Impact of acquisition and modeling parameters on test-retest reproducibility of edited GABA+

Abstract

Literature values vary widely for within-subject test-retest reproducibility of gamma-aminobutyric acid (GABA) measured with edited magnetic resonance spectroscopy (MRS). Reasons for this variation remain unclear. Here we tested whether three acquisition parameters: 1) sequence complexity (two-experiment MEGA-PRESS versus four-experiment HERMES), 2) editing pulse duration (14 versus 20 ms), and 3) scanner frequency drift (interleaved water referencing (IWR) turned ON versus OFF), and two linear combination modeling variations: 1) three different co-edited macromolecule models (called ‘1to1GABA’, ‘1to1GABAsoft’, and ‘3to2MM’ in the Osprey software package) and 2) 0.55 versus 0.4 ppm spline baseline knot spacing affected the within-subject coefficient of variation of GABA + macromolecules (GABA+). We collected edited MRS data from the dorsal anterior cingulate cortex from 20 participants (30.8 ± 9.5 years; 10 males). Test and retest scans were separated by removing the participant from the scanner for 5-10 minutes. Each acquisition consisted of two MEGA-PRESS and two HERMES sequences with editing pulse durations of 14 and 20 ms (referred to here as: MEGA-14, MEGA-20, HERMES-14, and HERMES-20; all TE = 80 ms, 224 averages). We identified the best test-retest reproducibility following post-processing with a composite model of the 0.9 and 3 ppm macromolecules (‘3to2MM’); this model performed particularly well for the HERMES data. Furthermore, sparser (0.55 compared with 0.4 ppm) spline baseline knot spacing yielded generally better test-retest reproducibility for GABA+. Replicating our prior results, linear combination modeling in Osprey compared with simple peak fitting in Gannet resulted in substantially better test-retest reproducibility. However, reproducibility did not consistently differ for MEGA-PRESS compared with HERMES, for 14 compared with 20 ms editing pulses, or for IWR-ON versus IWR-OFF. These results highlight the importance of model selection for edited MRS studies of GABA+, particularly for clinical studies which focus on individual patient differences in GABA+ or changes following an intervention.

Graphical Abstract

Test and retest data (n=20) were collected for four edited MR spectroscopy sequences (MEGA-14, MEGA-20, HERMES-14, and HERMES-20) and post-processed in Osprey using several modeling options (0.4 vs 0.55 ppm spline baseline knot spacing and three models for co-edited macromolecules). GABA+ reproducibility was best when using 0.55 ppm spline baseline knot spacing and the ‘3to2MM’ model.

1. Introduction

Gamma-aminobutyric acid (GABA), the brain’s primary inhibitory neurotransmitter, is an important metabolite for normal brain function. GABA levels can be quantified in vivo noninvasively using proton magnetic resonance spectroscopy (¹H MRS). However, due to its low concentration within the brain and overlapping spectral signatures of higher concentration metabolites, spectral editing is required for accurate in vivo quantification of GABA at 3 Tesla. Prior work has reported differences in brain GABA levels across the healthy lifespan^1,2, as well as in various psychiatric^3-5 and neurodegenerative conditions^6-8.

One common spectral editing technique for GABA is MEscher-GArwood Point RESolved Spectroscopy (MEGA-PRESS⁹), which requires a relatively lengthy (e.g., 10-minute) acquisition to target one metabolite (e.g., GABA). Newer sequences allow for simultaneous editing of multiple low-concentration metabolites; for instance, the Hadamard Encoding and Reconstruction of MEGA-edited Spectroscopy (HERMES) sequence permits simultaneous detection of both GABA and glutathione (GSH, one of the most abundant antioxidant sources in the central nervous system) using a Hadamard-encoded editing scheme¹⁰. Due to the finite selectivity of the editing pulses, the edited 3-ppm GABA signal is confounded by a co-edited macromolecular (MM) resonance. Since these GABA and MM signals cannot be reliably separated, their composite ‘GABA+’ is commonly reported.

Reports of intrasubject variability in MRS-measured GABA+ vary widely. Reported test-retest within-subject coefficients of variation (CVs) vary from 1% to almost 30%^11-15. Specific reasons for this wide range of reproducibility remain unclear. Maximizing test-retest reproducibility is becoming increasingly critical, for example as edited GABA+ measures are being used as markers to differentiate patients from controls at a single cross-sectional time point^16-19 and as a longitudinal outcome measure for clinical trials^20-24 (often with methods limited to only one pre-intervention time point^20,22-24, rather than multiple pre-intervention time points to demonstrate a stable GABA+ baseline).

