Metabolite T₂ relaxation times decrease across the adult lifespan in a large multi-site cohort

Abstract

Purpose

Relaxation correction is crucial for accurately estimating metabolite concentrations measured using in vivo magnetic resonance spectroscopy (MRS). However, the majority of MRS quantification routines assume that relaxation values remain constant across the lifespan, despite prior evidence of T₂ changes with aging for multiple of the major metabolites. Here, we comprehensively investigate correlations between T₂ and age in a large, multi-site cohort.

Methods

We recruited approximately 10 male and 10 female participants from each decade of life: 18–29, 30–39, 40–49, 50–59, and 60+ years old (n=101 total). We collected PRESS data at 8 TEs (30, 50, 74, 101, 135, 179, 241, and 350 ms) from voxels placed in white-matter-rich centrum semiovale (CSO) and gray-matter-rich posterior cingulate cortex (PCC). We quantified metabolite amplitudes using Osprey and fit exponential decay curves to estimate T₂.

Results

Older age was correlated with shorter T₂ for tNAA, tCr_3.0, tCr_3.9, tCho, Glx, and tissue water in CSO and PCC; r_s = −0.21 to −0.65, all p<0.05, FDR-corrected for multiple comparisons. These associations remained statistically significant when controlling for cortical atrophy. T₂ values did not differ across the adult lifespan for mI. By region, T₂ values were longer in the CSO for tNAA, tCr_3.0, tCr_3.9, Glx, and tissue water and longer in the PCC for tCho and mI.

Conclusion

These findings underscore the importance of considering metabolite T₂ changes with aging in MRS quantification. We suggest that future 3T work utilize the equations presented here to estimate age-specific T₂ values instead of relying on uniform default values.

1. Introduction

Understanding brain changes across the healthy adult lifespan is critical for preserving brain health in the quickly aging global population and uncovering possible mechanisms of age-related neurological disease. Proton magnetic resonance spectroscopy (¹H MRS) is the only methodology that allows non-invasive measurements of endogenous brain metabolite concentrations. However, MRS data are often acquired at echo times (TEs) that are non-negligible compared to metabolite transverse relaxation rates (T₂). This results in T₂-weighting of the signal, such that metabolite amplitude changes associated with normal aging might be caused by changes in relaxation but misinterpreted as changes in metabolite concentration. This confound exists even for short-TE MRS, but is particularly a concern for J-difference-edited MRS¹, which relies on TEs of >65 ms (constrained by the duration of frequency-selective editing pulses and the J-evolution of target metabolites). Thus, recent consensus^1–3 suggests that it is critical to address the confound of T₂ relaxation (including for reference signals), particularly in studies of aging and neurodegeneration. However, despite this, most MRS quantification procedures (likely incorrectly) use static reference values, which assume that metabolite T₂ relaxation remains constant across the adult lifespan.

Prior work has reported varied relationships between age and T₂, primarily for the singlet resonances total N-acetyl aspartate (tNAA), creatine (tCr), and choline (tCho). A majority of prior work at 3 and 4 T has reported shorter metabolite T₂s with older age both for metabolites^4–7 and tissue water^4,7. A few studies^8,9 at 1.5 T have reported the opposite effect of longer metabolite T₂s with older age; however, one of these studies⁹ was complicated by overlap among water and metabolite signals, and the other⁸ examined only the frontal lobe and included only males in the sample. One study¹⁰, also at 1.5 T, found longer NAA T₂ in the centrum semiovale of older adults; however, this study utilized a linewidth-based approach which has not been validated or used again since publication in 2005. With the exception of work by Brooks and colleagues⁸, each of these prior studies involved comparison of discrete age groups (young versus older adults) rather than continuous sampling across the adult lifespan, and each used small sample sizes (all n < 20 per age group, with the exception of work by Deelchand and colleagues⁴ which included 32 young and 26 older adults). Therefore, in the present study, we leveraged a large, multi-site cohort in order to more comprehensively investigate whether metabolite and tissue water T₂ values differ across the normal adult lifespan, and to provide statistical models for calculating age-specific T₂ values for future integration into MRS quantification procedures.

