Stratification of multiple sclerosis patients using unsupervised machine learning: a single-visit MRI-driven approach

• DOI: 10.1007/s00330-022-08610-z
• Publication Date: 2022-03-14T04:12:31Z
• Authors: Giuseppe Pontillo; Simone Penna; Sirio Cocozza; Mario Quarantelli; Michela Gravina; Roberta Lanzillo; Stefano Marrone; Teresa Costabile; Matilde Inglese; Vincenzo Brescia ...; Daniele Riccio; Andrea Elefante; Maria Petracca; Carlo Sansone; Arturo Brunetti
• Keywords: Neuroradiology; Fluid-attenuated inversion recovery

To stratify patients with multiple sclerosis (pwMS) based on brain MRI-derived volumetric features using unsupervised machine learning.The 3-T MRIs of relapsing-remitting pwMS including 3D-T1w and FLAIR-T2w sequences were retrospectively collected, along Expanded Disability Status Scale (EDSS) scores long-term (10 ± 2 years) clinical outcomes (EDSS, cognition, progressive course). From the MRIs, volumes demyelinating lesions 116 atlas-defined gray matter regions automatically segmented expressed as z-scores referenced to external populations. Following feature selection, baseline biomarkers entered Subtype Stage Inference (SuStaIn) algorithm, which estimates subgroups characterized by distinct patterns biomarker evolution stages within subgroups. The trained model was then applied longitudinal MRIs. Stability subtypes stage change over time assessed via Krippendorf's α multilevel linear regression models, respectively. prognostic relevance SuStaIn classification ordinal/logistic analyses.We selected 425 (35.9 9.9 years; F/M: 301/124), corresponding 1129 MRI scans, healthy controls (N = 148; 35.9 13.0 77/71) 80; 40.4 11.9 56/24) reference Based 11 surviving two identified, designated "deep (DGM)-first" subtype 238) "cortex-first" 187) according atrophy pattern. Subtypes consistent (α 0.806), significant annual increase (b 0.20; p < 0.001). EDSS associated DGM-first (p ≤ 0.02). Baseline predicted disability, transition course, cognitive impairment 0.03), latter also 0.005).Unsupervised learning modelling provides a biologically reliable prognostically meaningful stratification pwMS.• can provide single-visit patients. • so-obtained tends be captures disease-related damage progression, supporting biological reliability model. predicts secondary course.