Presentation Type
Oral Presentation
Category
STEM (science, technology, engineering, mathematics)
Abstract/Artist Statement
Individual differences in second-language (L2) proficiency are not always well captured by traditional voxel-wise fMRI analyses. Network-level representations provide a more stable and interpretable framework for characterizing how the brain supports L2 processing, particularly in the presence of substantial inter-individual variability. In this study, we investigated whether large-scale brain networks can differentiate advanced and intermediate L2 speakers, and which network features carry the strongest proficiency-related signal.
Using fMRI data from Spanish–Basque bilinguals performing a semantic animacy judgment task, we applied group-level Principal Component Analysis (PCA) across all participants and both language conditions to identify shared spatial brain networks. Subject-specific component time courses were extracted and used to construct within-subject Δ(L2−L1) features, following the logic of classic bilingual contrasts (L2>L1 / L1>L2). This formulation acts as a normalization step, reducing stable individual differences in overall signal magnitude and variance, and allowing network effects to be interpreted as language-specific shifts associated with proficiency.
We evaluated static network engagement and connectivity features using logistic regression with rigorous cross-validation and permutation testing. Static Δ(L2−L1) network features reliably predicted proficiency above chance, whereas full connectivity models did not. A targeted connectivity analysis focusing on a single, empirically identified network further improved classification performance, demonstrating that selective network interactions can add meaningful information without introducing excessive noise.
Together, these findings indicate that proficiency differences are most strongly reflected in how specific large-scale networks are differentially engaged during L2 relative to L1, rather than in broad connectivity patterns. This work highlights the value of within-subject, contrast-based network features for studying bilingual language processing and provides a principled framework for decoding proficiency from fMRI data.
Mentor Name
Lucy Owen
A Network-Based fMRI Approach to Decoding Second-Language Proficiency in Advanced and Intermediate Speakers
UC 331
Individual differences in second-language (L2) proficiency are not always well captured by traditional voxel-wise fMRI analyses. Network-level representations provide a more stable and interpretable framework for characterizing how the brain supports L2 processing, particularly in the presence of substantial inter-individual variability. In this study, we investigated whether large-scale brain networks can differentiate advanced and intermediate L2 speakers, and which network features carry the strongest proficiency-related signal.
Using fMRI data from Spanish–Basque bilinguals performing a semantic animacy judgment task, we applied group-level Principal Component Analysis (PCA) across all participants and both language conditions to identify shared spatial brain networks. Subject-specific component time courses were extracted and used to construct within-subject Δ(L2−L1) features, following the logic of classic bilingual contrasts (L2>L1 / L1>L2). This formulation acts as a normalization step, reducing stable individual differences in overall signal magnitude and variance, and allowing network effects to be interpreted as language-specific shifts associated with proficiency.
We evaluated static network engagement and connectivity features using logistic regression with rigorous cross-validation and permutation testing. Static Δ(L2−L1) network features reliably predicted proficiency above chance, whereas full connectivity models did not. A targeted connectivity analysis focusing on a single, empirically identified network further improved classification performance, demonstrating that selective network interactions can add meaningful information without introducing excessive noise.
Together, these findings indicate that proficiency differences are most strongly reflected in how specific large-scale networks are differentially engaged during L2 relative to L1, rather than in broad connectivity patterns. This work highlights the value of within-subject, contrast-based network features for studying bilingual language processing and provides a principled framework for decoding proficiency from fMRI data.