr differences in children with T1D across multiple sites. On an individual level, we identified higher NAA/Cr in the WM than in deep GM, consistent with prior studies in healthy adults.²⁴ On the population level, we found negative associations with complications of T1D in WM, specifically reduced NAA/Cr with chronic hyperglycemia and DKA, as discussed in following sections.

In children with T1D, MRS has been used as single-voxel spectroscopy in single-site case reports.^13⇓⇓-16 Single-voxel spectroscopy has the benefit of allowing investigators to place a voxel in tissue as desired to measure metabolite ratios in a single location in the brain with higher signal-to-noise ratios.²⁵ However, this approach makes it difficult to resolve tissue or region differences in metabolite ratios across the brain with a single acquisition or to retroactively analyze other locations that were not prospectively defined.²⁵ We demonstrate that 2D MRSI, even with a lower signal-to-noise ratio than single-voxel spectroscopy, when combined with weighted signal averaging, overcomes these limitations in pediatric T1D samples at multiple sites. This suggests that future, large-scale, multisite MRS studies that are better equipped to sample diverse participants from this population can benefit from 2D imaging.

As it stands, there is no widely accepted methodology for identifying region-specific metabolite ratios from 2D MRSI, largely due to 3 key challenges. The first is accounting for partial volume effects among different brain tissues. The second is accounting for magnetic field inhomogeneities, eddy currents, and other nonlinearities that may exist across large structures or spatial dependencies like those due to chemical shift displacement. The third is accounting for noise that can exist in every single step of MRSI analysis, including signal acquisition, frequency transformation, spectral baseline modeling, basis fitting, and peak computation. Existing techniques using linear regression or source separation methods have primarily focused on overcoming partial volume effects.^24,26,27 However, we note that these techniques are performed later in the MRSI workflow and risk compounding and propagating noise, yielding measurements with reduced meaning. Thus, we propose first performing the weighted averaging operation in the signal domain without explicit partial volume modeling to exchange increased tissue specificity for reduced noise artifacts. Accordingly, we emphasize that the results reported here are “weighted” by tissue as opposed to being absolutely “specific” to tissue. We note that this design choice was made in large part due to considerations for noisy and variable data that may arise from multisite studies. Furthermore, we note that no approach has yet successfully modeled spatial inhomogeneities, dependencies, and nonlinearities. Thus, in the future, we hope for sequences with higher signal-to-noise ratios and to extend this idea of regional mapping in the signal domain with posterior matrix inversions to better incorporate partial volume effects and to add direct consideration of spatial nonlinearities, extending emerging methodologies.^24,26,27

When computing metabolite ratios with LCModel, we leveraged “default” 3T basis sets provided by LCModel instead of computing “custom,” scanner-specific sets using pulse sequences and envelope parameters.^28,29 This design choice was made for the following reasons: 1) to simplify both the present and future studies across multiple sites where MRSI pulse sequences and envelope parameters may not be readily available or reliably queried from scanners (as was the case for sites B and C presently); 2) because custom basis sets leveraging techniques like density operator simulation or the Versatile Simulation, Pulses, and Analysis (VeSPA; https://www.opensourceimaging.org/project/vespa-versatile-simulation-pulses-analysis/) toolkit primarily impact multiplet signals, whereas NAA and Cr are largely singlet;^28⇓-30 and 3) because initial attempts using these custom sets did not dramatically affect results.

We used data from multiple sites acquired with different TE/TRs to maximize the sample size available for analysis. Thus, this approach required empiric calibration for site-wise effects via phantom studies for this pilot study. However, scanner calibration is an open problem, and many other calibration options exist. Careful consideration of site-wise effects should be used in future studies, including the use of matched TE/TR acquisitions and scanner hardware when possible.

