QL10
A New Machine Learning Approach to Predict Disease Course in Multiple Sclerosis

Thursday, May 25, 2017
B2 (New Orleans Convention Center)
Giampaolo Brichetto, MD, PhD , Scientific Research Area, Italian MS Society, Genoa, Italy, Genova, Italy
Samuele Fiorini, PhD , DIBRIS, University of Genoa, Genova, Italy
Michela Ponzio, PhD , Scientific Research Area, Italian MS Society, Genoa, Italy, Genova, Italy
Annalisa Barla, PhD , DIBRIS, University of Genoa, Genova, Italy
Alessandro Verri, PhD , DIBRIS, University of Genoa, Genova, Italy
Andrea Tacchino, PhD , Scientific Research Area, Italian MS Society, Genoa, Italy, Genova, Italy
Andrea Tacchino, PhD , Scientific Research Area, Italian MS Society, Genoa, Italy, Genova, Italy



Background: Measuring Multiple Sclerosis (MS) is an hard challenge mainly for the functional problems variety that may affect the individual patients. An universally accepted measurement instrument precise, reliable, easy to administer, able to capture the key neurological domains affected by MS, sensitive at various disability levels and accurately reflecting neurological and neuropsychological disability is still lacking. Therefore, the development of improved clinical outcome measures is a significant unmet need. Achieving an accurate clinical course description in MS is crucial for communication, prognosis, treatment decision-making, design and recruitment of clinical trials.

Objectives: This work aims at using Patient Reported Outcomes (PROs), Clinical Scales (CS) and anthropometric measures, collected within the Italian MS Foundation (FISM) initiative PROMPRO-MS dataset, for the detection of MS courses by means of machine learning (ML) techniques.

Methods: The set of PRO and CS was: EDSS; FIM; Edinburgh Handedness Inventory; Abilhand; MoCA; PASAT 3; SDMT; HADS; LSI; OAB-q; MFIS for a total of 126 features. 852 people with MS (PwMS), followed by Italian MS Society Rehabilitation Centre (Genoa, Padova, Vicenza), have been enrolled in the study. Data are acquired longitudinally every 3-4 months from January 2014. The proposed ML analysis is conducted on a dataset presenting the first time point of the 852 subjects and the 126 features.

Results: The applied ML techniques showed that PwMS diagnosed as relapsing-remitting (RR) could be isolated from other clinical courses (ALL). In particular, nine “top” questions were selected by the ‘Features Selection’ supervised (FS) algorithm: one question from LSI, three items from FIM; one from MFIS, one from HADS, one from MoCA, one from SDMT and one from Abilhand.

Conclusions: To the very best of our knowledge this is the first study which  predicted  MS  course  taking  only  into account a small subset of anthropometric and questionnaires variables, which could be proposed as a novel questionnaire, tailored for RR detection.