You are viewing the SNL 2018 Archive Website. For the latest information, see the Current Website.

Poster E41, Saturday, August 18, 3:00 – 4:45 pm, Room 2000AB

Classification of fMRI Data in Aphasia Based on Task, Time Point, and Subject

E. Susan Duncan1, Steven L. Small2;1Louisiana State University, 2University of California, Irvine

Introduction: Functional magnetic resonance imaging (fMRI) investigations of people with aphasia typically try to identify group characterizations of behavioral deficits or therapeutic improvement. Previous work in healthy individuals has demonstrated that fMRI can reliably distinguish among them across different tasks and time points. In the present study, we hypothesized that task-(in)dependent fMRI could reliably distinguish among individuals with aphasia at various time points during a treatment study. We seek to identify whether we can train a classifier to reliably separate fMRI samples based on subject, task, or time point (pre- vs. post-therapy). This work is ultimately intended to better understand patterns of functional connectivity associated with recovery. Methods: Nineteen subjects with post-stroke aphasia were scanned under two conditions (rest, speech observation) on ≥ 6 occasions over 18 weeks as part of a larger behavioral therapy study. Up to 3 baseline and 3 post-therapy scans were collected for each subject. We used FreeSurfer and the Connectome Mapping Toolbox to parcellate anatomical data into 82 cortical/subcortical regions, facilitated by Virtual Brain Transplant. Functional data were pre-processed using AFNI and FSL and nonlinearly aligned to a common template (ANTs). We applied a General Linear Model to remove nuisance regressors (motion; signal from white matter, ventricles, lesion) as well as stimulus timing for task-dependent fMRI. These residual time series were bandpass filtered (0.01-0.1 Hz). Scans were excluded if they had ≤ 3 minutes of uncensored data (<0.3 mm or degrees of motion), resulting in 132 data sets from 14 subjects. We extracted average time series from each of the 82 brain regions for each of the functional runs. Each of these 132 time series was mean-corrected and used to calculate a correlation matrix using the statistical language R. Ninety percent (n=118) of these correlation matrices were randomly selected to train a support vector machine (linear kernel; scikit-learn) to classify the remaining 14 matrices according to either task (n=2; rest, observation), time point (n=2; pre vs. post-therapy), or subject (n=14). Due to their probabilistic nature these analyses were repeated 100 times using different sets of training/testing data. Results: Identification by task was inconsistently above chance (M= 59.8%; SD= 4.1; range= 48.6-74.3), as was identification by time point (M= 61.4%; SD= 3.7; range= 52.1-70.0). Identification by subject was nearly perfect (M= 96.7%; SD= 1.8; range= 91.4-100). Discussion: The intrinsic organization of the brain is strongly subject to individual variation. However, it appears to be less distinctive in response to task, suggesting that observations from resting state fMRI are relevant for understanding changes in brain organization that underlie behavioral improvement on a task. This is beneficial given the potential limitations of individuals with aphasia to participate in task-based fMRI. Failure to divide the group based on time point (pre- vs. post-therapy) suggests that patterns of plasticity are not necessarily common across participants and should be viewed in relation to behavioral change. These findings highlight the importance of characterizing treatment- or recovery-induced changes in functional connectivity on an individual basis and suggest a foundation for personalized medicine.

Topic Area: Language Therapy