Interpreting Shared--Private Neural Representations Reveals Preferential Cross-Region Reorganization under Deep Brain Stimulation

Learned neural representations are increasingly used to summarize high-dimensional brain activity, but whether interpreting these representations can yield reliable scientific insight remains unclear. We study this question in human intracranial recordings during deep brain stimulation (DBS), asking whether stimulation preferentially reorganizes activity that is shared across brain regions or activity that is specific to individual regions. We use SPIRE (Shared-Private Inter-Regional Encoder), a nonlinear, time-resolved model that decomposes multi-region recordings into shared and private latent trajectories. We validate the shared-private interpretation using synthetic ground truth and held-out human recordings; model fitting and selection use only Off-stimulation activity. Applying the learned representation to stimulation recordings reveals significantly larger Off-to-On changes in shared than private latent structure during both GPi and STN stimulation, while GPi stimulation condition is also decoded more accurately from shared representations. A matched CTAE analysis reproduces the shared-over-private decoding effect, suggesting that this finding is not specific to SPIRE. Together, these results show that, within the recorded GPi-STN network, DBS-related changes are preferentially expressed in cross-region rather than region-private neural structure.