Premium
Automated Categorization of Parkinsonian Syndromes Using Magnetic Resonance Imaging in a Clinical Setting
Author(s) -
Chougar Lydia,
Faouzi Johann,
Pyatigorskaya Nadya,
YahiaCherif Lydia,
Gaurav Rahul,
Biondetti Emma,
Villotte Marie,
Valabrègue Romain,
Corvol JeanChristophe,
Brice Alexis,
Mariani LouiseLaure,
Cormier Florence,
Vidailhet Marie,
Dupont Gwendoline,
Piot Ines,
Grabli David,
Payan Christine,
Colliot Olivier,
Degos Bertrand,
Lehéricy Stéphane
Publication year - 2021
Publication title -
movement disorders
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 3.352
H-Index - 198
eISSN - 1531-8257
pISSN - 0885-3185
DOI - 10.1002/mds.28348
Subject(s) - progressive supranuclear palsy , magnetic resonance imaging , receiver operating characteristic , parkinsonism , cohort , medicine , parkinson's disease , atrophy , psychology , artificial intelligence , nuclear medicine , pathology , radiology , disease , computer science
Background Machine learning algorithms using magnetic resonance imaging (MRI) data can accurately discriminate parkinsonian syndromes. Validation in patients recruited in routine clinical practice is missing. Objective The aim of this study was to assess the accuracy of a machine learning algorithm trained on a research cohort and tested on an independent clinical replication cohort for the categorization of parkinsonian syndromes. Methods Three hundred twenty‐two subjects, including 94 healthy control subjects, 119 patients with Parkinson's disease (PD), 51 patients with progressive supranuclear palsy (PSP) with Richardson's syndrome, 35 with multiple system atrophy (MSA) of the parkinsonian variant (MSA‐P), and 23 with MSA of the cerebellar variant (MSA‐C), were recruited. They were divided into a training cohort (n = 179) scanned in a research environment and a replication cohort (n = 143) examined in clinical practice on different MRI systems. Volumes and diffusion tensor imaging (DTI) metrics in 13 brain regions were used as input for a supervised machine learning algorithm. To harmonize data across scanners and reduce scanner‐dependent effects, we tested two types of normalizations using patient data or healthy control data. Results In the replication cohort, high accuracies were achieved using volumetry in the classification of PD–PSP, PD–MSA‐C, PSP–MSA‐C, and PD‐atypical parkinsonism (balanced accuracies: 0.840–0.983, area under the receiver operating characteristic curves: 0.907–0.995). Performances were lower for the classification of PD–MSA‐P, MSA‐C–MSA‐P (balanced accuracies: 0.765–0.784, area under the receiver operating characteristic curve: 0.839–0.871) and PD–PSP–MSA (balanced accuracies: 0.773). Performance using DTI was improved when normalizing by controls, but remained lower than that using volumetry alone or combined with DTI. Conclusions A machine learning approach based on volumetry enabled accurate classification of subjects with early‐stage parkinsonism, examined on different MRI systems, as part of their clinical assessment. © 2020 International Parkinson and Movement Disorder Society