Search

Predicting the transition from relapsing–remitting (RR) to secondary-progressive (SP) multiple sclerosis (MS) from early in the disease course is challenging. To construct prediction models for SPMS using sociodemographic and self-reported clinical measures that would be available at/near MS onset, with specific considerations for MS genetic risk factors. We conducted a retrospective cross-sectional study based on 1295 white, non-Hispanic individuals. Cox proportional hazard prediction models were generated for three censored SPMS outcomes (ever transitioning, transitioning within 10 years, and transitioning within 20 years) using sociodemographic, comorbid health information, symptomatology, and other measures of early disease activity. HLADRB1*15:01 and HLA-A*02:01, as well as a genetic risk score, were iteratively considered in each model. We also explored the relationships for all 200 MS risk variants located outside the major histocompatibility complex. Nomograms were generated for the final prediction models. An older age of MS onset and being male predicted a short latency to SPMS, while a longer interval between the first two relapses predicted a much longer latency. Comorbid conditions and onset symptomatology variably predicted the risk for transitioning to SPMS for each censored outcome. The most notable observation was that HLA-A*02:01, which confers decreased risk for MS, also contributed to decreased hazards for SPMS. These results have the potential to advance prognostication for a person with MS using information available at or near onset, potentially improving care and quality of life for those who live with MS.

(1) Background: Complex genetic relationships, including gene-gene (G × G, epistasis), gene(n), and gene-environment (G × E) interactions, explain a substantial portion of the heritability in multiple sclerosis (MS). Machine learning and data mining methods are promising approaches for uncovering higher order genetic relationships, but their use in MS have been limited. (2) Methods: Association rule mining (ARM), a combinatorial rule-based machine learning algorithm, was applied to genetic data for non-Latinx MS cases (n = 207) and controls (n = 179). The objective was to identify patterns (rules) amongst the known MS risk variants, including HLA-DRB1*15:01 presence, HLA-A*02:01 absence, and 194 of the 200 common autosomal variants. Probabilistic measures (confidence and support) were used to mine rules. (3) Results: 114 rules met minimum requirements of 80% confidence and 5% support. The top ranking rule by confidence consisted of HLA-DRB1*15:01, SLC30A7-rs56678847 and AC093277.1-rs6880809, carriers of these variants had a significantly greater risk for MS (odds ratio = 20.2, 95% CI: 8.5, 37.5, p = 4 × 10−9). Several variants were shared across rules, the most common was INTS8-rs78727559, which was in 32.5% of rules. (4) Conclusions: In summary, we demonstrate evidence that specific combinations of MS risk variants disproportionately confer elevated risk by applying a robust analytical framework to a modestly sized study population.