Slide Slam A10
Prediction of post-stroke aphasia treatment outcomes is significantly improved by inclusion of local resting-state fMRI measures
Robert Wiley1,5, James Higgins2, David Caplan3, Swathi Kiran4, Todd Parrish2, Cynthia Thompson2, Brenda Rapp5; 1University of North Carolina Greensboro, 2Northwestern University, 3Massachusetts General Hospital, 4Boston University, 5Johns Hopkins University
Introduction: While the use of neural-based measures for predicting response to treatment in post-stroke aphasia (PSA) is of interest for basic science, its utility for clinical purposes is qualified by the relative difficulty and expense of collecting such measures. Recent work with rs-fMRI (e.g., Iorga et al., 2021; Demarco & Turkeltaub, 2020; Guo et al., 2019) indicates that local rs-fMRI analyses (as opposed to rs-fMRI connectivity based approaches) distinguish between healthy and lesioned tissues and index domain-specific language deficits. We investigated a set of local rs-fMRI measures in terms of their ability to contribute unique information for the purposes of predicting response to treatment, above and beyond what can be predicted on the basis of demographic or behavioral measures alone. Methods: 64 individuals with PSA subsequent to a single left-hemisphere stroke were treated for deficits in naming (n=28), spelling (n=22), or syntax (n=14), and completed rs-fMRI scans prior to beginning treatment. Response to treatment was measured as percentage of maximum gain from pre-to-post assessments on trained items. The rs-fMRI data were used to measure the fractional Amplitude of Low Frequency Fluctuations (fALFF; Zou et al., 2008) across the 96 anatomical gray-matter parcels of the Harvard-Oxford Atlas (Desikan et al, 2006). Four predictors were derived from the fALFF: with or without normalization within-participants, and with lesioned tissue excluded or assigned a value of 0. Response to treatment was first predicted, separately for each language domain, using the best set of demographic and behavioral measures (determined by exhaustive search through all variables, e.g., pre-treatment accuracy, age, sex) and prediction accuracy was assessed with cross-validation. The process was repeated including lesion volume and in turn each of the four fALFF predictors, with the best set predictors selected via elastic net regression (Zou & Hastie, 2005). The difference in the precision and 80% prediction intervals of the two sets of models (behavioral only versus behavioral and neural measures) were statistically assessed using Monte Carlo analysis. Results: Median absolute errors (MAEs) for predictions based on behavioral/demographic measures ranged from 11-17% across language domains. MAEs were significantly improved to 1-3% when including fALFF normalized within-participants (excluding lesioned tissue), for all three language domains. Similarly, 80% prediction intervals around the response to treatment narrowed from ±22-32% to ±4-6% (p’s < 0.05), indicating the predictions were both more precise and expressed more certainty. The alternative neural measures (non-normalized FALFF or lesioned tissue expressed as fALFF = 0) did not significantly improve predictions of response to treatment beyond behavioral/demographic measures. Monte Carlo procedures demonstrated these improvements were not attributable to chance or “over-fitting” due to including additional predictors. Conclusions: These results are the first to statistically assess whether local rs-fMRI measures (fALFF) improve predictions of treatment outcomes in aphasia beyond demographic and behavioral measures. For the three language domains tested, normalized fALFF (excluding lesioned tissue) significantly improves precision and provides narrower prediction intervals over demographic and behavioral measures. We suggest this is the type of evaluation that should be applied in considering neural measures for clinical applications.