Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Embodiment hierarchy predicts neural decoding accuracy of Chinese word semantics from SEEG: a 9-patient intracranial BCI study using an embodied priming paradigm and LLM semantic embeddings
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
Jingwen Yang1, Leyan Gao1, Weiyi Li2, Rui Feng2, Shuo Lu1; 1Shenzhen University, 2Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China
Introduction Embodied cognition theory suggests that action/sensory words recruit sensorimotor re-activation, whereas abstract words rely more on amodal association cortices. This dissociation should affect the decodability of lexical semantics from intracranial signals, a key issue for naturalistic language BCIs. Using SEEG in native Chinese speakers with a structured embodied priming paradigm, we tested the hypothesis that words with higher embodiment strength yield systematically higher decoding accuracy, dependent on sensorimotor cortical contributions. Methods Nine refractory epilepsy patients (4 female, mean age 29.1 years, normal language/cognition) undergoing SEEG monitoring at Huashan Hospital were implanted with electrodes covering left language and sensorimotor regions (Broca, Wernicke, premotor, M1, S1, insula, angular gyrus, anterior temporal lobe). Stimuli: 60 Chinese target words organized into 12 semantic categories (hand, mouth, foot, eye, fire, water, etc.). Each category had 5 target words presented in fixed embodiment order: strong (L1, direct bodily action/sensory, e.g., “grasp”), medium (L2, indirectly embodied, e.g., “read”), low (L3, weak perceptual metaphor, e.g., “arrange”), and abstract (L4, two words per category, e.g., “understand”, “communicate”). A prime word (category label) preceded each block to activate embodied semantics. Trial structure: prime (1s) → ISI (1s) → five target words each for 3s with 1s ITI. Patients silently read and mentally rehearsed meaning. SEEG high-gamma power (70-150 Hz) was extracted from 200-800 ms post-stimulus. A patient-specific CNN-LSTM decoded the word's semantic embedding extracted from a Chinese-adapted Llama 2 7B model (4096-dimensional). Decoding accuracy measured as cosine similarity between predicted and true Llama embeddings. A linear mixed-effects model (LME) tested embodiment level (4 levels) effect with random intercepts for patient and word. Results Across 9 patients (660 trials; 15 strong, 15 medium, 15 low, 30 abstract words after balancing), mean decoding cosine similarity was: strong 0.79±0.09, medium 0.70±0.11, low 0.62±0.12, abstract 0.55±0.13 (F(3,30)=28.7, p<0.0001, repeated-measures ANOVA). Pairwise comparisons: strong > medium (p=0.02), medium > low (p=0.01), low > abstract (p=0.04). LME showed a linear embodiment effect: each 1-step increase in embodiment level (1=abstract to 4=strong) raised cosine similarity by 0.074 (β=0.074, SE=0.008, p<0.01), controlling for word frequency and stroke number. Time-resolved decoding revealed embodiment advantage emerging at 150-250 ms (strong vs. abstract: Δ=0.13, p=0.003) and persisting beyond 600 ms. When SEEG contacts in sensorimotor cortex (pre/postcentral gyrus and insula) were excluded, decoding accuracy for strong words dropped by 36% (0.79→0.51), whereas abstract words dropped only 6% (0.55→0.52; interaction p<0.001). No significant effect of lesion side or electrode number was found. Conclusion Strongly embodied Chinese words are decoded from SEEG signals with significantly higher accuracy and earlier temporal onset, critically dependent on sensorimotor cortical activity. Using a Llama-based semantic embedding space, we provide the first direct neural decoding evidence that embodiment strength predicts semantic decodability. These findings offer a principled, theory-driven strategy for optimizing language BCIs - prioritizing high-embodiment vocabulary for high-performance decoding and developing embodiment-adaptive decoders for abstract semantics.
Topic Areas: Meaning: Lexical Semantics,