Use of a Machine Learning Model to Predict Disability Progression in Patients with Relapsing Multiple Sclerosis Treated with Ozanimod
Ludwig Kappos1, Douglas Arnold2, Chahin Pachai3, Nathanial Eddy3, Jason Osik3, Chun-Yen Cheng3, Sarah Harris3, Erik Deboer3, Leorah Freeman4, Cristina Granziera5, Massimo Filippi6
1Research Center for Clinical Neuroimmunology and Neuroscience Basel (RC2NB), Departments of Head, Spine and Neuromedicine, Clinical Research, Biomedicine, and Biomedical Engineering, University Hospital and University of Basel, Basel, Switzerland, 2NeuroRx Research and Montréal Neurological Institute, McGill University, Montréal, Quebec, Canada, 3Bristol Myers Squibb, Princeton, New Jersey, 4Dell Medical School, The University of Texas at Austin, Austin, Texas, 5Translational Imaging in Neurology (ThINk) Basel, Department of Biomedical Engineering, Faculty of Medicine, University Hospital Basel and University of Basel, Basel, Switzerland; Department of Neurology, University Hospital Basel, Switzerland; Research Center for Clinical Neuroimmunology and Neuroscience Basel (RC2NB), Departments of Head, Spine and Neuromedicine, Clinical Research, Biomedicine, and Biomedical Engineering, University Hospital and University of Basel, Basel, Switzerland, 6Neuroimaging Research Unit, Division of Neuroscience, Neurology Unit Neurorehabilitation Unit, and Neurophysiology Service, IRCCS San Raffaele Scientific Institute, and Vita-Salute San Raffaele University, Milan, Italy
Objective:
To use a machine learning (ML) model to predict long-term clinical disability progression (CDP) in patients with relapsing multiple sclerosis (RMS) treated with ozanimod.
Background:
ML models may predict CDP using baseline demographic, clinical, MRI, and biomarker data.
Design/Methods:
Patients with RMS were treated with ozanimod 0.46 or 0.92mg/d or interferon β-1a 30μg/wk for 24 months in RADIANCE (NCT02047734) or ≥12 months in SUNBEAM (NCT02294058); completers could enroll in an open-label extension trial (DAYBREAK‒NCT02576717) of ozanimod 0.92mg/d for up to ~6 years (database lock: 4/7/2023). Patients were pooled across studies and treatment arms. CDP-6 was defined as a ≥1-point increase from baseline in Expanded Disability Status Scale (EDSS) score, confirmed after 6 months. Predictors tested included demographics, disease duration, relapse history, prior treatment, EDSS, Kurtze Functional Systems Scores (FSS), Timed 25-Foot Walk Test (T25W), 9-Hole Peg Test (9HPT), low-contrast letter acuity (LCLA), MRI lesion counts and volumes, brain volumes, serum glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL) at baseline. An Extreme Gradient Boosting (XGBoost)-based ML model was used. Model parameters were optimized using a stratified cross-validation strategy based on 80% (n=1804) of patients. Individualized risk scores for CDP-6 were derived for the remaining 20% (n=452) and categorized into risk tertiles (low, medium, high); incidence of CDP-6 was assessed for each tertile.
Results:
The trained XGBoost model reliably predicted patient risk (c=0.7). Top 10 impactful predictors included the 9HPT, FSS-Bowel/Bladder, EDSS, NfL, T25W, GFAP, body mass index, FSS-Brainstem, LCLA (z-score), and time from diagnosis. Low-risk patients had negligible incidence of progression (~2% at 24 months, ~6% at 60 months), while in high-risk patients, this incidence was ~17% and ~26% at those respective times.
Conclusions:
The proposed ML model, trained on a large validation set, reliably identified risk groups for long-term CDP-6 in RMS. Additional validation, ideally using external databases, is warranted.