Several acquisition factors, including editing pulse duration and editing scheme complexity (MEGA-PRESS versus HERMES) may contribute to intrasubject variability in GABA+ measurements. Longer editing pulses are more selective (narrower bandwidth) and thus result in smaller contributions from co-edited MM signals²⁵ and greater sensitization to motion and scanner frequency drift^26,27. The extent to which 14 versus 20 ms editing pulses for MEGA-PRESS and HERMES sequences affect intrasubject reproducibility of GABA+ measurements has not been explicitly reported. HERMES interleaves four sub-experiments into the same acquisition to simultaneously edit for GABA+ and GSH¹⁰ compared to the two-step MEGA-PRESS scheme. One key part of post-processing is the alignment of sub-spectra before calculating Hadamard combinations in order to reduce subtraction artifacts in the GABA− and GSH-edited difference spectra^28-30. Aligning four HERMES sub-spectra which lack a common strong ‘reporter signal’ might be more challenging and variable than aligning two MEGA sub-spectra that share the same residual water peak. Though we previously reported an intrasubject CV of 9.9% for GSH measured with HERMES, compared with 8.8% for GSH measured with MEGA-PRESS¹⁴ (using the Osprey software package³¹ for data modeling), the extent to which HERMES might increase intrasubject variability of GABA+ is unclear.

Another potential factor influencing reproducibility is scanner frequency drift. Edited sequences such as MEGA-PRESS and HERMES are sensitive to scanner frequency drift^27,32, caused by instability of the superconducting magnet, gradient heating or cooling, or subject motion that can differ between individual scanners^27,33. This B₀ instability may lead to increased linewidths and imperfect subtraction of signals in the OFF and the ON spectra in edited MRS experiments, but also changes to editing efficiency. In order to address B₀ frequency drift during acquisition, a prospective frequency correction method called interleaved water referencing (IWR) was proposed and demonstrated to reduce between-subject variance in MM-suppressed GABA measurements³². Here we investigate whether IWR impacts intrasubject variability of GABA+.

Though our initial goals for this work were to characterize the effect of acquisition factors on GABA+ intrasubject variability, we decided post hoc to also examine the effects of modeling choices on GABA+ reproducibility. Our recent work¹⁴ reanalyzing the test-retest dataset reported in Prisciandaro et al. (2020)¹² found that linear-combination modeling with the Osprey software package³¹ significantly improved HERMES GSH intrasubject reproducibility over simple peak fitting in Gannet³⁴, reducing the within-subject CV from 19 to 9.9%. Here, in addition to replicating our finding of the benefit of Osprey over Gannet processing¹⁴, we aimed to examine how Osprey linear-combination modeling decisions affect GABA+ reproducibility. Among 61 GABA-edited MEGA-PRESS datasets with TE = 68 ms, we previously found that sparser spline knot spacing and inclusion of co-edited MMs in the linear combination model basis function significantly improved intersubject CVs³⁵. Although the spline baseline is needed to account for lipid contamination and poor water suppression, it can also be a potential source of overfitting if given too many degrees of freedom. Highly flexible baseline models are likely to absorb substantial portions of the edited 3-ppm GABA+ signal, leading to overall lower and more variable estimates. Sparser spline knot spacing (0.55 ppm over 0.4 and 0.25 ppm) better estimated the background in the GABA-edited difference spectra for TE = 68 ms edited GABA+ data³⁵; however, differences in the MM contribution for TE = 80 versus TE = 68 ms GABA+ data could interact with the spline baseline modeling. Thus, here we investigate if sparser spline baseline knot spacing similarly performs better for TE = 80 ms GABA+ data. Linear combination modeling of GABA-edited spectra is improved by modeling of the co-edited MM signal at 3 ppm (MM_3co)³⁵. Here we fit three different models that yielded the lowest intersubject CVs in the recommended wide (0.5-4 ppm) modeling range with sparse (0.55 ppm) knot spacing in our previous work³⁵, referred to in Osprey as ‘1to1GABA’, ‘1to1GABAsoft’, and ‘3to2MM’ (described in detail in the Methods). In the present work, we examined how these modeling decisions affected intrasubject reproducibility of GABA+ at TE = 80 ms, which has lower SNR due to reduced editing efficiency and greater T₂ relaxation of GABA and MM signals compared with TE = 68 ms GABA+ data and subtly different signal lineshapes. For completeness, we also examined the impact of acquisition and analysis factors on intrasubject reproducibility of HERMES-measured GSH estimates. We aimed to replicate our finding of better GSH test-retest reproducibility using Osprey compared with Gannet processing¹⁴ and to determine whether editing pulse duration or dense versus sparse spline baseline knot spacing differentially affects GSH compared with GABA+ reproducibility.

