Reproducibility of Three Whole-Brain N-Acetylaspartate Decline Cohorts in Relapsing-Remitting Multiple Sclerosis

Patients

Sixteen RRMS patients (12 women, 4 men, median age 38 [27–55] years) were recruited serially based on 4 criteria: 1) clinically definite (CD) disease,²⁴ 2) MS of less than 12 years’ duration, 3) no relapse experienced in the last 3 months, and 4) no contraindications to MR imaging. Their age, sex, disease duration, Expanded Disability Status Scale (EDSS) scores, and medications are shown in Table 1. All patients were briefed on the procedure they were to undertake and gave their Institutional Review Board–approved written consent.

Because months and even years may elapse between the 2 clinical events that confirm MS diagnosis,²⁴ we defined 2 disease durations: ΔY₁, the time from the second (confirming) event: 1)

and ΔY₂, the time from the first symptom (FS), which better approximates true disease duration: 2)

MR Imaging and Brain Volume

Patients’ brain volumes (V_B) were segmented from their T1-weighted sagittal magnetization-prepared rapid acquisition of gradient-echo images (TE/TR/TI of 7.0/14.7/300 ms, 128 sections 1.5 mm thick, 256² matrix, 210 × 210 mm² FOV) by using the MIDAS package.²⁵ Specifically, a “seed” region was placed in periventricular white matter. Following selection of all pixels at or above the gray matter intensity, a brain mask was constructed in 3 steps: 1) morphologic erosion, 2) recursive region growth-retaining pixels connected to the seed region, and 3) morphologic inflation to reverse the erosion. The masks were truncated at the foramen magnum to incorporate the brain stem and cerebellum but not the cord, and V_B was obtained by multiplying the number of pixels in the mask by their volume.

MR Spectroscopy and WBNAA Quantification

The amount of WBNAA, Q_NAA, was obtained in the same session. Following auto-shimming that yielded consistent 15 ± 3-Hz full-width-at-half-maximum head water linewidth, a nonlocalizing, nonecho (TR/TE of 10.0/0.0 seconds) ¹H-MR spectroscopy sequence yielded the whole-head NAA signal intensity.^26–28 Absolute quantification was done against a reference 3-L sphere of 1.5 × 10⁻² moles NAA in water. Patient and reference NAA peak areas, S_P and S_R, were integrated and Q_NAA was estimated: 3) where V_R^180° and V_S^180° are the voltages into 50 Ω of radiofrequency power needed for nonselective 1-millisecond, 180° inversion on the reference and patient, reflecting their relative receive sensitivities.

Finally, to account for natural variations in brain size, Q_NAA from Equation (3) was divided by the patient’s V_B to obtain the WBNAA concentration: 4) which is independent of brain size and is therefore suitable for cross-sectional comparison. The intra- and interpatient variability of such concentrations in healthy adults have been shown to be approximately ±0.6 mmol/L.^26,29

WBNAA Decline Rate Estimates and Group Partitioning

The original study estimated individual WBNAA decline rates with only one measurement by using 3 assumptions: 1) The second time point can be obtained by assuming each patient’s WBNAA to be the average of healthy controls (13.2 mmol/L)^23,26 before disease onset. 2) NAA loss begun at either CD or FS (ie, its duration, ΔY₁ or ΔY₂) is known. 3) WBNAA changes during ΔY₁ or ΔY₂ were linear. Consequently, the decline rate for the ith patient from either CD (R_1i) or FS (R_2i) was²³ 5a) or 5b)

Patients were classified as stable if their R was less than 0 mmol/L/y, moderate if between 0 and 1.7 mmol/L/y, and rapid if greater than 1.7 mmol/L/y, as shown in Figs 1 and 2.²³

• Download figure
• Open in new tab
• Download powerpoint

Fig 1.

Dot-plot distribution of individual patients as a function of average rate of WBNAA decline per year of CD disease duration (R₁), as defined by Equations (1) and (5a). The vertical dotted lines at R₁ = 0 and 1.7 mmol/L/y partition the cohort into stable (○), moderate (□), and rapid (▿) according to criteria defined in the text.²³

• Download figure
• Open in new tab
• Download powerpoint

Fig 2.

Same as Fig 1, except that the average individual rates of WBNAA decline per year, R₂, were estimated for disease duration from FS, ΔY₂, as defined by Equations (2) and (5b). The vertical dotted lines at R₁ = 0 and 1.7 mmol/L/y partition the cohort into stable (○), moderate (□), and rapid (▿) decline rates according to criteria defined in the text.²³

Statistical Analyses

Constrained least-squares regression was used to construct a model to predict WBNAA as a linear function of elapsed time from either CD or FS, subject to the first assumption above; that is, that at either time point of CD or FS, the WBNAA level was within the normal range of 13.2 ± 3 SD, 3 × 0.6 mmol/L.^26,29,30 This cutoff was chosen to ascertain 99% likelihood that a given patient was outside the normal WBNAA range. Comparison of 2 independent binomial proportions was used to calculate the similarity of group distributions between the current and previous datasets.

Number of Decline Subgroups.