2. Methods

2.1. Participants

101 healthy adults provided written informed consent to participate at one of two sites: the Johns Hopkins University School of Medicine (n = 51) and the University of Florida (n = 50). The sample included approximately 10 females and 10 males from each of the following decades: 18–29, 30–39, 40–49, 50–59, and 60+ years (Table 1). The Johns Hopkins University and University of Florida Institutional Review Boards approved all study procedures. Participants first completed the Montreal Cognitive Assessment (MoCA)¹¹, followed by a 1-hour MRI protocol. Of note, four individuals scored below the cut-off score of 23 out of 30 (i.e., indicative of possible mild cognitive impairment¹²). However, MoCA score was not an a priori exclusion criterion for this study. Moreover, each of these individuals scored 22 (just below the cut-off), and 3 of these 4 reported that English was not their primary language which can negatively impact MoCA performance¹³ (and it was not feasible to conduct the MoCA in a language other than English). Therefore, we presumed that cognitive impairment was not likely and opted to retain these individuals in the cohort and statistical analyses.

Table 1.

Participant Demographics

Variable	18–29 years	30–39 years	40–49 years	50–59 years	60+ years
Sex, #	10 F, 13 M	10 F, 10 M	10 F, 10 M	9 F, 9 M	11 F, 9 M
Age, Mean (SD),	24.2 (3.3)	35.0 (2.9)	44.2 (3.0)	55.0 (2.9)	67.5 (4.1)
[Min, Max], years	[18.7, 29.8]	[30.3, 39.9]	[40.0, 49.2]	[50.2, 59.8]	[60.8, 75.4]
MoCA, Mean (SD)a	27.9 (2.0)	28.0 (2.3)	27.6 (2.5)	26.3 (1.8)	27.7 (2.5)
[Min, Max]	[23, 30]	[23, 30]	[22, 30]	[23, 29]	[22, 30]

^aOne point was added to the MoCA score for n = 9, as these individuals reported having completed 12 years of education or less¹¹.

2.2. MRS Acquisition

All scans were performed using a 32-channel head coil on either the Johns Hopkins University 3 T Philips dStream Ingenia Elition MRI scanner or the University of Florida 3 T Philips MR7700 MRI scanner. For voxel positioning, we first collected a T₁-weighted structural MRI scan using the following parameters: MPRAGE, TR/TE 2000 ms/2 ms, flip angle 8°, slice thickness 1.0 mm, 150 slices, voxel size 1 mm³ isotropic, total time 2 min 46 sec. Next, we acquired TE series data from two 30 × 26 × 26 mm³ voxels: the white matter (WM) rich centrum semiovale (CSO) and the gray matter (GM) rich posterior cingulate cortex (PCC; Figure 1). Scan parameters for the TE series included: PRESS localization, TR 2000 ms, 8 logarithmically-spaced TEs 30, 50, 74, 101, 135, 179, 241, and 350 ms, 24 transients per TE sampled at 2000 Hz with 1024 points, and CHESS water suppression (115 Hz bandwidth). Within each series, the TE steps were neither interleaved nor randomized. We also collected a separate series of unsuppressed water reference data at each of the 8 TEs with the same parameters, but with 2 transients per TE and no water suppression. Of note, these voxel sizes and locations were selected to match those collected in our recent cohort of short-TE PRESS metabolite data in 102 individuals ranging from their 20s to their 60s¹⁴.

Figure 1. Voxel Placement.

TE series data were acquired from the centrum semiovale (CSO) and posterior cingulate cortex (PCC). Each participant’s native space binary voxel mask for their CSO and PCC voxels 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).

2.2. MRS Data Processing

MRS data were analyzed within MATLAB R2021b using the open-source analysis toolbox Osprey (v2.5.0; https://github.com/schorschinho/osprey/)¹⁵. All analysis procedures followed consensus-recommended guidelines^1,3. Briefly, analysis steps included: loading the vendor-native raw data (which had already been coil-combined, eddy-current-corrected, and averaged on the scanner at the time of data collection), removing the residual water signal using a Hankel singular value decomposition (HSVD) filter¹⁶, and modeling the metabolite peaks at each TE separately as described previously^15,17 using TE-specific custom basis sets. The basis sets were simulated by the MRSCloud tool¹⁸ (https://braingps.mricloud.org/mrs-cloud). MRSCloud using a localized 2D density-matrix simulation of a 101 × 101 spatial grid (voxel size 30 × 30 × 30 mm³; field of view 45 × 45 × 45 mm³) and vendor-specific refocusing pulse shape, duration, and sequence timings based on the MATLAB simulation toolbox FID-A¹⁹. The basis sets consisted of 18 basis functions: ascorbate (Asc), aspartate (Asp), creatine (Cr), negative creatine methylene (-CrCH₂), gamma-aminobutyric acid (GABA), glycerophosphocholine (GPC), glutathione (GSH), glutamine (Gln), glutamate (Glu), lactate (Lac), myo-inositol (mI), N-acetyl aspartate (NAA), N-acetyl aspartyl glutamate (NAAG), phosphocholine (PCh), phosphocreatine (PCr), phosphoethanolamine (PE), scyllo-inositol (sI), and taurine (Tau), as well as 5 macromolecule signals (MM09, MM12, MM14, MM17, MM20) and 3 lipid signals (Lip09, Lip13, Lip20) included as parameterized Gaussian functions¹⁷.