In summary, in the present work we acquired a new test-retest dataset to investigate the effects of sequence complexity (MEGA-PRESS versus HERMES), editing pulse duration (14 versus 20 ms), scanner frequency drift (IWR-ON versus IWR-OFF), MM model (1to1GABA, 1to1GABAsoft, and 3to2MM), spline baseline knot spacing (0.40 versus 0.55 ppm), and use of linear-combination modeling (Osprey versus Gannet simple peak fitting) on the test-retest reproducibility (within-subject CV) of GABA+. We predicted that less sequence complexity (MEGA-PRESS) and shorter editing pulse duration (14 ms) would result in better GABA+ test-retest reproducibility (lower within-subject CVs). We examined the effects of Osprey modeling choices post hoc and thus did not have hypotheses for these exploratory aims; however, in line with our previous work¹⁴, we predicted that Gannet processing would result in poorer reproducibility compared with Osprey.

2. Methods

2.1. MRS Acquisition

Twenty healthy adults (mean age: 30.8 ± 9.5 years; 10 males) provided their written informed consent. Participants completed two consecutive 45-minute scans, separated by brief (~5-10 minutes) removal from the scanner. All scans were conducted on the same 3.0 Tesla Philips Ingenia Elition MRI scanner (Philips Healthcare, The Netherlands) using a 32-channel head coil. For voxel positioning, we first collected a T₁-weighted structural MRI scan using the following parameters: compressed SENSE, TR/TE = 2 ms/2 ms, flip angle = 8°, slice thickness = 1.0 mm, 170 slices, voxel size = 1 mm³ isotropic, total time = 2 min 19 sec. Next, we acquired metabolite spectra from a 30 x 30 x 30 mm³ voxel in the bilateral dorsal anterior cingulate cortex (dACC; Figure 1). We positioned the dACC voxel on the mid-sagittal slice, just superior to the genu of the corpus callosum.

Fig 1. Voxel Placement.

MR spectra were acquired from the dorsal anterior cingulate cortex (dACC); each participant’s native space binary voxel mask for their test (left) and retest (right) scans was normalized to standard (MNI) space and overlaid onto the spm152 template. Warmer colors indicate areas of greater overlap between participants (color bar = number of subjects overlapped).

We collected four consecutive metabolite spectra per scan (i.e., 4 “test” and 4 “retest” spectra per participant): two GABA-edited MEGA-PRESS sequences (i.e., with 14 or 20 ms editing pulses) and two HERMES sequences (i.e., with 14 or 20 ms editing pulses, each of which edited for both GABA and GSH). The order of sequences was randomized and balanced across subjects. Common parameters for the four sequences included: TR/TE = 2000 ms/80 ms; 224 averages in total, total duration 7 min 44 sec; and MOIST water suppression (140 Hz bandwidth). The MEGA-PRESS sequences applied editing pulses at 1.9 ppm and 7.46 ppm; the HERMES sequences applied editing pulse lobes at 1.9 ppm and 4.56 ppm. Editing pulse duration was either 14 ms (referred to here as MEGA-14 or HERMES-14) or 20 ms (referred to here as MEGA-20 or HERMES-20). IWR was set to “ON” for 10 participants and to “OFF” for the remaining 10 participants (for both their test and retest acquisitions).

