Noise correction increased readout agreement for 20–500 ms SEEG from five pulses with 3-bit dither encoding and lossless overflow in eight patients

We tested whether a known digital noise correction preserves the choice of a repeatable SEEG readout. We used public stimulation data from eight patients with epilepsy. Five early pulses selected a bipolar readout at each site. Five separate later pulses tested its waveform forecast. The primary test used 884 candidate pair instances at 84 sites and 32 dither repeats. With 3-bit symbols and lossless float32 overflow, agreement with the source choice rose from 83.20% to 97.76%. The mean gain was 14.56 percentage points. The descriptive patient bootstrap interval was 7.84 to 21.00 points. Mean later forecast skill changed by +0.892 points, but it fell in 6 of eight patients. The payload used 4.085 bits per input sample before headers. The package contains all candidate voltage samples, a proof of the energy correction, source hashes, scores, figures, and an offline reconstruction recipe; GPL-3.0 scripts are excluded. This is a test of software encoding on processed clinical SEEG. It does not establish benefit in a live brain-computer interface.

SEEG · signal processing · dither · noise correction · readout selection · brain-computer interface