We extracted amplitudes for 6 metabolites of interest: tNAA, tCho, tCr_3.0 (Cr + PCr), tCr_3.9 (Cr + PCr - (-CrCH₂)), mI, and Glx (Glu + Gln). We multiplied each metabolite amplitude by Osprey’s internal MRSCont.fit.scale factor for each TE and participant to make the metabolite amplitudes directly comparable across TEs. This scaling factor is applied to the data to ensure an optimal dynamic range between the data and basis set during modeling. It is defined as the ratio of the maximum of the real part of the data and the basis set in the model range. Next, we used lsqcurvefit in MATLAB to fit monoexponential T₂ decay functions to the TE series metabolite amplitudes (Equation 1) in order to obtain the T₂ decay constant and an R² value of model fit for each participant for each metabolite.

$$y_i=A_i\mathrm{*}{e}^{\left(-\frac{TE}{T_{2}i}\right)}$$ [1]

In Equation 1, ${\mathrm{y}}_{\mathrm{i}}$ represents the metabolite amplitude, ${\mathrm{A}}_{\mathrm{i}}$ is a scaling constant, and ${\mathrm{T}}_{2\mathrm{i}}$ is the relaxation time to be calculated for the i^th subject. Lastly, we created a binary mask of the two MRS voxels in subject space, co-registered these masks to each participant’s T₁-weighted structural scan, and segmented the structural scans using SPM12²⁰, in order to calculate the volume fractions of white matter (fWM), gray matter (fGM), and cerebrospinal fluid (fCSF) in each participant’s voxels in subject space (for use in statistical models to control for cortical atrophy and for estimation of tissue water T₂).

We then repeated a similar procedure for tissue water. To estimate the water amplitudes, the unsuppressed water data at each TE was modeled using a linear combination model with a simulated water signal¹⁸. We used MATLAB’s lsqcurvefit to fit a biexponential decay function (Equation 2) to obtain tissue water T₂ and R² values for each participant.

$$y_i=A_i\mathrm{*}{e}^{\left(-\frac{TE}{T_{2wi,tissue}}\right)}\mathrm{*}0.4211\mathrm{*}(1-fCSF)+A_i\mathrm{*}{e}^{\left(-\frac{TE}{T_{2wi},CSF}\right)}\mathrm{*}0.5789\mathrm{*}fCSF$$ [2]

In Equation 2, ${\mathrm{y}}_{\mathrm{i}}$ represents the water amplitude, ${\mathrm{A}}_{\mathrm{i}}$ is a scaling constant, ${\mathrm{T}}_{2\mathrm{w}\mathrm{i},\mathrm{t}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{u}\mathrm{e}}$ is the tissue water relaxation time, and ${\mathrm{T}}_{2\mathrm{w}\mathrm{i},\mathrm{C}\mathrm{S}\mathrm{F}}$ is the CSF water relaxation time to be calculated for the i^th subject. fCSF is the fraction of CSF within the voxel for the i^th subject; 0.4211 weights the first term by the approximate molal concentration of water for non-CSF tissue (40/(40+55)), and 0.5789 weights the second term by the approximate molal concentration of water for CSF (55/(40+55)). The ${\mathrm{A}}_{\mathrm{i}}$ and ${\mathrm{T}}_{2\mathrm{w}\mathrm{i},\mathrm{t}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{u}\mathrm{e}}$ terms were unconstrained, and the ${\mathrm{T}}_{2\mathrm{w}\mathrm{i},\mathrm{C}\mathrm{S}\mathrm{F}}$ term was constrained to the range of 50–3000 ms^21–23. A data acquisition error occurred at the University of Florida site for the water data; therefore, we included only the Johns Hopkins University participants (n = 51) in statistical analyses of the water T₂ data.

