Poster Presentation

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

INVERSE MODELING OF AUDITORY CORTICAL-SUBCORTICAL DYNAMICS FROM SCALP EEG SIGNALS VIA JOINT PSD NETWORKS

Shourya Verma¹, Adarsh Mukesh¹, Michael G. Heinz¹, Ananth Grama¹; ¹Purdue University

Presenter: Shourya Verma

Separating cortical and subcortical contributions from scalp EEG remains fundamentally challenging due to volume conduction and the ill-posed inverse problem inherent to source localization. Traditional methods such as dynamic statistical parametric mapping (dSPM) require subject-specific MRI scans, precise electrode localization, detailed head models, and anatomical preprocessing which limit practical deployment for clinical screening and real-time applications. We propose a dual-decoder contrastive learning framework that learns both spectral and structural features directly from scalp topology, trained on source-localized supervision to enable rapid inference on new recordings. Our architecture maps 63-channel EEG power spectral densities to physiologically structured 256-dimensional embeddings through residual convolutional blocks with channel attention mechanisms. Independent decoders reconstruct cortical (10 ROIs including bilateral Heschl's gyri, superior and middle temporal gyri) and subcortical (3 ROIs: brainstem nuclei and bilateral thalamic medial geniculate nuclei) activity across five frequency bands (delta, theta, alpha, beta, gamma). Training combines mean squared error reconstruction loss with L1 sparsity regularization and a novel power-weighted contrastive loss that enforces structural separation between cortical and subcortical embedding centroids by maximizing their cosine distance. We trained on 20 participants across naturalistic listening contexts (clean/noisy/competing-speaker audiobooks) with held-out participant validation (16/4 split, 2000 non-overlapping windows). The contrastive approach achieved 31% lower mean absolute reconstruction error, 8.6-fold improved correlation versus baseline autoencoder, and near-perfect cluster separability (ROC AUC 1.0). Frequency-band analysis reproduced expected neurophysiological patterns, with cortical dominance progressively decreasing from low to high frequencies. This work demonstrates that meaningful cortical-subcortical representations can be learned from scalp recordings alone, with potential applications in rapid clinical screening, interpretable biomarker discovery across auditory processing hierarchies, and natural extensions to higher-density recording modalities (MEG, HD-EEG, iEEG).

Topic Area: Methods, Tools, Theory & Neural Coding