Publication
Predicting personalised risk of disability worsening in multiple sclerosis with machine learning
Publisher:
Cold Spring Harbor Laboratory
Date:
03-02-2022
DOI:
10.1101/2022.02.03.22270364
Abstract: Multiple sclerosis is a heterogeneous disease with an unpredictable course. We applied machine learning to generate in idualised risk scores of disability worsening and stratify patients into subgroups with different prognosis. Clinical data and MRI scans from published randomised clinical trials in patients with relapsing-remitting and progressive MS were ided into training (n=5,483) and external validation data sets (n=2,668). We processed brain MRI scans to obtain 18 measures for lobar grey matter, deep grey matter and lesion volumes, and T1-/T2-weighted ratio of the normal-appearing white matter regions. We developed a machine learning model, called subpopulation risk stratification (SunRiSe), that combines multi-parametric clinical and MRI data to estimate in idualised risk scores and stratify patients into subgroups on the basis of this risk in particular, we entered MRI measures, the Expanded Disability Status Scale, age and gender to generate risk scores of disability worsening (i.e., the time to confirmed disability worsening). Based on SunRiSe risk scores, high-, medium-, and low-risk subpopulations were defined at study entry. We assessed whether selecting patients at high risk of disability worsening reduces s le size compared to when all risk groups were s led together. In both the training and external validation data sets, SunRiSe-stratified patients in three groups associated with different levels of risk of disability worsening. In the external validation data set, patients at high risk were mainly progressive MS and had more disability events compared to those at medium-risk (hazard ratio [HR]=1.34, p .0001) and low-risk (HR=1.51, p .0001). At study entry, male gender, older age, higher lesion load, higher disability, lower lobar cortical grey matter, lower normal-appearing white matter T1/T2 ratio and lower deep grey matter volumes, were the most important variables in defining the SunRiSe risk score. The inclusion of patients predicted to be at high risk, reduced (i) duration of an event-driven trial by an average of 4.5 months (±2.1 months) (ii) the number of participants in a randomised trial by approximately 200, with 80% statistical power to detect a 30% treatment effect. Machine learning provides a personalised risk score that can identify patients who have the greatest risk of disability worsening and therefore should be treated with the most effective medications and monitored more closely. Risk stratification allows the enrichment of clinical trials with patients more likely to worsen, and thereby reduces trial duration and s le size.