2.3. Statistical Analyses

We conducted all statistical analyses using R 4.3.2²⁴ within RStudio²⁵. First, we calculated descriptive statistics (mean, standard deviation) by age group for the T₂ values for each of the 6 metabolites of interest and tissue water. Next, we examined the correlation between T₂ and age for each metabolite and voxel separately. As multiple variables did not satisfy the Pearson correlation normality assumption (Shapiro test p < 0.05), we instead report nonparametric Spearman correlations. To account for multiple comparisons, we applied the Benjamini-Hochberg false discovery rate (FDR) correction to the p-values for each voxel²⁶. Secondly, we ran a series of linear models, setting each metabolite T₂ as the outcome variable and age as the predictor: ${\mathrm{T}}_2={\mathrm{\beta }}_0+{\mathrm{\beta }}_1*(\mathrm{A}\mathrm{g}\mathrm{e}-30)$. We centered age around 30 years, so that the intercept (β₀) from this model would represent the predicted metabolite T₂ value at 30 years old, and the slope (β₁) would represent the change in T₂ for each year of age. The aim of this model was to provide an equation to calculate predicted T₂ value for a given metabolite given the age of a participant. As a follow-up analysis, we reran each of these linear models controlling for the potential effects of cortical atrophy with aging: ${\mathrm{T}}_2={\mathrm{\beta }}_0+{\mathrm{\beta }}_1*\left(\mathrm{A}\mathrm{g}\mathrm{e}-30\right)+{\mathrm{\beta }}_2\mathrm{*}\mathrm{T}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{u}\mathrm{e}$. As in our recent work examining metabolite T₁ changes with aging²⁷, we calculated cortical atrophy as the relative tissue fraction within the voxel, $\mathrm{f}\mathrm{G}\mathrm{M}/(\mathrm{f}\mathrm{W}\mathrm{M}+\mathrm{f}\mathrm{G}\mathrm{M})$. The purpose of this follow-up model was to ensure that cortical atrophy effects were not a major contributing factor to the observed T₂ relationships with age.

In addition, we conducted a series of paired t-tests (followed by FDR correction of the p-values²⁶) to examine differences in metabolite T₂ values between the CSO and PCC voxels. We also computed one linear mixed effects model per metabolite in which we set T₂ (across both the CSO and PCC voxels) as the outcome variable, age, voxel, and the interaction of age with voxel as the predictors, and a random intercept (${\mathrm{u}}_{\mathrm{i}}$) for each subject: ${\mathrm{T}}_2={\mathrm{\beta }}_0+{\mathrm{\beta }}_1*\left(\mathrm{A}\mathrm{g}\mathrm{e}-30\right)+{\mathrm{\beta }}_2\mathrm{*}\mathrm{V}\mathrm{o}\mathrm{x}\mathrm{e}\mathrm{l}+{\mathrm{\beta }}_3\mathrm{*}(\mathrm{A}\mathrm{g}\mathrm{e}-30)\mathrm{*}\mathrm{V}\mathrm{o}\mathrm{x}\mathrm{e}\mathrm{l}+{\mathrm{u}}_{\mathrm{i}}$. The primary aim of this model was to test for any Age*Voxel interaction effects (i.e., whether the age slope differed by brain region in any cases). The linear mixed effects model and random subject intercepts were necessary because this modeling approach structured the data as ‘repeated measures’ in which each participant had two measurements (CSO T₂ and PCC T₂).

3. Results

3.1. Data Quality

Creatine (Cr) linewidths were well within the range of consensus-recommended standards (i.e., < 13 Hz for 3 T³) for all spectra except one individual’s CSO voxel (33-year-old male, Cr linewidth = 14.1 Hz). In addition, for one participant (19-year-old male), the PCC voxel was mistakenly positioned at the wrong location. Thus, these datasets (1 CSO and 1 PCC) were excluded before any statistical analyses. Additional consensus-recommended data quality metrics are presented in Appendix A. Example single-subject spectra at each TE and decay functions for each metabolite are presented in Figure 2. The mean R² value across the whole cohort for the goodness of fit of the T₂ decay model was ≥0.80 for each of the 6 metabolites of interest and tissue water. Table 2 presents descriptive statistics by age group for the T₂ values.

