Author Of 1 Presentation
P0012 - Machine-learning optimized Combinatorial MRI scale (COMRISv2) correlates highly with MS disability (ID 1438)
Abstract
Background
Volumetric biomarkers derived from brain MRI correlate only mildly to moderately with disability scales in multiple sclerosis (MS) patients. We previously addressed this issue by employing machine learning (ML) to select semi-quantitative MRI (semi-qMRI) features and their weights in the Combinatorial MRI Scale (COMRISv1). COMRISv1 correlated strongly with physical and cognitive disability in an independent validation cohort.
Objectives
Building on this work, we aimed to test the hypothesis that more powerful ML algorithm (i.e., Random Forests; RF) to both fully quantitative (qMRI) and semi-qMRI biomarkers in COMRISv2 will outperform COMRISv1 in the ability to predict physical and cognitive disability in an independent cohort of MS patients.
Methods
The prospectively acquired MS patients (n=283, divided 2:1 into training and validation cohorts) underwent brain MRI imaging within days of clinical evaluation. Neurological examination was transcribed to NeurEx app that automatically computes disability scales. Semi-qMRI features were determined weekly by consensus of MS-trained neurologists, while qMRI features were computed by lesion-TOADS algorithm implemented to QMENTA platform. All measurements were acquired as part of an IRB-approved clinical protocol, and support was provided by the National Institute of Allergy and Infectious Disease Division of Intramural Research.
Results
All RF-based COMRISv2 models validated (p<0.0001 for all) in the independent cohort. The predictions were stronger for models of physical disability, from which the model based on granular CombiWISE scale achieved the highest correlation (Spearman Rho = 0.855; Linh’s concordance coefficient that reflects 1:1 concordance between predicted and measured outcome; CCC = 0.824). COMRISv2 model of cognitive disability predicted measured symbol digit modalities tests (SDMT) with Rho = 0.493. Unexpectedly, formal comparison of the models consisting only from qMRI or semi-qMRI features demonstrated stronger predictive power of the latter.
Conclusions
COMRISv2 predicts clinical outcomes with strong accuracy but models of physical disability favor qMRI biomarkers reflecting disease burden in the infratentorial compartment, which is currently not measurable as qMRI biomarkers with sufficient accuracy. Addition of qMRI biomarkers of telencephalon damage only strengthened the performance of cognitive disability COMRISv2 model.