Slide Slam A7
Using Baseline Task fMRI BOLD Activity to Predict Treatment Gains in Persons with Aphasia Undergoing Language Therapy
Serena Song1,2, Lisa Krishnamurthy1,3, Amy Rodriguez1,4, Joo Han1,3, Clara Glassman1,5, Bruce Crosson1,4, Venkatagiri Krishnamurthy1,4,6; 1Center for Visual and Neurocognitive Rehabilitation, Atlanta VA Medical Center, 2Neuroscience and Behavioral Biology, Emory University, 3Physics and Astronomy, Georgia State University, 4Neurology, Emory University, 5Nuclear and Radiological Engineering and Medical Physics, Georgia Institute of Technology, 6Medicine (Div. of Geriatrics and Gerontology), Emory University
Imaging is an increasingly popular tool to develop predictors of treatment outcomes in aphasia (Crosson et al., 2019). Structural imaging-based lesion-symptom mapping (LSM) is a widely-used approach for identifying predictors, but the lack of network information in LSM may omit the holistic details needed to create valid language-network prediction models. Instead, task-fMRI offers neurobiological inspection of lesion impact on the whole language network, simultaneously encoding both domain-general and domain-specific information. This study aims to develop multivariate and mass-univariate analysis on whole-brain task-fMRI to predict treatment outcomes in persons with aphasia (PWA) undergoing intention treatment (INT) that invokes both domain-specific and domain-general aspects of word retrieval using a left-hand circular motion when producing responses. Fourteen English-speaking PWA (>6 months post-stroke) underwent treatment focused on naming pictures and generating category exemplars. Participants were randomized into INT (N=7) or a control treatment (N=7) (Benjamin et al., 2014). Treatment gains were assessed by calculating change in performance on category exemplar generation (CEG) from baseline to 2-weeks post-treatment. Additionally, baseline task-fMRI images were acquired during a CEG task. Images were processed as described in (Krishnamurthy et al., 2021) to compute the z-transformed area-under-the-curve (ZAUC) of task activity for each voxel. Whole-brain ZAUC and CEG change scores for all 14 participants were entered into a mass-univariate or multivariate analysis to investigate which method may be more meaningful on network level data. Mass-univariate analysis involved voxel-wise linear regression between ZAUC and CEG change scores where each voxel is assumed as independent. Multivariate analysis involved sparse canonical correlation implemented in LESYMAP (Pustina et al., 2018), where all voxels are entered into the model simultaneously. All ROI clusters were corrected for multiple comparisons (p<0.05, cluster size=50). Significant gains in CEG were seen in 6/7 INT and 3/7 control participants (Benjamin et al., 2014). Significant task activity could be identified at the network level for each participant (Krishnamurthy et al., 2021). The mass-univariate brain-behavior analysis revealed domain-specific aspects of semantic processing, where greater ZAUC in the left middle temporal gyrus and fusiform gyrus predicted reduced gains in CEG (negative slope). The multivariate brain-behavior analysis was sensitive to domain-general functions, where greater ZAUC in the left precuneus of the default mode network and left superior parietal lobule involved in attention predicted more gains in CEG (positive slope). We have shown the feasibility of using whole-brain, network-level task-fMRI data to compute brain-behavior predictor models. Our preliminary results highlight the need for both mass-univariate and multivariate analysis on whole-brain data as each analysis provides unique neural correlates that are differentially sensitive to domain-specific and domain-general aspects of language processing. Importantly, approaches that allow identification of both domain-general and domain-specific biomarkers offer more precise prediction of language recovery, thus facilitating patient triaging and individualizing treatment planning. This proof of principle study may provide a platform for developing conceptual neurocognitive models that can describe language treatment-induced neuroplasticity in PWA undergoing various treatments. Future work will involve sensitizing the biomarkers to account for INT-specific treatment changes.