Figure 2. Single-Subject Example Spectra and CSO T₂ Decay.

A PRESS spectra for each of the 8 TEs for the CSO (left) and PCC (right) voxels for a representative subject (48-year-old female). This representative subject was determined by averaging the R² values of model fit across the 6 metabolites and tissue water for each person in the Johns Hopkins cohort (as water data were unavailable for University of Florida subjects) and then taking the group median of this average R² value. B Example T₂ decay plots for the 6 metabolites and tissue water from the CSO voxel for the same representative subject. Blue points represent the metabolite amplitude at each TE, and the dark blue line represents the calculated T₂ decay curve.

Table 2.

T₂ Descriptive Statistics by Age Group

CSO Voxel
Metabolite	18–29 years Mean (SD)	30–39 years Mean (SD)	40–49 years Mean (SD)	50–59 years Mean (SD)	60+ years Mean (SD)
tNAA	298.9 (22.5)	283.8 (24.3)	267.9 (24.2)	267.9 (24.6)	260.6 (23.8)
tCr3.0	159.5 (10.6)	153.0 (8.8)	147.7 (11.5)	148.3 (8.9)	147.9 (8.6)
tCr3.9	145.0 (10.7)	133.9 (16.2)	133.0 (10.3)	133.9 (8.6)	132.9 (9.3)
tCho	230.1 (18.6)	216.3 (24.4)	203.0 (19.7)	204.5 (17.7)	207.8 (19.7)
mI	146.7 (14.0)	141.5 (17.2)	132.7 (11.7)	144.8 (16.0)	140.6 (19.0)
Glx	206.1 (40.6)	213.7 (36.7)	186.3 (29.6)	195.6 (42.1)	181.5 (30.3)
Tissue Watera	64.8 (2.1)	64.1 (2.3)	62.7 (2.6)	62.5 (3.1)	62.4 (2.5)
PCC Voxel
Metabolite	18–29 years	30–39 years	40–49 years	50–59 years	60+ years
	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)
tNAA	255.2 (22.6)	237.1 (12.8)	224.0 (15.3)	221.1 (16.7)	215.6 (14.7)
tCr3.0	155.5 (8.9)	149.6 (7.8)	144 (8.6)	145.3 (5.8)	145.8 (8.5)
tCr3.9	131.1 (9.1)	126.4 (5.2)	125.8 (9.6)	123.6 (7.4)	127.6 (10.0)
tCho	258.2 (25.2)	240.3 (13.3)	221 (15.9)	224.2 (20.2)	228.0 (20.3)
mI	166.8 (15.0)	157.1 (13.7)	157.5 (14.2)	158.0 (13.6)	159.5 (16.3)
Glx	166.8 (19.2)	156.7 (21.6)	164.8 (29.3)	162.9 (23.8)	147.6 (31.2)
Tissue Watera	54.8 (4.0)	55.2 (2.5)	52.1 (2.5)	49.7 (5.9)	50.7 (2.7)

Note. This table lists T₂ values in ms for the 6 metabolites and tissue water.

^aTissue water includes only the n = 51 Johns Hopkins subjects.

3.2. T₂ Relationships with Age

Older age was significantly correlated with shorter T₂ values for tNAA, tCr_3.0, tCr_3.9, tCho, Glx, and tissue water in both the CSO and PCC; Spearman r = −0.21 to −0.65, p < 0.05, FDR-corrected for multiple comparisons (Figure 3; Table 3). Age was most strongly correlated with tNAA T₂. Age did not correlate with mI T₂ for either voxel.

Figure 3. Metabolite T₂ Correlations with Age.

T₂ correlations with age are shown for the CSO (above, blue) and PCC (below, orange). Each point represents one participant. The bottom right corner of each plot indicates the Spearman correlation coefficient (r_s) and statistical significance of the FDR-corrected p-value for the correlation, *p<0.05, **p<0.01, ***p<0.001. The gray line and shading represent the linear model and 95% confidence interval for the model using age to predict T₂ (produced using the geom_smooth function in R).

Table 3.

