Aphasia after stroke commonly impairs word retrieval, but cognitive models of word production often assume that underlying conceptual and semantic representations remain relatively intact. The current study evaluated whether concept-level semantic structure can still be decoded from functional MRI signals in individuals with chronic post-stroke aphasia. The investigators aimed to determine which semantic models best capture the neural representational geometry that arises when participants silently generate semantic features for pictured nouns.
The experiment included two groups: eight healthy adults and six individuals with chronic aphasia. Participants completed a dense-sampling fMRI protocol in which they viewed 57 pictured nouns and silently generated semantic features associated with each picture. The task was designed to evoke concept-level semantic processing while avoiding overt speech production, allowing comparison of neural representational geometry between controls and people with aphasia.
BOLD signals were acquired during the covert semantic feature generation task. The authors applied representational similarity analysis (RSA) to compare the geometry of neural response patterns with candidate semantic models. RSA permits assessment of how closely the structure of neural dissimilarities between concepts matches the structure predicted by different semantic models. Representational similarity decoding was also used to test whether concept identity could be recovered from neural patterns in each group.
Three classes of semantic models were evaluated against the neural representational geometry:
The purpose of comparing these models was to identify which type of semantic representation best explains brain activity during covert semantic feature generation in both healthy controls and individuals with chronic aphasia.
Beyond model fits, the authors performed representational similarity decoding to assess whether individual concept identities could be recovered from neural data. Decoding measured whether the pattern of neural responses contained reliable information about which concept a participant was silently considering. Decoding performance was examined separately for healthy adults and for participants with aphasia.
Representational similarity analysis indicated that the experiential model (Exp48) provided the best match to neural representational geometry in both people with aphasia and controls. Exp48 outperformed the taxonomic model (WordNet) and the distributional models (Word2Vec, GloVe) in explaining the structure of BOLD response patterns during the covert semantic task.
Representational similarity decoding demonstrated that concept identity could be recovered from neural patterns well above chance in both groups. This result shows that despite chronic aphasia and lesion-related anatomical changes, concept-level information remained present and decodable in the measured BOLD signals.
The authors reported no relationship between decoding accuracy and language measures obtained from individuals with aphasia. That is, decoding performance did not correlate with the language assessment metrics reported for the aphasia group in the source.
Although concept identity was decodable in participants with chronic aphasia, the study did not find an association between decoding accuracy and the language measures reported for those individuals. The source does not provide additional details about specific language tests or numerical correlation results beyond stating the absence of a relationship.
These findings support the view that concept-level experiential semantic structure remains robustly represented in the brain of people with chronic post-stroke aphasia. The study suggests preserved conceptual representations can coexist with lesion-related language impairments and altered anatomy. The superior fit of the experiential model (Exp48) across both groups underscores the role of experiential feature structure in shaping neural semantic geometry during covert semantic retrieval.
Clinically, the results indicate that interventions targeting conceptual-semantic processing may be informed by the persistence of decodable experiential representations, although the source does not provide or endorse specific therapeutic strategies. Because the study used covert semantic generation and decoding from BOLD data, translation to behavioral or rehabilitative outcomes would require further investigation.
This report is a preprint and has not undergone peer review. The sample sizes were small (eight controls, six individuals with aphasia) as reported in the source. The source does not report additional methodological or demographic details beyond those summarized here.
Funding sources declared in the preprint include NIH/NIDCD (F31DC021613), William Orr Dingwall Dissertation Fellowship, Academy of Aphasia Barbara Martin Aphasia Research Grant, University of Pittsburgh Clinical and Translation Science Institute pilot grant, and a Carnegie Mellon and University of Pittsburgh brain imaging center seed grant. The authors declared no competing interests in the source.
Note: All statements above summarize findings and details reported in the cited preprint. Because this article is a preprint, conclusions should be interpreted in light of the absence of peer review.