2.2. MRS Data Processing

MRS data were analyzed using the open-source analysis toolbox Osprey (v2.4.0; https://github.com/schorschinho/osprey/)³¹ within MATLAB R2021b. All analysis procedures followed consensus-recommended guidelines^26,36,37. Briefly, analysis steps included: loading the vendor-native raw data; removing the residual water signal using a Hankel singular value decomposition (HSVD) filter³⁸; eddy-current correction based on the water reference³⁹; and robust spectral registration⁴⁰ to align the individual transients within each sub-spectrum set (edit-ON and edit-OFF for MEGA-PRESS and sub-spectrum A (GABA-ON/GSH-ON), B (GABA-ON/GSH-OFF), C (GABA-OFF/GSH-ON), and D (GABA-OFF/GSH-OFF) for HERMES). Final alignment of the averaged MEGA-PRESS sub-spectra minimized the residual water peak in the difference spectrum before subtraction to generate the GABA-edited difference spectrum. The HERMES sub-spectra were aligned using three pairwise steps, adjusting the frequency and phase such that different target regions are minimized in the difference spectrum of each pair: the residual water was minimized for the GSH-OFF sub-spectra (aligning B and D), subsequently the 2 ppm tNAA signal was minimized to align the GABA-OFF/GSH-ON sub-spectrum C to the previously corrected sub-spectrum D, and finally the 3.2 ppm tCho peak was minimized to align the GABA-ON/GSH-ON sub-spectrum A to the previously corrected sub-spectrum C. The final GABA− and GSH-edited HERMES difference spectra were generated via Hadamard combination.

We modeled the metabolite spectra as detailed in our prior work^31,35,41, using a wide modeling range (0.5-4 ppm³⁵), either 0.4 or 0.55 ppm knot spacing, and custom basis sets for each sequence generated by our MRSCloud tool (https://braingps.mricloud.org/mrs-cloud⁴²). We generated a basis set for each of the four sequences using PRESS localization at TE = 80 ms with the sequence-specific timing and real RF pulse waveforms and 14- or 20-ms editing pulse durations. The basis sets consisted of 19 basis functions (ascorbate, aspartate, creatine (Cr), negative creatine methylene (−CrCH₂), GABA, glycerophosphocholine, glutathione (GSH), glutamine, glutamate, H₂O, lactate, myo-inositol, NAA, NAAG, phosphocholine, phosphocreatine, phosphoethanolamine, scyllo-inositol, and taurine). For the edit-OFF (MEGA-PRESS) and the sum spectrum (HERMES), 5 macromolecules (MM09, MM12, MM14, MM17) and 3 lipids (Lip09, Lip13, Lip20) were included as parameterized Gaussian basis functions⁴¹. The GABA-edited difference spectrum included parameterized Gaussian basis functions for the 0.9 and 3 ppm macromolecule signals³⁵. We modeled the co-edited MMs at 3 ppm using three different MM models³⁵:

• 1to1GABA: uses one single composite GABA + MM basis function by adding the edited GABA at 3 ppm and edited MM at 3 ppm (MM_3co) with a fixed 1:1 amplitude ratio. That is, the 1to1GABA model assumes that 50% of the 3 ppm GABA signal is attributed to coedited MM.

• 1to1GABAsoft: instead of using a hard 1:1 assumption for the relationship between GABA and MM_3co, this model uses separate GABA and MM_3co basis functions and imposes a soft 1:1 amplitude constraint on the ratio of the GABA and MM_3co basis functions during the optimization step.

• 3to2MM: uses separate basis functions for GABA and MM. The MM basis function in this model comprises both the MMs at 0.94 ppm (MM_0.94) and the MMs at 3 ppm (MM_3co), added in a 3:2 ratio. This model does not impose any amplitude assumptions or constraints on GABA.

Note that, although we model the MM signal in these ways, the primary outcome variable that we report is still “GABA+” (GABA+MM), as is standard for 3 T GABA-edited MEGA-PRESS data. This is necessary to compare the three MM models, as different levels of constraint on the GABA:MM ratio (ranging from none for 3to2MM to full in 1to1GABA) change the covariance between GABA and MM. Modeling of MM in some fashion is necessary to appropriately model GABA+-edited spectra, but the inclusion of two covarying components in the model does not assume that GABA and MM can be meaningfully resolved with this acquisition methodology.

Binary masks of the test and retest MRS voxels were reconstructed in subject space and coregistered to each participant’s corresponding test or retest T₁-weighted structural scan. We then segmented the structural scans using SPM12⁴³ and quantified metabolite levels with respect to the unsuppressed water scan. In order to investigate variance inherent to MRS (as opposed to additional variance introduced by factors such as tissue segmentation), similar to our prior test-retest work¹⁴, no further relaxation or tissue segmentation corrections were applied. For additional details, see Appendix A, in which we list all consensus-recommended parameters regarding our MRS data acquisition, processing, and quality⁴⁴.

