Learning Data-Driven Shared Coordinates for Cross-Subject Intracranial Recordings

Intracranial recordings are difficult to aggregate across subjects because electrode count, placement, and regional coverage vary from person to person. Anatomical coordinates such as MNI provide a shared reference frame, but spatial proximity does not guarantee similarity of recorded neural dynamics. We ask a narrower representation-learning question: can the neural data themselves provide a shared coordinate system for heterogeneous intracranial channels? We train a contrastive CNN on 10-s local field potential (LFP) segments from 20 subjects to map each channel into a 32-dimensional data-driven coordinate. Brain-region labels define coarse cross-subject positive and negative relationships during training, but at inference a channel coordinate is computed from its LFP alone; no region label is provided to the encoder. The resulting representation is region structured and generalizes to previously unseen electrodes (52.84\% held-out-channel k-NN accuracy with our selected objective). To test whether this geometry carries information beyond coarse anatomy, we hold out an entire target region and compare otherwise identical pooled transformers conditioned on region labels, MNI locations, or the learned coordinates. Data-driven coordinates significantly improve cross-region waveform prediction in both hemispheres. These results support the following claim: a signal-derived embedding can serve as a useful shared channel coordinate across subjects and preserve distinctions not expressed by region identity or MNI position in the cross region prediction task.