T₂ Correlations with Age

Metabolite	CSO		PCC
	rs	pFDR-corr	rs	pFDR-corr
tNAA	−0.46	< 0.001***	−0.65	< 0.001***
tCr3.0	−0.37	< 0.001***	−0.43	< 0.001***
tCr3.9	−0.34	0.001**	−0.21	0.047*
tCho	−0.38	< 0.001***	−0.51	< 0.001***
mI	−0.07	0.477	−0.17	0.095
Glx	−0.34	0.001**	−0.24	0.022*
Tissue Watera	−0.30	0.035*	−0.46	0.001**

Note. This table presents the Spearman r values and FDR-corrected p-values for the correlation of age with metabolite T₂ in the CSO and PCC voxels separately;

^*p<0.05,

^**p<0.01,

^***p<0.001.

^aTissue water includes only the n = 51 Johns Hopkins subjects.

Next, we fit linear models to predict T₂ using age for each metabolite and voxel combination: ${\mathrm{T}}_2={\mathrm{\beta }}_0+{\mathrm{\beta }}_1*\left(\mathrm{A}\mathrm{g}\mathrm{e}-30\right)$; Table 4. As we centered age around 30 years old, the intercept (β₀) represents the predicted T₂ value for each metabolite at 30 years of age (as opposed to age 0 which would represent an unhelpful extrapolation). The slope (β₁) represents the change in the predicted value of T₂ for each year of life. For example, for tNAA in the CSO, the predicted T₂ value for an individual age 30 years would be: T₂ = 288.27 + (30–30)*−0.81 = 288.27 ms, while the predicted T₂ value for an individual age 50 years would be T₂ = 288.27 + (50–30)*−0.81 = 272.07 ms. (Note that these predicted T₂ values also correspond to the gray linear model lines plotted in Figure 3). The slope and intercept values listed in Table 4 can thus be utilized to calculate a predicted T₂ for any age in a WM- or GM-rich voxel. Table 4 only includes slopes for the metabolites which were significantly correlated with age.

Table 4.

Model Coefficients for Estimating T₂

Metabolite	CSO		PCC
	Intercept	Slope (β1)	Intercept	Slope (β1)
tNAA	288.27	−0.81	244.35	−0.89
tCr3.0	155.21	−0.25	151.67	−0.24
tCr3.9	139.88	−0.26	128.53	−0.10
tCho	220.32	−0.51	245.66	−0.73
mI	142.39		162.13
Glx	206.54	−0.68	165.20	−0.37
Tissue Watera	64.11	−0.05	54.38	−0.12

Note. This table presents the intercept and slope (β₁) values for the linear model using age to predict T₂ values: T₂ = β₀ + β₁^*(Age-30). Slope is listed only for the metabolites which were significantly correlated with age. As Age is centered around 30 years in the equation, the intercept (β₀) represents the predicted T₂ value for each metabolite at Age = 30 years. The slope (β₁) represents the change in the predicted value of T₂ for each 1-year increase in age beyond 30. This model can be used in future work to predict a T₂ value for a given age for these metabolites.

^aTissue water includes only the n = 51 Johns Hopkins subjects.

Older age was significantly correlated with greater cortical atrophy (calculated as $\mathrm{f}\mathrm{G}\mathrm{M}/(\mathrm{f}\mathrm{W}\mathrm{M}+\mathrm{f}\mathrm{G}\mathrm{M})$) in the PCC (r_s = −0.25; p = 0.010) but not the CSO (r_s = −0.04; p = 0.680). As a follow-up to the linear models presented in Table 4, we reran each model controlling for cortical atrophy with aging: ${\mathrm{T}}_2={\mathrm{\beta }}_0+{\mathrm{\beta }}_1*\left(\mathrm{A}\mathrm{g}\mathrm{e}-30\right)+{\mathrm{\beta }}_2*\mathrm{T}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{u}\mathrm{e}$ (see Supplementary Table B1). Including this metric of cortical atrophy in the model did not change the statistical significance of any T₂ relationships with age, with the exception of tCr_3.9 in the PCC (for which the age-T2 relationship became non-significant, p = 0.094). Independent of the associations between age and T₂, greater cortical atrophy was significantly associated with longer metabolite T₂ values for tNAA, tCr_3.0, and tCr_3.9 (CSO only), as well as mI and tissue water (CSO and PCC).

