Independent contact models gave unequal bipolar prediction gains at 20 ms with 100 ms histories in two SEEG rest records

Two public SEEG rest records supplied contact power for independent contact models. The models gave unequal prediction gains between bipolar channels that share a contact. The main test used a 20 ms forecast and 100 ms histories of 4 ms voltage means. The first and second 120-second periods were tested separately. Mean relative error reductions over 32 shared directions ranged from 0.136% to 0.216%. The largest gain was 2.689%. Channels with separate contacts gave zero gain. A control that preserved contact variance but gave contacts the same normalized time course also gave zero gain. A proof and independent AR(1) simulation explain how shared contacts and different time scales can give one-way prediction. These are exact finite-history linear projection scores in data-based stationary models. They are not held-out decoder scores or evidence of neural pathways. This new analysis reuses a fixed cohort and contact selection from a prior sci package. Both full source EDF hashes and all 23,040,000 selected native samples were checked. A clean replay matched every result table and covariance array. The package preserves code, compact source samples, metadata, results, three figures, a proof, and repeat instructions. No stimulation was tested.

stereo-electroencephalography · bipolar montage · directed prediction · finite-history models · shared contacts · brain-computer interfaces