Contributions
Abstract: P553
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Introduction: Graph theory has emerged as a powerful tool to model complex systems by characterising relationships between distinct brain regions. Resting state fMRI (rs-fMRI) has also provided important insights into functional reorganization in subjects with Multiple Sclerosis (MS) at different stages of the disease. In this study we coupled graph theoretical analysis, based on rs-fMRI images, with a machine learning approach to assess the ability of graph metrics to discriminate relapsing remitting MS (RRMS) subjects with short disease duration (MSshort) from those with similar (mild) disability but longer disease duration (MSlong).
Methods: 36 RRMS patients with recent disease onset (≤5yrs, mean EDSS=1.4±0.9; MSshort) and 26 RRMS patients with later disease (>5yrs duration; mean EDSS=2.2±1.4; MSlong) underwent 3T MRI examination including rs-fMRI and 3D T1-weighted imaging. 29 healthy controls (HC) were also scanned.
For each subject, rs-fMRI images were preprocessed using FSL and then parcellated into 116 distinct regions using the automatic anatomical labelling (AAL) atlas. For each region, the mean rs-fMRI signal was extracted and used to calculate the graph edges defined as the Pearson's correlations between all pairs of AAL regions (graph nodes). The resulting cross-correlation matrix was thresholded and then processed to calculate graph metrics. Extracted graph metrics were used as input features to run a Support Vector Machine (SVM) classifier.
Results: SVM applied to graph metrics identified small sets of features (brain graph measures) to discriminate MSshort, MSlong and HC. SVM achieved the best classification performance when considering the discrimination between MSshort and MSlong, reaching a classification accuracy (ACC) equal to 93.3% using 10 features from 8 distinct nodes (Cbl-7b, Cbl-9, Vermis-45, Vermis-8, Cuneus, Inf. Par. Gyr, Mid. Cing. Gyr, Sup.Temp.Gyr) all located in the right hemisphere of the brain. SVM reached ACC=88.9% to classify MSshort from HC (using 8 features) and ACC=90.9% to classify MSlong from HC (using 11 features).
Conclusions: Results demonstrate the potential of machine learning (SVM) combined with graph methods to automatically classify RRMS patients' clinical profiles. The proposed method was particularly efficient in discriminating MSshort from MSlong. This warrants future longitudinal studies of early RRMS patients to assess the ability of the classifier to provide insights into their possible evolution.
Disclosure:
Gloria Castellazzi: is supported by ECTRIMS (ECTRIMS postdoctoral research fellowship exchange programme)
Laëtitia Debernard: nothing to disclose
Tracy R. Melzer: nothing to disclose
John C. Dalrymple-Alford: nothing to disclose
David H. Miller: has received honoraria through payments to UCL Institute of Neurology, for Advisory Committee and/or Consultancy advice in multiple sclerosis studies from Novartis and Mitsubishi Pharma Europe and compensation through payments to UCL Institute of Neurology for performing central MRI analysis of a multiple sclerosis trial from Novartis.
Egidio D'Angelo: nothing to disclose
Deborah F. Mason: has received honoraria and travel grants from Biogen Idec, Novartis and TEVA.
Claudia A.M. Gandini Wheeler-Kingshott: has received research grants (PI and co-applicant) from Spinal Research, Craig H. Neilsen Foundation, EPSRC, Wings for Life, UK MS Society, Horizon2020, NIHR/MRC.
Abstract: P553
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Introduction: Graph theory has emerged as a powerful tool to model complex systems by characterising relationships between distinct brain regions. Resting state fMRI (rs-fMRI) has also provided important insights into functional reorganization in subjects with Multiple Sclerosis (MS) at different stages of the disease. In this study we coupled graph theoretical analysis, based on rs-fMRI images, with a machine learning approach to assess the ability of graph metrics to discriminate relapsing remitting MS (RRMS) subjects with short disease duration (MSshort) from those with similar (mild) disability but longer disease duration (MSlong).
Methods: 36 RRMS patients with recent disease onset (≤5yrs, mean EDSS=1.4±0.9; MSshort) and 26 RRMS patients with later disease (>5yrs duration; mean EDSS=2.2±1.4; MSlong) underwent 3T MRI examination including rs-fMRI and 3D T1-weighted imaging. 29 healthy controls (HC) were also scanned.
For each subject, rs-fMRI images were preprocessed using FSL and then parcellated into 116 distinct regions using the automatic anatomical labelling (AAL) atlas. For each region, the mean rs-fMRI signal was extracted and used to calculate the graph edges defined as the Pearson's correlations between all pairs of AAL regions (graph nodes). The resulting cross-correlation matrix was thresholded and then processed to calculate graph metrics. Extracted graph metrics were used as input features to run a Support Vector Machine (SVM) classifier.
Results: SVM applied to graph metrics identified small sets of features (brain graph measures) to discriminate MSshort, MSlong and HC. SVM achieved the best classification performance when considering the discrimination between MSshort and MSlong, reaching a classification accuracy (ACC) equal to 93.3% using 10 features from 8 distinct nodes (Cbl-7b, Cbl-9, Vermis-45, Vermis-8, Cuneus, Inf. Par. Gyr, Mid. Cing. Gyr, Sup.Temp.Gyr) all located in the right hemisphere of the brain. SVM reached ACC=88.9% to classify MSshort from HC (using 8 features) and ACC=90.9% to classify MSlong from HC (using 11 features).
Conclusions: Results demonstrate the potential of machine learning (SVM) combined with graph methods to automatically classify RRMS patients' clinical profiles. The proposed method was particularly efficient in discriminating MSshort from MSlong. This warrants future longitudinal studies of early RRMS patients to assess the ability of the classifier to provide insights into their possible evolution.
Disclosure:
Gloria Castellazzi: is supported by ECTRIMS (ECTRIMS postdoctoral research fellowship exchange programme)
Laëtitia Debernard: nothing to disclose
Tracy R. Melzer: nothing to disclose
John C. Dalrymple-Alford: nothing to disclose
David H. Miller: has received honoraria through payments to UCL Institute of Neurology, for Advisory Committee and/or Consultancy advice in multiple sclerosis studies from Novartis and Mitsubishi Pharma Europe and compensation through payments to UCL Institute of Neurology for performing central MRI analysis of a multiple sclerosis trial from Novartis.
Egidio D'Angelo: nothing to disclose
Deborah F. Mason: has received honoraria and travel grants from Biogen Idec, Novartis and TEVA.
Claudia A.M. Gandini Wheeler-Kingshott: has received research grants (PI and co-applicant) from Spinal Research, Craig H. Neilsen Foundation, EPSRC, Wings for Life, UK MS Society, Horizon2020, NIHR/MRC.