1003 Dissecting Speech Planning and Articulation Circuits Using Seq2Seq Models

INTRODUCTION: Understanding the cortical mechanisms underlying speech articulation is crucial for developing efficient, tractable speech brain-computer interface (BCI) devices. This study investigates pre-articulatory and articulatory kinematics during single-word production, utilizing sequence-to-sequence (Seq2Seq) models to elucidate the spatiotemporal signatures of articulatory trajectories. METHODS: Intracranial recordings were obtained from patients implanted with depth electrodes in the subcentral and pre-central gyri during single-word speech production. A Seq2Seq model was employed to reconstruct phonemic sequences from neural activity during and before articulation. For higher, clinically translatable results using solely the neural data, we can employ Bayes’ rule and solve for the probability of phoneme sequences given neural data by tying together a phoneme probability model and a reverse Seq2Seq model that predicts neural activity given a phoneme sequence. RESULTS: Seq2Seq models leveraging phonemic transition probabilities model achieved a 21.63% accuracy in word classification for 33 words using pre-articulatory data, compared to 6.5% with a linear classifier. Distinct pre-articulatory and articulatory channels were identified in precentral and subcentral gyri, and temporal dynamics showed separable peaks around 250 milliseconds (ms) prior to articulation and 100ms post-articulation for these cortical regions. We then employed Bayesian decoding methods to improve pre-articulatory word classification decoding performances for words outside the training set. CONCLUSIONS: These findings advance our understanding of speech planning and production, demonstrating that recurrent Seq2Seq models capture intricate neural dynamics and reveal distinct roles for different cortical areas. This architecture provides interpretable knowledge crucial for neurosurgeons in functional surgery and speech mapping. Moreover, it creates an infrastructure for clinically translatable, high-performance brain-computer interfaces with generalizable decoders, enhancing the robustness and applicability of neural decoding models for improved patient outcomes in speech restoration therapies.