Because any unconstrained number of data points can be divided into separate groups, no statistical methodology can objectively determine an optimal number of subgroups. Visual inspection of individual WBNAA levels plotted against disease duration for either ΔY₁ or ΔY₂ in Figs 3 and 4, however, readily reveals 3 such subgroups. This determination is bolstered by the resulting regression lines that so closely match the respective ones from the original study on a different patient cohort.

• Download figure
• Open in new tab
• Download powerpoint

Fig 3.

Individual WBNAA levels as a function of ΔY₁ (from clinical diagnosis) of Equation (1) (labeled according to Fig 1). Solid lines are the regressions calculated for each subgroup. The dashed lines are the ±95% confidence intervals calculated for the subgroups of the previous cohort.²³ Note how well the current regressions fit into the previous cohort’s confidence intervals.²³

• Download figure
• Open in new tab
• Download powerpoint

Fig 4.

Individual WBNAA levels as a function of ΔY₂ (from fist symptom) of Equation (2) (labeled according to Fig 2). Solid lines are regressions for each subgroup. The dashed lines are the ±95% confidence intervals calculated for the subgroups of the previous cohort.²³ Note how well the regression lines for the current patients fit into the previous cohort’s confidence intervals.²³

Subgroup Cross-Sectional Decline Rates.

For each subgroup, the linear prediction models derived from the current data fall within the 95% confidence intervals of the previous patients’ data (Figs 3 and 4). Accordingly, a test of 2 binomial proportions determined that the group WBNAA decline rates were not different from those in the original study (P > .05).

Subgroup Populations as Fraction of Total Cohort.

A test of 2 binomial proportions determined that the group size percentages of the current and previous cohorts, shown in Table 2, are indistinguishable, despite the smaller sample size in this study. In the CD groups, most patients are defined as moderate, with fewer being found in the stable and rapid groups. Although the partitioning from FS is more uniform, its similarity to the group distribution percentage of the original data is nonetheless statistically significant.

These observations are bolstered by the known epidemiology of MS (eg, as described by Weinshenker), which supports the notion of 3 levels of severity: “benign,” “intermediate,” and “acute”.³¹ Ascertaining these levels for individual patients, however, is currently possible only retrospectively, years after diagnosis. This delay poses serious limitations for several reasons. First, substantial WBNAA losses (exceeding 22%) were observed in monosymptomatic patients only suspected of having MS, indicating, together with irreversible tissue loss, that early widespread damage occurs even before a confirmed diagnosis of MS.^30,32,33 Second, early drug treatment after the first clinical symptom has been shown to delay conversion to CD MS,^32,34 underscoring the need to triage patients for treatment based on some surrogate marker, especially considering the absence of a clinical indicator. Third, since the administration and pace of treatment (type, dose, and frequency of administration) must be weighed versus cost and side effects, the a priori knowledge of future disease course could identify patients who would benefit most from therapy and help determine its level. Finally, recruiting patients who exhibit the most common form of the disease (as opposed to its mild or virulent forms) is important in setting up clinical drug trials. Identification of such patients could help avert errors of the kind where an ineffective drug is assumed efficacious due to the influence of “stable” patients (a type I statistical error) or when an effective drug is judged to be a failure because of the undue influence of rapidly declining patients (a much more expensive type II error).

Note that although 3 assumptions were used for the analyses above, we argue that these assumptions are reasonable. First, that the WBNAA decline started from approximately the average level of healthy controls (13.2 mmol/L)^26,29,30 is easily justified by the low variability in the latter (±6%–7%).^26,29,30 The second assumption, that disease duration is known (either ΔY₁ or ΔY₂), is more equivocal; that is, because MS etiology is unknown, so is its true time of onset. Nevertheless, since ΔY₂ approximates disease duration better than does ΔY₁, the minimal changes in subgroup traits based on either one (only one patient [no. 7] in Table 1 switched assignment) indicate that either is sufficient. The fact that the moderate and rapid subgroup regression lines do not intersect the normal range at t = 0 (ie, we take this intercept to be the “true” time of disease onset) is just an overall indicator of the “softness” of the approximation that the disease started at CD or FS. Furthermore, the distinction between ΔY₁ and ΔY₂ continues to diminish as radiologic criteria to confirm MS after a single clinical episode become ubiquitous.^35–37 Finally, the assumption that the WBNAA decline was linear over ΔY₁ or ΔY₂, though the least robust of the 3 assumptions, is still reasonable. Although De Stefano et al reported that NAA declines fastest early in the course of the disease and slows down with the transition from RR into the secondary progressive form,¹² that process may, in most patients, take many years, even decades. A linear approximation initially, therefore, is reasonable when patients are younger, when the future course of their disease is least well known, and when they are more likely to benefit from treatment.

Importantly, although this study estimates rates of decline based on the above-mentioned assumptions, these rates can only be validated in a longitudinal study. In such a study, each patient would be assigned an estimated rate category based on his or her WBNAA at entry, and his or her true individual WBNAA decline rate would then be measured over time. Although clearly important, such a study would require much more time and a larger cohort to establish statistical power of prediction. It would be even more interesting if the estimated WBNAA decline rate predicted initially would not only be validated longitudinally, but would also be correlated with the clinical disability accumulated by these patients over time.