We also processed all MRS data using Gannet³⁴; see Appendix B for details. We present comparisons between within-subject reproducibility of metabolite concentrations generated by Osprey versus Gannet as a supplemental analysis. Gannet analyses are presented with the intention of replicating our previous results showing better within-subject reproducibility using Osprey’s linear-combination modeling procedures compared with Gannet’s simple peak fitting approach¹⁴.

2.3. Statistical Analyses

We conducted all statistical analyses using R 4.0.0⁴⁵ within RStudio⁴⁶. First, as in a majority of previous GABA-edited MRS test-retest work^11-15, we estimated the between-scan intrasubject reproducibility of GABA+ estimates from each acquisition and modeling condition using within-subject CVs. These CVs provide a metric of within-subject measurement agreement that is independent of the range of values in the sample. We calculated within-subject CVs and confidence intervals using the root-mean method according to (Bland, 2006: https://www-users.york.ac.uk/~mb55/meas/cv.htm); the R function for this calculation is available at https://github.com/khupfeld/within-subject-cv. We focus the Results section on the impact of acquisition and analysis on within-subject CVs (the most commonly reported test-retest repeatability metric in the MRS literature). However, for completeness, in Appendix B1 we also report two other reliability metrics found in the MRS test-retest literature¹¹: Pearson correlation (i.e., the direction and strength of relationship between the test and retest values) and intraclass correlation coefficient (ICC, a metric of how reliably an instrument distinguishes between subjects, accounting for both the consistency of within-subject values between test and retest, as well as the change in group-average values between test and retest⁴⁷). Single-rater, absolute-agreement, two-way mixed effect ICCs^11,48 were calculated using the irr package⁴⁹.

3. Results

3.1. GABA+ Data Quality

Three out of 80 total test-retest datasets (one MEGA-14 and two HERMES-20 datasets) were excluded before statistical analyses due to incorrect editing pulse parameters, unacceptably large lipid signal, and failure of sub-spectra alignment during post-processing. Thus, data presented in the Results include MEGA-14 (n=19), MEGA-20 (n=20), HERMES-14 (n=20), and HERMES-20 (n=18). Full consensus-recommended data quality metrics are presented in Appendix A. Group mean spectra for each sequence are presented in Figure B1. Creatine (Cr) linewidths (overall mean 5.4 Hz) were well within the range of consensus-recommended standards, i.e., < 13 Hz for 3 T³⁷ and did not differ between test and retest for any sequence, indicating consistent data quality across conditions (Table B1).

3.2. GABA+ Within-Subject Reproducibility

Table 1 and Figure 2 indicate within-subject CVs and CV confidence intervals for GABA+ for each sequence and post-processing modeling option, and Figure B2 depicts test and retest GABA+ values for each participant for each condition. Overall, there was not a consistent impact of acquisition sequence or editing pulse duration on GABA+ within-subject CVs, though HERMES CVs were generally more variable across acquisition and modeling choices than were MEGA CVs: MEGA-14 (mean 13.4%, range 8.1-21.0%), MEGA-20 (mean 12.4%, range 9.1-18.6%), HERMES-14 (mean 22.4%, range 11.1-34.4%), HERMES-20 (mean 13.4%, range 5.9-18.0%). Though HERMES-14 generally yielded the poorest reproducibility of the four acquisitions, this was the case only when the HERMES-14 data were fit using the 1to1GABA (29.2%, 34.4%) and 1to1GABAsoft (18.8%, 29.9%) models and not when using the 3to2MM model (11.1%, 11.3%).

Table 1.

Within-subject CVs for GABA+

		MEGA-14	MEGA-20	HERMES-14	HERMES-20
0.4 ppm knot	1to1GABA	21.0% (11.7-27.2%)	15.3% (12.4-17.8%)	34.4% (26.0-41.1%)	18.0% (11.1-23.0%)
	1to1GABAsoft	16.6% (8.1-22.1%)	18.6% (11.7-23.6%)	29.9% (19.2-37.7%)	17.2% (9.8-22.2%)
	3to2MM	16.7% (11.6-20.6%)	9.9% (8.2-11.3%)	11.3% (7.4-14.2%)	11.4% (9.0-13.4%)
0.55 ppm knot	1to1GABA	8.1% (2.6-11.2%)	9.1% (4.6-12.0%)	29.2% (23.5-33.9%)	15.3% (7.7-20.2%)
	1to1GABAsoft	9.1% (6.5-11.1%)	11.2% (7.9-13.7%)	18.8% (14.0-22.6%)	12.7% (7.8-16.2%)
	3to2MM	9.1% (5.1-11.8%)	10.0% (7.8-11.8%)	11.1% (6.8-14.1%)	5.9% (4.6-6.9%)

Note. Within-subject coefficients of variation (CVs) and 90% confidence intervals (CIs) were calculated according to the root mean square method described by (Bland, 2006: https://www-users.york.ac.uk/~mb55/meas/cv.htm).

