One of the most clinically frustrating realities in multiple sclerosis care is that disability can silently accumulate even when relapses are fully suppressed. Understanding what drives this so-called progression independent of relapse activity — PIRA — has become a top priority in MS research, because current therapies that excel at controlling relapses may leave patients unprotected against a parallel, smoldering disease process.
This population-based Swedish cohort study examined 2,837 relapsing-remitting MS patients who initiated rituximab, a potent B-cell depleting therapy, between 2010 and 2019. All participants were relapse-free throughout a six-year follow-up, making any observed disability worsening attributable solely to PIRA. Approximately 20% — 563 patients — experienced confirmed disability worsening on the Expanded Disability Status Scale. Researchers tested both hypothesis-driven analysis of pre-specified comorbidities and a data-driven sweep of all ICD-10 diagnostic codes from secondary care, applying machine learning models including elastic net, random forest, XGBoost, and neural networks. Among the pre-specified comorbidities — depression and anxiety (36%), hypertension (15%), and headache (8%) being the most prevalent — none emerged as statistically significant predictors of PIRA. Across all diagnostic codes, only neuromuscular bladder dysfunction reached significance after correction for multiple comparisons.
The null finding here is scientifically meaningful, not merely inconclusive. Comorbidities are increasingly recognized as contributors to MS disability, and several — particularly vascular and metabolic conditions — have been hypothesized to accelerate neurodegeneration through overlapping pathways. This study's failure to confirm that relationship for PIRA specifically suggests either that comorbidity effects on progression operate through relapse-mediated mechanisms rather than smoldering neurodegeneration, or that secondary-care ICD codes are too coarse an instrument to capture the relevant exposures. The machine learning ensemble approach adds methodological rigor, though AUC-based predictive performance was presumably modest given the null results. Limitations include the observational design, reliance on administrative diagnostic codes, and a predominantly Scandinavian cohort with potential ethnic homogeneity. Identifying reliable PIRA predictors remains an open and urgent clinical question.