WM NAA/Cr and Chronic Hyperglycemia

We found a negative association between WM NAA/Cr and hyperglycemia, as measured by HbA1c, that was not present in the deep GM. HbA1c is an average measure of chronic glucose control in the 3 months preceding the measurement.³¹ Thus, we interpret this negative association as evidence that prolonged hyperglycemia is associated with neuroaxonal loss in WM diffusely in children with T1D. This finding builds on prior DTI studies in this population that have suggested a relationship between hyperglycemia and WM biomarkers and represents a novel exploration of diffuse WM neurochemistry, as opposed to focal investigations performed in prior single-voxel MRS studies.^{8,13⇓⇓-16}

While small sample sizes can produce underpowered or overfit models, we are reassured that this effect was significant throughout the forward model selection process and was not suppressed by additional covariates. Furthermore, we found that the best model identified through model selection explained roughly 40% of the variance in the data, indicating high explainability in a noisy system with a small sample size. Thus, we hope these small-scale, exploratory results serve as evidence to motivate larger multisite studies to elaborate on these findings. Notably, larger-scale studies should further investigate the clinical implications of these findings and their associations with cognitive outcomes. Additionally, given that HbA1c is an average measurement of blood sugar across time, future studies could clarify which aspects of glycemic variability are most impactful, including time spent in the target glucose range, daily excursions, pre- and postprandial levels, and the number of hyper- or hypoglycemic episodes.³² Inclusion of continuous glucose monitoring data may allow these effects to be characterized.

We note the absence of an age effect in our results. Prior MRS studies have established age effects in NAA and Cr at low TEs (around 30 ms) in children.³³ However, different TEs, such as the long TEs used presently (>100 ms), affect different peaks differently, and changing NAA and changing Cr do not equate to changing NAA/Cr, especially in a small sample. Thus, to balance the need for modeling age against the differences between metabolite- and ratio-based analyses as well as the uncertainties of MRS at different TEs, we included age as a variable during model selection as opposed to explicitly acquiring a control group for age in this pilot study.

WM NAA/Cr and DKA

We found throughout the model selection process that participants with DKA at disease onset had a decreased WM NAA/Cr, similar to the HbA1c effect, and that the HbA1c effect was muted in the presence of DKA. However, we found that these effects were only revealed when considered jointly with and in the context of HbA1c via inclusion of an interaction term between the two. DKA is a well-known complication of T1D, marked by an acute elevation of blood sugar to dangerous levels coupled with the onset of acidosis due to the development of ketones, potential cerebral edema, and other metabolic changes.³⁴ Thus, these associations suggest that WM neuroaxonal integrity may be diffusely affected by episodes of DKA in a manner that outweighs chronic changes from prolonged hyperglycemia.

As mentioned previously, the primary limitation to this interpretation is the potential confounding by site. Differences in DKA frequency by site as well as site effects that may be larger in magnitude than glycemic effects make it difficult to untangle the two. Another limitation is that patients with frequent episodes of DKA may also have higher HbA1c levels.³⁵ Moreover, due to the multifactorial nature of DKA, it is unclear which of its aspects may be contributing to these findings. Furthermore, we investigated the presence of DKA at the onset of T1D, but due to sample size limitations, we were not able to account for DKA severity or subsequent episodes.

In short, we recognize that neurophysiologic changes in T1D are multifactorial and did not account for all possible variables in this pilot study. We note that this limitation is common in other neuroimaging studies examining brain structure and function in T1D, including those with larger sample sizes.³⁶ One large-scale study in which the number of DKA episodes, severity of DKA, hyperglycemia, hypoglycemia, and duration of T1D were simultaneously accounted for only considered behavioral outcomes without neuroimaging.³⁷ Despite our small sample size, however, we found significant effects that were not suppressed, even as additional covariates were added. Thus, these small-scale associations support that larger studies of WM health in T1D should do the following: 1) consider interactions between acute T1D complications, like DKA, and chronic glucose control, and 2) take into account balanced DKA groups per site and rigorous harmonization of different acquisition and scanner effects, as mentioned previously.

Last, we note that both the DKA and HbA1c effects were seen only in the WM and not in GM. This finding is consistent with other studies that have found reductions in WM volume in children 6 months after DKA and reductions in WM structural integrity.^38⇓-40 Thus, we would expect NAA/Cr to be reduced in WM related to neuroaxonal loss. Other studies have also found reductions in GM volume in individuals with T1D, and it has been posited that patients with long-standing T1D may have partial neuronal loss or dysfunction related to their disease.^41,42 Our findings may indicate that cerebral WM is more vulnerable to glycemic injury, and we suspect that a reduction in GM NAA/Cr would also be seen if our participants had long-standing T1D. Alternatively, the absence of effects in GM may simply be related to the small sample size in this pilot study.