Applying machine learning to multimodal neuroimaging data to classify multiple sclerosis patients with and without processing speed impairment
ECTRIMS Online Library. Buyukturkoglu K. 09/13/19; 278480; P1277
Korhan Buyukturkoglu
Korhan Buyukturkoglu
Contributions
Abstract

Abstract: P1277

Type: Poster Sessions

Abstract Category: Pathology and pathogenesis of MS - MRI and PET

K. Buyukturkoglu1, M. Nunez2, S. Lee2, J.F. Sumowski3, V.M. Leavitt1

1Department of Neurology, Columbia University Irving Medical Center, 2Department of Biostatistics and Psychiatry, Columbia University Mailman School of Public Health, 3Department of Neurology, Icahn School of Medicine at Mount Sinai, New York, NY, United States

Introduction: Processing speed (PS) is commonly impaired in multiple sclerosis (MS). Several neuroimaging correlates have been identified; however, well-defined neuroimaging biomarkers and predictors of this symptom are still lacking.
Objectives: To classify PS-impaired and preserved MS patients into different groups and extract a brain pattern related to PS impairment applying machine learning (ML) tools to neuroimaging data.
Methods: We segregated 179 out of 183 MS patients with Magnetic Resonance Imaging (MRI) data (RADIEMS cohort: 65.9% female, mean age 34.4 ± 7.6 years, disease duration 2.1 ± 1.4 years) into PS impaired (sample-based z-scores among MS patients <  -1.0, N = 33) and preserved groups based on Symbol Digit Modalities Test performance. Two supervised pattern classifiers (Elastic net and Random forest methods) were trained to classify patients into two groups using multimodal MRI data: global brain volume measures, thickness measures of 68 cortical regions, volume of 14 deep gray matter structures, hippocampal subfield volumes, T2 lesion load and location, and whole brain fractional anisotropy and mean diffusivity values. To compute classification performance, MRI data were randomly partitioned into training (80%) and test (20%) sets, followed by 5-fold cross-validation to identify optimal models. The process was repeated 1000 times. For each model, important variables and area under the curve (AUC) were calculated. The variable importance scores were computed as average of the standardized absolute value of the coefficients for Elastic net and average of the mean decrease in accuracy for Random forest across 1000 models.
Results: Average AUC for Random forest method was 0.71 (± 0.096). Elastic net performed a better classification with an average AUC of 0.78 (± 0.096). Both ML methods found T2 lesion load in the cerebellum to be the most important feature for classifying patients into PS impaired and preserved groups.
Conclusions: While the cerebellum has typically been associated with sensorimotor and vestibular functions, its pivotal role in cognition has been underappreciated. Our results revealed that cerebellar lesions played the most important role in classifying patients with PS impairment at the initial stages of MS disease. Future studies combining within and between cerebellum connectivity and detailed cerebellar atrophy measures with other MRI features warranted to be able to extract more precise brain patterns of PS impairment.
Disclosure: Nothing to disclose

Abstract: P1277

Type: Poster Sessions

Abstract Category: Pathology and pathogenesis of MS - MRI and PET

K. Buyukturkoglu1, M. Nunez2, S. Lee2, J.F. Sumowski3, V.M. Leavitt1

1Department of Neurology, Columbia University Irving Medical Center, 2Department of Biostatistics and Psychiatry, Columbia University Mailman School of Public Health, 3Department of Neurology, Icahn School of Medicine at Mount Sinai, New York, NY, United States

Introduction: Processing speed (PS) is commonly impaired in multiple sclerosis (MS). Several neuroimaging correlates have been identified; however, well-defined neuroimaging biomarkers and predictors of this symptom are still lacking.
Objectives: To classify PS-impaired and preserved MS patients into different groups and extract a brain pattern related to PS impairment applying machine learning (ML) tools to neuroimaging data.
Methods: We segregated 179 out of 183 MS patients with Magnetic Resonance Imaging (MRI) data (RADIEMS cohort: 65.9% female, mean age 34.4 ± 7.6 years, disease duration 2.1 ± 1.4 years) into PS impaired (sample-based z-scores among MS patients <  -1.0, N = 33) and preserved groups based on Symbol Digit Modalities Test performance. Two supervised pattern classifiers (Elastic net and Random forest methods) were trained to classify patients into two groups using multimodal MRI data: global brain volume measures, thickness measures of 68 cortical regions, volume of 14 deep gray matter structures, hippocampal subfield volumes, T2 lesion load and location, and whole brain fractional anisotropy and mean diffusivity values. To compute classification performance, MRI data were randomly partitioned into training (80%) and test (20%) sets, followed by 5-fold cross-validation to identify optimal models. The process was repeated 1000 times. For each model, important variables and area under the curve (AUC) were calculated. The variable importance scores were computed as average of the standardized absolute value of the coefficients for Elastic net and average of the mean decrease in accuracy for Random forest across 1000 models.
Results: Average AUC for Random forest method was 0.71 (± 0.096). Elastic net performed a better classification with an average AUC of 0.78 (± 0.096). Both ML methods found T2 lesion load in the cerebellum to be the most important feature for classifying patients into PS impaired and preserved groups.
Conclusions: While the cerebellum has typically been associated with sensorimotor and vestibular functions, its pivotal role in cognition has been underappreciated. Our results revealed that cerebellar lesions played the most important role in classifying patients with PS impairment at the initial stages of MS disease. Future studies combining within and between cerebellum connectivity and detailed cerebellar atrophy measures with other MRI features warranted to be able to extract more precise brain patterns of PS impairment.
Disclosure: Nothing to disclose

By clicking “Accept Terms & all Cookies” or by continuing to browse, you agree to the storing of third-party cookies on your device to enhance your user experience and agree to the user terms and conditions of this learning management system (LMS).

Cookie Settings
Accept Terms & all Cookies