3.2. T₂ Differences by Voxel

Paired t-tests revealed differences in T₂ values by voxel for all metabolites (as shown in Figure 4). T₂ values were higher in the CSO than in the PCC for tNAA, tCr_3.0, tCr_3.9, Glx, and tissue water, and lower in the CSO for tCho and mI.

Figure 4. Metabolite T₂ Differences by Voxel.

T₂ differences by voxel are shown for the CSO (blue) and PCC (orange). Each point represents one participant. The asterisks indicate the statistical significance of the FDR-corrected p-value for the paired t-test, *p<0.05, **p<0.01, ***p<0.001.

As a follow-up analysis, we computed one linear mixed effects model per metabolite (across both the CSO and PCC) to test whether the age slope differed by brain region for any metabolites or tissue water. The Age*Voxel interaction was significant only for tissue water (p=0.007), indicating that, in all but one case, the relationship of age with T₂ did not differ based on brain region (Supplementary Table B2). However, for tissue water, the relationship of age with T₂ was stronger for the PCC than the CSO.

4. Discussion

Here we present the largest analysis to date examining metabolite and tissue water T₂ changes across the healthy adult lifespan. Among 101 adults ages 18–75 and across two sites, metabolite and tissue water T₂ values in both the CSO and PCC were generally significantly shortened with age, even when controlling for age-related cortical atrophy. Moreover, T₂ values were longer in the CSO, with the exception of tCho and mI which exhibited longer T₂ in the PCC. Taken together, these results align with the majority of prior work which also reported T₂ declines with normal aging (but in much smaller cohorts). Moreover, the finding of T₂ differences based on age and brain region highlights the importance of measuring subject-level T₂ during data acquisition or employing estimation methods (such as the statistical models provided here) for calculating age- and region-appropriate T₂ values.

Older age was correlated with shorter metabolite T₂ values for tNAA, tCr_3.0, tCr_3.9, tCho, Glx, and tissue water in both the CSO and PCC. This aligns with most prior research in smaller samples which similarly found shorter tNAA, tCr, tCho, and tissue water T₂s with older age^4–7. As also seen in these prior studies, we identified the strongest age association for tNAA. Most prior reports did not examine Glx. Deelchand and colleagues (2020) reported reduced mI T₂ in CSO and PCC in older age, whereas we did not find an association between mI T₂ and age. However, it should be noted that Deelchand and colleagues⁴ compared two age groups (rather than treating age as a continuous variable), and their older cohort (ages 70–83 years) extended beyond our upper age range.

The specific mechanisms underlying these observed T₂ changes with aging remain unclear. With the exception of tCr_3.9 in the PCC, each of the identified T₂ associations with age remained statistically significant when controlling for cortical atrophy, suggesting that age-related atrophy is not a major factor in these findings. Instead, as metabolites are largely intracellular (glial or neuronal)²⁸, their T₂ relaxation times are likely influenced by changes in the cellular microenvironment⁴, i.e. cellular morphology, metabolism, or myelination^4,7. It is well established that neurons undergo morphological changes during aging—such as reduction in soma size and loss or regression of dendrites and dendritic spines²⁹— alongside a parallel metabolic shift in astrocytes associated with increased neuroinflammatory response and changes in oxidative metabolism^30,31. Furthermore, degeneration of myelin sheath³² and loss of axonal fiber³² with advancing age, accompanied by debris (e.g., protein aggregates) and degraded myelin accumulation^33,34 reported in white matter and further supported by in vivo diffusion tensor imaging^35,36.

The observed T₂ changes could also be influenced by the gradual deposition of iron, particularly Fe³⁺, in the brain with aging. Although iron is present in the brain in multiple forms, the intracellular non-heme iron (i.e., ferritin) in tissue is thought to cause dephasing of the proton spins and thus a faster T₂ decay³⁷. Several studies observed a strong linear correlation between iron concentrations and transverse relaxation (R₂)³⁸ values both in vivo and in post-mortem healthy and Alzheimer’s disease brain tissue³⁷, suggesting that faster T₂ relaxation is related to age-related iron deposition³⁹. Whilst the precise contribution of each of these mechanisms is unclear, the observed age relationships suggest that T₂ measurements are sensitive to various parallel changes in the cellular environment⁷.

