Keshav Balaji*, Shivashriganesh P. Mahato*, Divyansha Lachi, Eva L. Dyer
University of Pennsylvania
* Equal contribution
Machine Learning for Health (ML4H) 2026 — Findings Track
Note
Code coming soon. This repository will host the official implementation of iCLR.
Transfer across intracranial EEG (iEEG) recordings is challenging because electrode placement and anatomical coverage vary widely across patients. Learning representations of individual channels provides a way to map these heterogeneous electrode configurations into a common feature space, where channels can be characterized by stable anatomical and physiological properties. However, these properties cannot always be inferred from a channel's activity alone. A channel's relationship to the surrounding recorded population can provide additional context about its anatomical identity. We therefore develop a contrastive learning approach that combines pretrained temporal features with spatial and temporal attention to learn channel embeddings informed by the broader neural population. These embeddings integrate local signal characteristics with relationships to simultaneously recorded channels. Using the Omni-iEEG benchmark, we evaluate our representations on two anatomical localization tasks and show that population-aware representations achieve competitive or state-of-the-art performance for cross-patient channel localization without adaptation. Our results demonstrate that incorporating population context provides a powerful basis for learning transferable channel representations across heterogeneous patients and electrode configurations.