Learned eight-contact references reduced reference error, but decoder agreement gains were uncertain in eight public SEEG records

We tested whether early calibration can preserve a fixed SEEG decoder after its reference is reduced to eight contacts. We used eight public OpenNeuro ds004473 records. We fit nonnegative sum-one weights on early full-contact data and tested 564 later trials. Learned weights reduced later normalized reference error in every person. The equal-person mean error fell from 0.2153 to 0.1243, a 42.2% reduction. The primary decoder decision disagreement fell from 23.68% with equal weights to 20.82% with learned weights. The paired reduction was 2.87 percentage points, with a descriptive 95% person bootstrap interval of -0.28 to 6.20 and exact paired p=0.15625. The decoder agreement gain remains uncertain. Accuracy rose by 2.07 points, with an interval of -0.29 to 4.65. These are reused records and an offline task-state test. Full-contact calibration is required. The package contains checked source recipes, compact signal data, code, fitted models, all later predictions, and three figures. All raw file hashes passed. Fresh source inputs and repeated results matched the archive. The result concerns brain signal reading, not stimulation.

stereoelectroencephalography · brain-computer interface · signal reference · decoder agreement · calibration · signal processing