Poster No:
25
Submission Type:
Abstract Submission
Authors:
Linyang He1, Alan Bush1, Latane Bullock1, Yanming Zhu1, Yuanning Li2, Robert Richardson1
Institutions:
1Massachusetts General Hospital, Boston, MA, 2ShanghaiTech University, Shanghai, Shanghai
First Author:
Linyang He
Massachusetts General Hospital
Boston, MA
Co-Author(s):
Alan Bush
Massachusetts General Hospital
Boston, MA
Introduction:
The basal ganglia (BG), long associated with motor control, has been less emphasized in language processing compared to cortical regions. However, research has established links between the BG and language-related cortical areas such as the inferior frontal gyrus and other prefrontal regions (Ullman, 2006), indicating its potential involvement in language functions. This has spurred interest in the role of the BG, particularly the subthalamic nucleus (STN), in language processing. While recent studies have identified the STN's participation in motor aspects of speech production and lexical semantics (Chrabaszcz et al., 2021; Lipski et al., 2018), its extent in high-level language aspects like syntax and contextual semantics remains an area of exploration.
On another front, deep language models (DLMs), emerging as a new tool in computational neuroscience, have become a powerful lens for exploring brain functions related to language processing (Goldstein et al., 2022; Li et al., 2023). The current project aims to leverage DLMs to investigate the role of the STN in language processing.
Methods:
We used local field potential (LFP) recordings during deep brain stimulation (DBS) surgery targeting the left STN in two Parkinson's Disease patients. Patients engaged in a sentence repetition task: repeating 10 sentences from the Harvard Psychoacoustic Sentences set. We analyzed theta, beta, and high gamma frequency bands of the STN-LFP data when patients were articulating sentences, correlating LFP with linguistic features derived from GPT-2 large (Brown et al., 2020).
Four types of embeddings from GPT-2 were used: full, lexical, syntactic, and residual contextual. An L2-regularized linear regression model reconstructed the LFP signals from linguistic features, with the correlation coefficient (R score) quantifying the degree of STN's potential involvement in corresponding language aspects. 5-fold cross-validation was applied to obtain reliable R scores.
Results:
Our analysis revealed significant correlations across all linguistic features in theta and beta bands for all patients compared to permutation baseline (p < 1e-5). For patient one, the theta band showed the most robust correlation (R=0.39±0.04, mean±std) and the beta and high gamma bands showed average R scores of 0.22±0.04 and 0.28±0.02, respectively. The second patient, despite a dominant theta band in language processing, showed a lower high-gamma R score of 0.07±0.02, possibly due to speech impairment.
Linguistic feature analysis indicated that lexical embedding had lower R scores (0.272 across frequency bands), while syntactic, residual contextual, and full embeddings exhibited similar higher R scores (0.299, 0.301, and 0.299, respectively). Paired-samples t-test indicated a difference between the lexical and three other features (p = 0.0025). Different from cortical studies, we found all Transformer layers of DLM encoded STN features similarly.
Temporal dynamics analysis, extending word onset to 100~600ms pre-onset, showed that lexical features' R scores remained relatively stable, whereas the scores for other higher-level linguistic features exhibited a strong downward trend. This may suggest that the STN's role in lexicon processing is persistent throughout speech production, whereas its involvement in higher-level language processing is more immediate and transient, differing from cortical processing patterns.
Conclusions:
Through the lens of DLM, we found STN theta power can be predicted from both lexical-level and high-level language features. Interestingly, these features predict theta band power better than either beta or gamma power. Our results also suggest that the STN exhibits distinct temporal dynamics and correlations with DLM features compared to the cortex. This study is the first to apply DLMs in dissecting the neural substrates of language within the BG, offering a novel methodological approach that could broaden our understanding of subcortical structures' role in language processing.
Brain Stimulation:
Deep Brain Stimulation 1
Language:
Language Comprehension and Semantics 2
Speech Production
Modeling and Analysis Methods:
Other Methods
Novel Imaging Acquisition Methods:
Imaging Methods Other
Keywords:
Basal Ganglia
Cognition
Computational Neuroscience
ELECTROPHYSIOLOGY
Language
Machine Learning
Other - deep language model
1|2Indicates the priority used for review
Provide references using author date format
Brown, T.B., et al. (2020). Language Models are Few-Shot Learners. arXiv.
Caucheteux, C., et al. (2021). "Disentangling syntax and semantics in the brain with deep networks." In Proceedings of the 38th International Conference on Machine Learning, PMLR, 1336–1348.
Chrabaszcz, A., et al. (2021). "Simultaneously recorded subthalamic and cortical LFPs reveal different lexicality effects during reading aloud." Journal of Neurolinguistics, 60, 101019.
Goldstein, A., et al. (2022). "Shared computational principles for language processing in humans and deep language models." Nature Neuroscience, 25, 369–380.
Li, Y., et al. (2023). "Dissecting neural computations in the human auditory pathway using deep neural networks for speech." Nature Neuroscience.
Lipski, W.J., et al. (2018). "Subthalamic Nucleus Neurons Differentially Encode Early and Late Aspects of Speech Production." Journal of Neuroscience, 38, 5620–5631.
Tankus, A., et al. (2019). "Degradation of Neuronal Encoding of Speech in the Subthalamic Nucleus in Parkinson’s Disease." Neurosurgery, 84, 378–387.