Investigating resting state connectivity alterations in temporal lobe epilepsy with machine learning
Stand-By Time
Wednesday, June 28, 2017: 12:45 PM - 2:45 PMSubmission No:
3128
Submission Type:
Abstract Submission
On Display:
Wednesday, June 28 & Thursday, June 29
Authors:
Gyujoon Hwang1, Jed Mathis2, VEENA NAIR1, Ferdaus Kawsar2, Rosaleena Mohanty1, Gengyan Zhao1, Megan Rozman2, Taylor McMillan1, Dace Almane1, Andrew Nencka2, Mohsen Mazrooyisebdani1, Elizabeth Felton1, Aaron Struck1, Rasmus Birn1, Rama Maganti1, Lisa Conant2, Colin Humphries2, Bruce Hermann1, Manoj Raghavan2, Edgar DeYoe2, Jeffrey Binder2, Beth Meyerand1, Vivek Prabhakaran1
Institutions:
1University of Wisconsin-Madison, Madison, WI, 2Medical College of Wisconsin, Milwaukee, WI
First Author:
Introduction:
Resting-state functional MRI (fMRI) has been a widely studied topic in the recent years, due to its simplicity and reliability. The NIH-sponsored Epilepsy Connectome Project (ECP) aims to characterize connectivity changes in people with temporal lobe epilepsy (TLE). MRI protocols follow those used in the Human Connectome Project (HCP), and include 40 minutes of resting-state fMRI acquired at 3T using 8-band multiband imaging. In this ongoing study, we applied binary linear support vector machine (SVM) linear classifier analysis to resting state connectivity data of TLE patients and healthy controls, to identify connections (features) that distinguish these groups. Two prior studies reported 83.9% and 83.3% accuracy in classifying epilepsy patients vs. controls using lower resolution images.
Methods:
30 TLE patients (age = 41.0 years, 13 males), and 32 healthy controls (age = 30.1 years, 14 males) were tested. All images were acquired with a 3T GE scanner using simultaneous multi-slice ("multi-band") imaging. Four 5-minute resting state scans (TR = 802ms, 2.0mm isotropic) were combined to generate connectivity matrices. Participants fixated on a cross during the scans. For spatial alignment, a T1-weighted anatomical image was acquired using a magnetization prepared gradient echo sequence (0.8mm isotropic).
Preprocessing included image registration, bandpass filtering (0.01-0.1 Hz), and removal of white matter, cerebrospinal fluid, global signal, and motion regressors. FreeSurfer-based preprocessing scripts developed by HCP were used to project the data to surface space and extract time series data from 360 regions defined by HCP's Glasser parcellation scheme. Pearson's correlations between every pair of parcels were computed and normalized with Fisher transformation to generate connectivity matrices. Correlations with less than 30% of the maximum strength were considered as having no correlation.
Bottom triangular regions of these matrices were reshaped into a row matrix with 64,620 features and used as inputs to a linear SVM classifier. Classification was performed with the Spider Machine Learning Toolbox 3 as well as custom MATLAB scripts, using a linear kernel with misclassification margin parameter, C=1. Each feature was first evaluated for its individual performance in classifying, and the top performing features were selected. The ideal number of these 'top features' was searched by sweeping through a range of numbers. For cross validation, leave-one-out-cross-validation (LOOCV) was used.
Preprocessing included image registration, bandpass filtering (0.01-0.1 Hz), and removal of white matter, cerebrospinal fluid, global signal, and motion regressors. FreeSurfer-based preprocessing scripts developed by HCP were used to project the data to surface space and extract time series data from 360 regions defined by HCP's Glasser parcellation scheme. Pearson's correlations between every pair of parcels were computed and normalized with Fisher transformation to generate connectivity matrices. Correlations with less than 30% of the maximum strength were considered as having no correlation.
Bottom triangular regions of these matrices were reshaped into a row matrix with 64,620 features and used as inputs to a linear SVM classifier. Classification was performed with the Spider Machine Learning Toolbox 3 as well as custom MATLAB scripts, using a linear kernel with misclassification margin parameter, C=1. Each feature was first evaluated for its individual performance in classifying, and the top performing features were selected. The ideal number of these 'top features' was searched by sweeping through a range of numbers. For cross validation, leave-one-out-cross-validation (LOOCV) was used.
Results:
The SVM classifier's best LOOCV accuracy was 85.5% (sensitivity = 80.0%, specificity = 90.6%) when 132 top features were used in the training. SVM assigns weight to each feature according to the contribution of the feature in driving the classifier. Since the combination of the weighted features defines the classifier, a high absolute weight does not necessarily indicate that the individual connection has a significant group difference. Features that had the highest absolute weights and their HCP-defined labels and regions are summarized in Table 1. The negative weights indicate that the connections were more positive for the patients, and vice versa. These results are also visualized in Figure 1.
Conclusions:
In this preliminary analysis, SVM showed good accuracy in differentiating TLE patients from healthy subjects using resting state connectivity matrices based on the Glasser parcellation. The performance will likely improve with larger sample sizes and separation of TLE patients into left and right hemisphere seizure focus subgroups. Future work will explore different machine learning classifier methods, analyze data in misclassified subjects, and explore multi-class classification using patient subsets.
Disorders of the Nervous System:
Epilepsy 1
Imaging Methods:
BOLD fMRI
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
fMRI Connectivity and Network Modeling
Task-Independent and Resting-State Analysis
Keywords:
Epilepsy
FUNCTIONAL MRI
Machine Learning
Other - Functional Connectivity
1|2Indicates the priority used for review
Would you accept an oral presentation if your abstract is selected for an oral session?
I would be willing to discuss my abstract with members of the press should my abstract be marked newsworthy:
Please indicate below if your study was a "resting state" or "task-activation” study.
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute the presentation in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels or other electronic media and on the OHBM website.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Internal Review Board (IRB) or Animal Use and Care Committee (AUCC) Approval. Please indicate approval below. Please note: Failure to have IRB or AUCC approval, if applicable will lead to automatic rejection of abstract.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references in author date format
Glasser, M.F., at el. (2016), ‘The Human Connectome Project’s neuroimaging approach’, Nature Neuroscience, vol. 19, pp. 1175-1187.
Pereira, F., Mitchell, T., Botvinick, M. (2009), ‘Machine learning classifiers and fMRI: a tutorial overview’, NeuroImage, vol. 45, pp. S199-S209
Zhang, J., at el. (2012) ‘Pattern classification of large-scale functional brain networks: identification of informative neuroimaging markers for epilepsy’, PLoS One, vol. 7, no. 5, e36733.
Zhengyi, Y., at el. (2015) ‘Lateralization of temporal lobe epilepsy based on resting-state functional magnetic resonance imaging and machine learning’, Frontiers in Neurology, vol. 6, pp. 184
Feinberg, D.A., Setsompop, K. (2013), ‘Ultra-fast MRI of the human brain with simultaneous multi-slice imaging’, Journal of Magnetic Resonance, vol. 229, pp. 90-100.
Weston, J., at el. (2005), ‘The Spider Machine Learning Toolbox’,
Boser, B.E., Guyon, I.M., Vapnik, V.N. (1992), ‘A training algorithm for optimal margin classifiers’, 5th Annual ACM Workshop on COLT, pp. 144-152.