Sina Dabiri
BME PhD Proposal Presentation

Date: 2025-08-28
Time: 2pm
Location / Meeting Link: Whitaker Bldg, McIntire Conference Rm

Committee Members:
Zachary Danziger, Thackery Brown, Farzaneh Najafi, May Wang, Tansu Celikel, Audrey Sederberg


Title: Recovering Frequency-Aware Latent Dynamics of Repetition Suppression in Multisensory EEG

Abstract:
Temporal predictability modulates cortical dynamics and underlies repetition suppression (RS), yet linking macroscopic electrophysiological signatures to circuit-level dynamics remains methodologically challenging. Using an existing EEG dataset collected with auditory-cue/tactile-probe protocols that manipulate sensory predictability, this dissertation will develop and validate frequency-aware dimensionality-reduction analyses to recover network-specific dynamics that mediate bottom-up sensory processing and top-down attentional control. We initially validated PCA/TCA, ICA, UMAP, and Autoencoder pipelines using synthetic neural data driven by dynamic latent variables (i.e., temporal smoothing (τ) options in Morrell Model), which guided our selection of, dimensionality-reduction method and reconstruction diagnostics. Preliminary analyses on empirical and synthetic data indicate that commonly used linear approaches (PCA/ICA) reliably capture low-frequency structure but systematically underrepresent power above ≈8 Hz, produce negative R² in the gamma band (30–50 Hz) for τ = 0.5 and 1.0 s, and yield modest coherence between recovered components and ground truth (≈0.15–0.4). Perturbation analyses show that reconstruction fidelity is highly sensitive to τ, and that simple per-frequency confidence-thresholding across parameter sweeps fails to rescue high-frequency recovery. In the proposed work, I will first examine the recoverability and interpretability of latent variables inferred from a synthetic EEG data and then spatiotemporal EEG data across the multisensory and uncertainty contrasts in the dataset, with particular emphasis on delineating parameter regimes and empirical constraints that determine when recovered dynamics can be linked to early sensorimotor processing (central regions) versus later frontal attentional engagement. Outcomes will include systematic characterizations of method performance across frequency bands, empirically grounded limits on interpreting high-frequency latent structure, and open-source code and notebooks to support reproducible analysis. These contributions aim to reduce methodological ambiguity in latent-state studies of spatiotemporal EEG data and to examine the link between latent electrophysiological states and cognitive variables arising from temporal and stimulus uncertainty.