Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Replicable linguistic neural encoding at the individual-subject level during speech comprehension
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Maya Inbar1, Yi Li1, Laura Gwilliams2, Alexandra Woolgar1; 1University of Cambridge, 2Stanford University
Language function demands the dynamic orchestration of multiple levels of representation. Advances in neuroimaging and machine learning offer a way to characterize receptive language processing during passive listening, providing insight into the representation of these levels and their interrelations. For example, recent group studies of neurotypical individuals reveal that language processing follows a spatiotemporally hierarchical dynamic code (Gwilliams et al., 2025; Karunathilake et al., 2025). Passive listening paradigms hold potential to shed light on receptive language skills without requiring overt responses which may be useful clinically. However, to adopt these methods in translational research, we need to shift perspective from group studies to individual-participant characterizations. In particular, we have little data on the robustness of these methods for characterizing receptive language processing at the individual participant level. Here, we study linguistic encoding across a rich array of levels in individual neurotypical and neurologically healthy participants. We investigate the extent to which different levels of language representation in the brain can be characterized from a limited amount of individual-participant data, and the stability of these characterizations across two recording sessions. Participants (N=18) listened to fictional short stories in American English over the course of two 1h long MEG sessions (Gwilliams et al., 2023). We annotated each word in the stimulus set with 36 linguistic features spanning different levels of linguistic representation: phonetic form, word form, part of speech, syntax and semantics. We employed a back-to-back regression algorithm (King et al., 2020) to characterize the timepoint-by-timepoint neural representation of the 36 linguistic features while controlling for their co-variation. We analyzed MEG data from each 1h recording session separately and compared the results within each individual against an empirical null distribution created by permutation. We carefully designed our algorithm and permutation process to reflect linguistic dependencies across features and words. Each of the 36 linguistic features was represented in at least one single session (1 hour) of individual-participant MEG activity. However, the prevalence of significant neural representation in individual datasets varied across features. For example, we found significant neural representation of semantics in 55-100% of the datasets, whereas we found significant neural representation of syllable count in <25% of the datasets. When we assessed test-retest reliability, we found that individual-subject representational dynamics of significantly represented features were stable across the two recording sessions. This was true of features that are tightly related to sensory processing, such as phonetic features, and features of greater abstraction, such as the part of speech of a word. Our results suggest that passive listening is a promising avenue for tracking linguistic processing in individuals, with implications for both basic science and clinically-oriented domains. The data suggest that as little as 1 hour of MEG data can be used to characterize individual representation strength and dynamics across a broad range of linguistic levels during passive listening. Furthermore, determining the detection rates on a feature-by-feature basis in neurotypicals may provide a baseline for future comparisons in clinical populations.
Topic Areas: Speech Perception, Computational Approaches