Current inputs from up to eight other shafts increased mean voltage error for simulated 50 ms loss of one SEEG shaft in 35 records after a 2 s ridge fit

We tested voltage recovery when one electrode shaft lost samples. We used 35 public SEEG records from 21 people in OpenNeuro ds006519 version 1.0.0. The test used 20 separate 50 ms loss simulations per record. Each model used two seconds for fitting and a sparse native-sample grid. The main ridge penalty was 0.1. The joint model added current signals from up to eight other shafts to 100 ms of past target voltage. Mean error increased from 0.993 to 1.828 after division by fit target variance. This was an 84.2% increase in the ratio of group means. Twenty of 21 people had higher error. The paired effect was -0.836, with a 95% person-bootstrap interval of [-1.391, -0.452]. All four planned sensitivity tests had the same direction. Holding the last target sample gave lower mean error than either fitted model. The package preserves source hashes, input samples, code, fitted states, predictions, scores, four figures, and reproduction instructions. Independent checks and a full replay passed. The result concerns this fixed method and simulated loss. It does not show performance during stimulation or a live BCI task.

SEEG · intracranial EEG · voltage reconstruction · signal loss · ridge regression · benchmark · brain-computer interface