Fig 2. Within-subject CVs for GABA+.

Within-subject %CVs and 90% confidence intervals are shown for each acquisition and post-processing condition.

Indeed, across all four experiments, the 3to2MM model generally yielded the best reproducibility (mean 10.7%, range 5.9-16.7%) compared with the 1to1GABA (mean 18.8%, range 8.1-34.4%) and 1to1GABAsoft models (mean 16.8%, range 9.1-29.9%). This benefit of the 3to2MM model was particularly evident for the HERMES data. Across all conditions, 0.55 ppm spacing generally performed better (mean 12.5%, range 5.9-29.2%) than 0.4 ppm knot spacing (mean 18.4%, range 9.9-34.4%). The benefit of sparser knot spacing was somewhat larger for the MEGA-PRESS data (0.55 ppm mean 9.4%; 0.4 ppm mean 16.4%) compared with the HERMES data (0.55 ppm mean 15.5%; 0.4 ppm mean 20.4%). ICC values and Pearson correlation coefficients generally corresponded with within-subject CVs and provide further support for these conclusions (Tables B2.1-B2.2).

We encountered minimal drift in either IWR condition (IWR-ON: 0.029 ± 0.004 ppm; IWR-OFF: 0.040 ± 0.016 ppm) and drift did not differ between the IWR-ON versus IWR-OFF or test versus retest conditions (both p > 0.05; Table B3.1). For completeness, we present within-subject CVs by IWR condition in Tables B3.2-B3.3; however, given the lack of significant difference in scanner stability between IWR conditions, these results should be interpreted with caution.

3.3. GABA+ Results Using Gannet Modeling

For completeness and comparison with prior literature, Gannet modeling results are presented in Appendix B4. For each acquisition, within-subject CVs for GABA+ were higher (indicating poorer test-retest reproducibility) for Gannet compared with Osprey. This effect was strongest for 20 ms editing pulses (see Appendix B4 for details).

3.4. GSH Within-Subject Reproducibility

For completeness, we followed the same procedures to calculate within-subject CVs for GSH measured with the HERMES sequences (Appendix B5). Editing pulse duration did not have a clear impact on GSH CV, but contrary to the case for GABA+, less sparse 0.4 ppm spline baseline knot spacing resulted in lower CVs for GSH, though this effect was more pronounced for HERMES-20 compared with HERMES-14. Similar to GABA+, within-subject CVs for GSH were worse following Gannet compared with Osprey. See Appendix B5 for details.

4. Discussion

Our key findings are summarized in Table 2. While sequence complexity and editing pulse duration did not consistently affect GABA+ reproducibility, we identified the overall best test-retest reproducibility (lowest within-subject CVs) following post-processing modeling of co-edited MMs using the 3to2MM model and using sparser (0.55 ppm) spline baseline knot spacing. The 3to2MM model particularly benefited the HERMES data; 0.55 ppm baseline knot spacing benefited MEGA-PRESS somewhat more than HERMES data. Linear-combination modeling with Osprey resulted in better reproducibility for GABA+ than the simple peak fitting approach implemented in Gannet.

Table 2.

Key findings regarding GABA+ reproducibility

Factor				Result
Acquisition
Sequence complexity	MEGA-PRESS	HERMES		No difference
Editing pulse duration	14 ms	20 ms		No difference
Scanner frequency drift	IWR-ON	IWR-OFF		No differencea
Analysis
MM model	1to1GABA	1to1GABAsoft	3to2MM	Better reproducibility with 3to2MM, especially for HERMES
Spline baseline knot spacing	0.4 ppm	0.55 ppm		Better reproducibility with 0.55 ppm, especially for MEGA-PRESS
Overall modeling approach	Osprey (LCM)	Gannet (simple peak fitting)		Better reproducibility with Osprey

Note. This table summarizes our key findings regarding whether within-subject coefficient of variation (CV) for GABA+ differed based on the listed data acquisition and analysis factors. In cases where one acquisition or analysis factor yielded better reproducibility, this factor is bolded on the left side of the table; further details regarding the result is provided on the right side of the table. IWR = interleaved water referencing; MM = macromolecular; LCM = linear combination modeling.