In this dataset, T₂ relaxation times were predominantly longer for tNAA, tCr_3.0, tCr_3.9, Glx, and tissue water in the WM-rich CSO, whereas tCho and mI exhibited longer T₂ in the GM-rich PCC. There is relatively limited literature considering GM/WM differences in metabolite T₂, and this prior work differs in voxel location, cohort, acquisition and quantification methodology, and statistical approach. Of the seven references we identified^6,40–45, five reported longer T₂ for NAA in WM as we did^6,40–42,45, while one revealed no significant tissue effect⁴³ and one showed the reverse effect⁴⁴. For tCr_3.0, three references found our result of longer T₂ in WM^6,41,42, three showed no difference^40,43,45, and the same study showed the reverse effect⁴⁴. Only a few studies have measured T₂s of tCr_3.9 or Glx. For tCr_3.9, one paper showed longer T₂s in WM⁴⁵ (as we found) and one no difference⁴²; for Glx, one study that separated Glx as Glu and Gln with J-PRESS found no differences by tissue type in the major component, Glu⁴². For tCho, four studies found no difference^40,41,43,45, two found longer T₂ in GM^42,44 (as we did), and one longer T₂ in WM⁶. T₂ of mI remains less investigated, but the two studies that measured mI T₂ also found longer T₂ in GM^42,44. For tissue water, GM is generally found to have longer T₂ values than WM in multi-echo MRI experiments^46–48, although the extent to which CSF confounds this result depends on resolution. Regional T₂ differences may relate to greater micro- and macro-structural organization in myelinated WM compared to GM⁴⁹; however, further work is needed to fully understand the mechanisms that govern metabolite T₂ relaxation.

We recently performed a meta-regression analysis of 75 manuscripts⁵⁰ containing 629 unique values to derive a general predictive T₂ model, with linear factors for: metabolite, field strength, species, tissue, pulse sequence, and Carr-Purcell Meiboom-Gill filter. The average bias between the (30-year-old intercept) values reported in the present study and the model predicted values was +9 ms (i.e., on average the model predicts shorter T₂s than measured here). The average absolute difference was 23 ms which is smaller than the average absolute difference between the predicted model and the T₂ training dataset (42 ms). On this basis, we assert that our results are consistent with the diverse T₂ literature.

2D modeling of interrelated MRS data has recently gained interest in the MRS community^51–53. Most notably, it was found that 2D modeling of synthetic multi-TE MRS data with overlapping peaks led to improved precision due to improved model parsimony achieved through reparametrization⁵¹. Applying 2D modeling to our in vivo datasets may improve the T₂ estimation of the metabolites reported here and could potentially allow for the T₂ estimation of additional low-SNR metabolites. However, it will also require careful reparameterization of the T₂ relaxation constants, lineshape estimates, and baseline terms, which will be part of future studies.

There are several limitations to this work. First, we acknowledge that our T₂ measurement here is a complex mix of pure T₂ and some inhomogeneous broadening factors that are not fully refocused by the two PRESS spin echoes. The goal of the present work was to improve the accuracy of T₂ relaxation correction in quantification procedures by understanding age effects on our measure of T₂ (rather than to accurately measure pure T₂). Second, future work could expand upon the age range to include those younger than 18 and older than 70 years, as well as targeting both normal and pathological aging (e.g., Alzheimer’s and other neurodegenerative diseases). Given the potential of T₂ to reflect both micro- and macrostructural organization, the measure may show utility as an early indicator of these changes, as suggested by Kirov and Tal⁵⁴. Though this was a large cohort with systematic recruitment across the adult lifespan, we only enrolled a few individuals older than age 70 years (the timeframe at which aging effects drastically accelerate). Lastly, we were limited to collecting only two voxels (WM-rich CSO and GM-rich PCC); however, prior evidence suggests that neurochemical changes with aging are highly region-dependent⁵⁵, and therefore future work might consider probing T₂ changes in other brain regions, or across the entire brain.

5. Conclusions

Consistent with prior literature, in a large multi-site cohort sampled systematically across the adult lifespan, we identified a clear age-related decrease in T₂ for multiple metabolites and tissue water, as well as differences in T₂ between the WM-rich CSO and GM-rich PCC. Together, these findings highlight potential changes in the brain’s cellular microenvironment with normal aging and underscore the critical importance of considering metabolite T₂ differences across the adult lifespan in MRS quantification procedures. We suggest that future MRS work leverage the models presented here to estimate age- and region-specific T₂ values instead of relying on uniform default values.