Spider-MS: an individualized polyhedral prediction of multiple sclerosis prognosis

Multiple sclerosis (MS) is a potentially disabling disease that shows marked variability in its severity and underlying mechanisms. Early prediction of patients at risk of specific unfavourable outcomes may be key to treatment stratification and management. Here, we propose a prognostic framework, called the Spider-MS model, to predict a wide range of relevant outcomes at the individual level, at symptom onset. We included patients from the Barcelona first-attack cohort, i.e. with a first demyelinating attack suggestive of multiple sclerosis, younger than 50 years old at symptom onset and seen in the clinic within 3 months of the first attack, to build the Spider-MS model (original cohort). Patients with a first demyelinating attack from the Royal Melbourne Hospital were used to validate the model externally (validation cohort). All patients were prospectively assessed clinically, with the Expanded Disability Status Scale (EDSS) and relapse tracking, and through brain and, in some cases, spinal cord MRI. Spider-MS was built as a set of eight accelerated failure models with Weibull distribution, one for each outcome, including McDonald 2017 diagnosis, second attack, yearly rate of new T2 lesions > 2, relapse-associated worsening (RAW) at the first attack and at subsequent attacks, confirmed and sustained disability worsening, progression independent of relapse activity (PIRA) and EDSS 3.0. Model predictors included age, sex, first attack topography, brain and spinal cord lesions, CSF oligoclonal bands and percentage of time on high-/moderate-efficacy treatment before the outcome. We included 1180 patients from Barcelona (mean age 32.37 years, 810 females) and 108 from Melbourne (32.41 years, 78 females). Median follow-up times were 10.80 and 11.10 years, respectively. In the original cohort, 797/1180 (67.5%) fulfilled the McDonald criteria, 121/1180 (10.3%) developed RAW at subsequent relapses and 290/1180 (24.6%) developed PIRA over the follow-up period. The prediction models reached moderate-high accuracy levels (Harrell's C values: 0.653-0.823) and showed that older age, cord involvement at first attack, greater number of brain and cord lesions, presence of CSF oligoclonal bands and lower percentage of time on treatment predicted a greater risk of unfavourable outcomes, with varying effects depending on the outcome. When the Spider-MS model was applied to the external cohort, for all outcomes except for RAW at first or subsequent attacks, the prediction models reached moderate-high accuracy values (Harrell's C = 0.623-0.766). In general, patients with the highest predicted risks experienced acute inflammatory activity or disability earlier than lower-risk patients. In conclusion, our Spider-MS model can be considered a promising individual predictive tool with the potential to inform clinical practice.