^aDue to the lack of statistically significant difference in scanner stability between IWR conditions, these results should be interpreted with caution.

Overall, the more complex, 4-experiment HERMES sequence did not result in consistently different test-retest reproducibility of GABA+ compared with the 2-experiment MEGA-PRESS sequence. We anticipated HERMES CVs to be worse than MEGA-PRESS CVs for multiple reasons; for example, HERMES requires successful alignment of more sub-spectra and thus is more susceptible to artifacts from subject movement. Similarly, editing pulse duration did not consistently influence GABA+ reproducibility. We anticipated that shorter, less selective 14 ms editing pulses would result in less susceptibility to motion and scanner drift artifacts²⁶ and result in better GABA+ reproducibility. However, given that this was a healthy control cohort, we did not encounter significant motion in the participants. Moreover, we did not have significant frequency drift; thus, susceptibility to drift likely did not have a substantial effect on this cohort (though drift could have more influence in a patient or pediatric cohort or for scans acquired on a different scanner or during an EPI-heavy protocol).

HERMES-14 generally yielded the poorest reproducibility of the four acquisitions. HERMES-14 is an experiment that was included in this protocol to complete the 2x2 design, without any further development or optimization. It is thus perhaps not surprising that is performed poorly, as we had previously assumed that HERMES would not work with 14 ms pulses. However, both HERMES-14 and HERMES-20 reproducibility were poorest when the data were fit using the 1to1GABA and 1to1GABAsoft models compared with the 3to2MM model. HERMES CVs were also consistently more variable than were MEGA-PRESS CVs. It thus appears that modeling choices may exert a greater influence on reproducibility of HERMES GABA+ data than MEGA-PRESS data, and that the 3to2MM model may fit HERMES GABA+ data most consistently—though it remains unclear why specifically this is the case.

In addition to yielding the best reproducibility for HERMES GABA+, the 3to2MM model generally resulted in the smallest GABA+ CVs (best reproducibility) across all acquisitions. The 3to2MM model adds separate basis functions for GABA+ and MM (both MM_0.94 and MM_3co) in a 3:2 ratio and does not impose amplitude constraints on GABA+. Benefits of the 3to2MM model include the addition of the MM_0.94 peak, which provides a non-overlapped anchoring reference for the amplitude of the MM_3co peak and exploits the fact that MM profiles are relatively consistent across healthy participants and ages⁵⁰. In our prior work in MEGA-PRESS TE = 68 ms GABA+ data³⁵, we found some of the lowest intersubject CVs using the 3to2MM model; this model performed well without overfitting the data. Though we hesitate to make any definitive recommendations regarding the optimal modeling decisions from the current sample size, these TE = 80 ms GABA+ data suggest that 3to2MM is likely a good model to use for GABA+ across a variety of acquisition parameters, at least in the absence of pathological MM changes. It is not easy to say which model is more valid—our ability to estimate the correct GABA:MM ratio determines our ability to estimate the correct MM_3.0:MM_0.9 ratio. It is notable that the most reproducible model (3to2MM) has the greatest freedom (the GABA and MM components are not relatively constrained) to model signal at 3 ppm. It is also interesting that the 1to1GABA constraints perform poorly across both 14 and 20 ms editing pulse durations, as the relative contribution of MM to the GABA signal is lower (and more variable) for the more selective (20 ms) editing pulses.

Sparser 0.55 ppm spline baseline knot spacing yielded better GABA+ reproducibility overall, and particularly for MEGA-PRESS GABA+. This is similar to our prior report that 0.55 ppm knot spacing resulted in lower intersubject CVs for TE = 68 ms GABA+ data³⁵. It is logical that the sparser (i.e., more rigid) baseline better estimated GABA+, as the more rigid baseline did not tend to overfit the signal, i.e., did not incorrectly assign some of the GABA+ signal to baseline (as was the case for the less sparse 0.4 ppm baseline which tended to bend upwards into the GABA+ signal). Contrarily, we found that more flexible 0.4 ppm knot spacing yielded better reproducibility for GSH. This effect was particularly evident for HERMES-20 compared with HERMES-14 GSH; this finding will require further investigation but likely allows the model to better accommodate the less flat baseline behavior in GSH-edited spectra. The results here highlight primarily that spline baseline knot spacing decisions can greatly affect reproducibility of metabolite estimates and that this modeling decision should likely differ depending on the metabolite of interest for a given study.

