Confounder-adjusted MRI-based predictors of multiple sclerosis disability

Abstract

ABSTRACTIntroductionBoth aging and multiple sclerosis (MS) cause central nervous system (CNS) atrophy. Excess brain atrophy in MS has been interpreted as accelerated aging. Current paper tests an alternative hypothesis: MS causes CNS atrophy by mechanism(s) different from physiological aging. Thus, subtracting effects of physiological confounders on CNS structures would isolate MS-specific effects.MethodsStandardized brain MRI and neurological examination were acquired prospectively in 649 participants enrolled in ClinicalTrials.gov Identifier: NCT00794352 protocol. CNS volumes were measured retrospectively, by automated Lesion-TOADS algorithm and by Spinal Cord Toolbox, in a blinded fashion. Physiological confounders identified in 80 healthy volunteers were regressed out by stepwise multiple linear regression. MS specificity of confounder-adjusted MRI features was assessed in non-MS cohort (n=160). MS patients were randomly split into training (n=277) and validation (n=132) cohorts. Gradient boosting machine (GBM) models were generated in MS training cohort from unadjusted and confounder-adjusted CNS volumes against four disability scales.ResultsConfounder adjustment highlighted MS-specific progressive loss of CNS white matter. GBM model performance decreased substantially from training to cross-validation, to independent validation cohorts, but all models predicted cognitive and physical disability with low p-values and effect sizes that outperforms published literature based on recent meta-analysis. Models built from confounder-adjusted MRI predictors outperformed models from unadjusted predictors in the validation cohort.ConclusionGBM models from confounder-adjusted volumetric MRI features reflect MS-specific CNS injury, and due to stronger correlation with clinical outcomes compared to brain atrophy these models should be explored in future MS clinical trials.HighlightsRegressing out physiological confounders affecting volume of CNS structures in healthy volunteers, strengthened correlations between white matter volumes and disability outcomes in MS cohortsAggregating volumetric features into generalized boosting machine (GBM) models outperformed correlations of individual MRI biomarkers with clinical outcomes in MSDeveloped more sensitive and reliable models that predict MS-associated disabilityIndependent validation cohorts show true model performancesDeveloped GBM models should be explored in future MS clinical trials