We retrospectively identified 103 patients () from the Comprehensive Longitudinal Investigation of MS at Brigham study.11
Inclusion criteria: 1) age 18–60; 2) Brain MRI performed using our MS-designated protocol; 3) baseline EDSS scoring10
by an MS-neurologist within three months of MRI; 4) follow-up EDSS scoring 2.5 to 3.5 years later; 5) Established MS12
at baseline [relapsing-remitting (RR), secondary progressive (SP) or primary progressive (PP)].13
Clinically isolated syndromes were not included, as many of them will not develop MS.14
All patients except 12 were treated with immunotherapy during the observation period (). Baseline characteristics () were comparable to a large population-based MS cohort.15
This study was approved by our institutional review board.
Ninety-six patients (93%) had clinical follow-up 3.2 ± 0.3 years later. Follow-up disability was stable or progressed [1-point progression on follow-up EDSS if the baseline score was less than 6 (or 0.5 point progression if baseline was 6 or higher), sustained for three months] ().
Patients underwent 1.5T MRI with axial T1-weighted spin-echo (TR/TE: 725/20) and dual-echo T2-weighted (3000/80/30) images [voxel size 0.9375 × 0.9375 × 3 mm]. T1-imaging was repeated five minutes after 0.1 mmole/kg intravenous gadolinium. Our analysis, based on a hypothesis of a more informative composite measure, was tested without prior knowledge of the component MRI measures in the sample.
determined whole brain T2 hyperintense lesion volume (T2LV) and brain parenchymal fraction (BPF, an estimate of whole brain atrophy)17
from the dual-echo T2-weighted images using an automated template driven approach with partial volume correction. 16
Although FLAIR images would have likely yielded a higher lesion load than dual echo T2 images, we have not yet developed an automated FLAIR segmentation method and thus relied on our established method. T1 hypointense lesions (black holes) were the consensus of two trained observers,18
and showed at least partial hyperintensity on T2 images, but no gadolinium-enhancement (to reduce the likelihood of including transient lesions).19
TDS+ achieves an intra-class correlation of 0.994, inter-scan coefficient of variation (COV) of 4.98%, and volume bias of 0.01±0.68 mL.16
Our T1 hypointense lesion volume measurement shows intraobserver and interobserver COVs of 1.7% and 4.5%.18
In the current cohort, MRDSS achieved intraobserver and interobserver COVs of 2.3% and 4.4%.
All calculations necessary to derive MRDSS score from the MRI data were scaled using the cohort studied, not from external data. MRI data were rounded to two decimal places. When non-rounded data was used instead, the results did not change (data not shown). We did not use absolute T1 hypointense lesion volume for the MRDSS because of high collinearity with T2LV (r=0.79). Because the distributions of T2LV and T1/T2 were skewed, log (T2LV) or logistic (T1/T2) transformation was performed. T2LV, BPF and T1/T2 were then standardized (z
*) by subtracting the mean and dividing by standard deviation. The individual components of the MRDSS had only moderate inter-correlation with one another, both before and after transformations (r=0.53–0.58). Patients with zero T1 volume were not included in the normalization, but were assigned a value more extreme (−2.5) than the most extreme non-zero standardized T1/T2 (−1.92), similar to the MS Functional Composite.20
We selected the more extreme value because patients with zero T1 hypointensities have meaningfully less severe disease than patients with a small amount. The value −2.5 approximated the magnitude of the difference between the two smallest values on the standardized scale. The individual standardized scores were equally weighted and summed for each patient:
Each subject’s composite value was then transformed to a continuous 0–10 MRDSS score (zero is lowest severity).
Spearman correlation or the Wilcoxon test assessed the association between baseline MRDSS score and other variables. Univariate logistic regression tested the association between baseline MRI (BPF, T2, T1/T2 and MRDSS) and 3-year clinical progression. Multivariate regression controlled for co-variates. The area under the curve (AUC) in the receiver operator characteristic (ROC) curve investigated the predictive ability of the model. The 95% confidence interval for AUC was the percentiles of a bootstrap distribution.21,22
Comparing MRDSS to the raw MRI data demonstrated the value of the MRDSS over conventional approaches, while comparisons to standardized scores investigated the effect of combining measures into a composite scale. A p<0.05 was considered significant.