We found better within-subject reproducibility for both GABA+ and GSH following linear-combination modeling in Osprey compared with simple peak fitting in Gannet. This is similar to our prior test-retest work¹⁴, which reported significant improvement in HERMES GSH test-retest reproducibility with Osprey (CV 9%) over Gannet (CV 19%). In the present work, we found this CV improvement with Osprey for both GABA+ and GSH and for both HERMES and MEGA-PRESS, albeit with worse overall reproducibility. In Song et al.¹⁴, we saw comparable GABA+ CVs to the present work but did not see significant improvement in HERMES GABA+ CVs between the two modeling approaches (Osprey CV 15%, Gannet CV 17%)—though this prior work used only the 1to1GABAsoft model and Osprey default 0.4 ppm spline baseline knot spacing, which was found here to perform worse for GABA+. While there were acquisition differences between the two studies (e.g., different scanner vendor) it remains unclear precisely why GSH reproducibility was overall worse in the present work compared with Song et al. The ability of low-n test-retest studies to truly characterize reproducibility is limited. Nonetheless, our prior and present results do tend to support implementing consensus-recommended linear combination modeling³⁶ for quantifying edited metabolites collected using various acquisitions.

There are several limitations to this work. The dACC is a moderately challenging region for MRS which is more susceptible to artifacts and low SNR than e.g., posterior cingulate cortex; however, we selected this region because of its clinical relevance to cognition⁵¹ and to match with our prior test-retest work^12,14. Here we compared the performance of one dense (0.4 ppm) and one sparse (0.55 ppm) spline baseline knot spacing; future work might benefit from testing additional spline baseline knot spacing settings, including those sparser than 0.55 ppm. Moreover, in the present study, we analyzed data using two analysis programs (Osprey and Gannet); while prior work has extensively compared agreement between different analysis algorithms for single timepoint edited GABA+ data⁵², future test-retest studies might consider making similar comparisons. Though (as anticipated) IWR-ON resulted in slightly better reproducibility compared with IWR-OFF, we did not encounter substantial scanner frequency drift in this dataset. Benefits of using IWR would likely become more evident for within-subject measurements taken over multiple days, on scanners with greater drift, or in patient cohorts with greater movement. There is no gold standard of metabolite level estimation for GABA+ to validate the results. Though the performance of MRS acquisition and modeling approaches is often judged by the amount of within-subject variance^11-15, lower variance does not necessarily reflect greater accuracy (but is one pre-condition for it). However, as these healthy control participants did not undergo any intervention between the test and retest phases of the experiment, CVs should predominantly reflect variance introduced by the acquisition and modeling parameters. Finally, this was a limited-sample-size dataset, and we therefore focus our interpretations on differences in descriptive statistics (primarily, differences in within-subject CVs). We designed this experiment with the intention of comparing acquisition parameters, but post hoc decided to additionally explore the impact of modeling decisions on test-retest reproducibility. Thus, given the sample size, we do not suggest that our findings should be interpreted as definitive recommendations for optimal GABA+ acquisition or modeling (although they are perhaps the best evidence available on which to make such decisions). Rather, we intend to highlight that reproducibility differs based on acquisition and modeling parameters and that such parameters could significantly affect interpretation of results in clinical work. Future studies in larger datasets should be conducted to provide more concrete recommendations for acquisition and modeling of edited GABA+ data at TE = 80 ms.

Abbreviations used

GABA

gamma-aminobutyric acid

MRS

magnetic resonance spectroscopy

MEGA-PRESS

MEscher–GArwood Point RESolved Spectroscopy

HERMES

Hadamard Encoding and Reconstruction of MEGA-Edited Spectroscopy

IWR

interleaved water referencing

GABA+

GABA + macromolecules

GSH

glutathione

MM

macromolecule

CV

coefficient of variation

dACC

dorsal anterior cingulate cortex

HSVD

Hankel singular value decomposition

Cr

creatine