Can we predict responses to language interventions? Should we?
Organizers: Thomas M.H. Hope1 & Jenny Crinion1; 1University College London
Presenters: Jade Digman, Nadine Martin, Janina Wilmskoetter, Anne Billot, Sigfus Kristinsson, Thomas Hope
Traditional group studies of language (re)learning often fail to capture individuals’ variance in response and in the case of aphasia (acquired language disorder post stroke) we frequently see much stronger benefits in individuals than the group-level effects might suggest. If these individual effects are systematic and repeatable, they suggest that language learning (treatment) studies’ efficiency might be improved, perhaps dramatically, by targeting interventions to those who are most likely to benefit. This symposium aims to bring together those working (or interested in working) in this emerging field, seeking consensus on its key methodological and ethical challenges. We will ask whether and how we can derive credible, predictive results from the typically small samples employed in this domain. And we will consider how best to address patients’ concerns that these results might eventually be used to withhold treatments from those judged unlikely to benefit from them.
Behavioural predictors of aphasia recovery and treatment response
Jade Digman1, Amy Rodriguez2, Kate O’Brien1, Penni Burfein1,3, David Copland1; 1University of Queensland, Australia, 2Centre for Visual and Neurocognitive Research, Atlanta, USA, 3Surgical Treatment and Rehabilitation Service, Herston, Australia
Individual response to aphasia treatment is highly variable and there is no current method that can be applied clinically to accurately identify treatment response for individuals prior to therapy. In addition to baseline demographic, behavioural and brain-related predictors, measuring early behavioural treatment response may provide a simple method for predicting overall treatment outcome. We examined the relationship between early naming probe accuracy (after 3 hours of impairment therapy) and treatment outcome for 32 individuals with chronic post-stroke aphasia. Participants received 48 hours of impairment, functional, computer and group-based therapy with therapy primarily targeting word retrieval. Linear regression models including the factors lexical-semantic function, age, and early probe accuracy demonstrated that early probe accuracy was a significant predictor of treatment outcome. These findings will be discussed in relation to clinical applications and other potential predictors including functional and structural brain imaging.
A computational account of naming impairments in aphasia, including model-based diagnosis, treatment, and post-treatment assessment.
Nadine Martin1, Julie Schlesinger1, Jessica Obermeyer2, Robert Wiley2, Gary Dell3; 1Temple University, USA, 2University of North Carolina, USA, 3University of Illinois, USA
Recent treatments for word production impairments in aphasia have targeted properties of lexical activation, following the Semantic-Phonological model’s characterization of naming impairments as slowed transmission or poor maintenance of activation that supports word retrieval. We demonstrated effects of impairment to these activation parameters using picture naming after a response delay. We examined the Semantic-Phonological model’s account further in a treatment for BD82, whose naming accuracy increased and rate of nonword productions decreased after a 5-second delay (the hallmark of a transmission deficit). We identified parameters of the model that simulated her pre-treatment response profile at the 1-second delay. After treatment, naming accuracy increased after the 1-second delay and nonword errors decreased significantly. We fit this pattern to the model by increasing the weight of the Phonological parameter from its pre-treatment setting. We discuss the potential of this model to predict changes in naming following treatments that directly target lexical activation.
Using residual neural networks to predict aphasia recovery after stroke
Janina Wilmskoetter1, Leonardo Bonilha1; 1Medical University of South Carolina, USA
Individuals with post-stroke aphasia oftentimes present significant variations in recovery, i.e., response to language treatment. Stroke lesion characteristics (such as location, volume, chronicity) only partially explain these observed inter-individual differences. Here, we postulate that aphasia severity and recovery depend on the degree of brain health beyond the stroke lesion, with brain health being defined as the integrity of preserved brain tissue. We will discuss common structural brain pathologies (i.e., small vessel brain disease), and their effect on aphasia recovery that take place independently from the primary stroke lesion. Further, we will discuss how these pathologies may cause variations in the topology of residual neural networks (i.e., decline in long-range white matter fibers, direct connections; and the controllability of brain networks). Our research indicates that personalized rehabilitation may be possible by shifting our focus from what is lost (lesioned brain areas) to what is preserved (residual brain areas).
Multimodal behavioral and neural data predict response to treatment in post-stroke aphasia
Anne Billot1, Stan Lai1, Maria Varkanitsa1, Emily Braun1, Brenda Rapp2, Todd Parish3, Ajay Kurani3, James Higgins3, David Caplan4, Cynthia Thompson1, Prakash Ishwar1, Margrit Betke1, Swathi Kiran1; 1Boston University, USA, 2Johns Hopkins University, USA, 3Northwestern University, USA, 4Harvard Medical School, USA
In this study, we use machine learning models to compare the independent and complementary prognostic role of initial severity of language impairments, demographics and structural/functional integrity of the brain in predicting response to language treatment after a stroke. 55 patients with aphasia were were characterized as responders or nonresponders to treatment based on their percent change in treatment probe accuracy. Support Vector Machine (SVM) and Random Forest (RF) models were constructed to predict treatment response labels. Input features sets included aphasia severity, cognitive composite scores, demographic and multimodal neuroimaging data such as lesion volume, percent spared in gray and white matter regions, diffusion-based fractional anisotropy and resting-state functional MRI (rs-fMRI) data. Our results show that combined behavioral, imaging and demographic models outperform single feature predictors. Functional connectivity at rest seems to be an important predictor of responsiveness to treatment, both alone and when combined with other patient-related data.
Predicting response to impairment-based aphasia therapy: Who doesn’t respond?
Sigfus Kristinsson1, Dirk den Ouden1, Christopher Rorden1, Argye Hillis2, Leonardo Bonilha3, Gregory Hickok4, Julius Fridriksson1; 1University of South Carolina, USA, 2Johns Hopkins University, USA, 3Medical University of South Carolina, USA, 4University of California, Irvine, USA
The question of who responds to impairment-based aphasia therapy has garnered little attention in the aphasia therapy literature. This may, in part, stem from ethical concerns regarding how such research findings will affect clinical management of aphasia. Notwithstanding, identifying factors that dissociate between those who respond and do not respond to conventional impairment-based therapy may be a critical step toward developing more efficient therapy approaches for the latter group. Here, we used LASSO regression to classify 106 individuals with chronic aphasia as therapy (i) responders or (ii) nonresponders based on baseline lesion data, functional activity, resting-state, and structural connectivity measures acquired prior to six weeks of impairment-based aphasia therapy. The predictive value of each neuroimaging modality was examined separately, and model accuracy was statistically compared across prediction models. A multimodal prediction model combining all neuroimaging modalities together was subsequently developed. Our findings will be discussed with reference to ethical concerns.
Predicting individual responses to treatments for aphasia: one method to rule them all?
Thomas Hope1, Ajay Halai2, Matthew Lambon-Ralph2, Jenny Crinion1; 1University College London, UK, 2University of Cambridge, UK
There is often significant inter-individual variance in studies of treatments for acquired language impairments (aphasia). If this variance can be predicted pre-treatment, the implication might be that we could target particular treatments to those patients most likely to benefit from them. Here, we use Partial Least Squares (PLS) regression models to predict individual responses to four different treatments (two for naming impairments, and one each for reading and comprehension impairments, respectively), in four independent samples of patients. Our approach appears effective in that it yields predictions that significantly out-perform a null model, which simply predicts the mean treatment response for every patient. Our predictions also significantly out-perform prior results for one of the four datasets (focused on reading impairment), found using step-wise feature selection and multiple linear regression. These results suggest that individual variance in many aphasia treatment studies might be systematic and predictable.