The neurocognitive mechanisms underlying semantic feature generation in persons with aphasia

Word production deficits are a cardinal feature of aphasia, a communication disorder affecting 2.4 million Americans (Simmons-Mackie & Cherney, 2018). Semantic Feature Analysis (SFA; Boyle & Coelho, 1995) is one of the most used treatments for word production deficits in aphasia (Tierney-Hendricks et al., 2021). The goal of SFA is to improve spoken word production by guiding persons with aphasia to produce semantic features related to treatment target (traditionally, a concrete noun). The efficacy of SFA is well supported by meta-analyses including over 50 participants (Efstratiadou et al., 2018; Quique et al., 2019). Feature generation is a key active ingredient of SFA (Boyle, 2010; Evans et al., 2021, Cavanaugh, Swiderski et al., 2022) and the treatment’s mechanism of action is hypothesized to be spreading activation between semantic concepts (Boyle et al., 2022). However, the nature of the information encoded during spreading activation remains unknown and characterization of the brain networks supporting this activity is incomplete. One way to address this knowledge gap is to test whether competing semantic models (taxonomic, distributional, or experiential) models better fit neural data elicited by semantic feature generation. Taxonomic, distributional, and experiential models represent concepts by category membership, patterns of co-occurrence, and sensory, motor, and affective experiences, respectively. Using representational similarity analysis, Fernandino and colleagues (2022) found that experiential feature-vectors correlated more strongly with neural data elicited by a concept-familiarity task than did taxonomic and distributional feature-vectors. Distributional feature-vectors have also been used to match BOLD signal elicited by an SFA-analogous covert semantic feature generation task to individual concept labels with over 80% accuracy (Anderson et al., 2016). We propose to extend these findings to people with aphasia and age-matched healthy adults. Aim 1 will identify the semantic model that most strongly correlates with BOLD signal elicited by semantic feature generation. We hypothesize that experiential and distributional models will show the highest agreement with the neural data for both participant groups. In Aim 2 we will decode the BOLD signal elicited by this task associated with single object concepts; we hypothesize that decoder accuracy will exceed 80% for healthy adults and be well above chance for participants with aphasia. In aim 3 we will identify relationships between decoder accuracy and theoretically and clinically relevant psycholinguistic abilities along with critical patient characteristics in persons with aphasia. We hypothesize that decoder accuracy will correlate more strongly with semantic than phonological language tasks and negatively correlate with patient characteristics such as lesion volume. This study will improve our understanding of the neurocognitive mechanisms underlying SFA and support future developments, including the creation of in silico models enabling the study of normal and impaired lexical semantic processing and its downstream effects on other language systems.