A rank from the past 60 seconds did not show lower mean background command rate error in eight SEEG test streams at an early target of 3 per eligible minute

We tested a rank of a fixed decoder score among scores from the past 60 seconds. The aim was to reduce later background command rate error without new labels. We used public OpenNeuro ds004473 version 1.0.2 and eight existing SEEG streams. The first 40 percent of stream endpoints supplied model training. The next 20 percent set thresholds. The last 40 percent supplied an offline test. The early target was 3 commands per eligible minute. Mean absolute rate error was 1.27 with fixed scores and 1.60 with ranks. The paired error reduction was -0.34 commands per eligible minute. The 95 percent patient bootstrap interval was [-1.12, +0.23]. This test did not show lower mean rate error. It did not establish that ranks are worse. Fixed scores detected 61 of 340 responses. Ranks detected 62. All nine planned history and target settings are reported. Independent code checked all ranks, commands, and hits. A fresh EDF audit checked all 67,309 feature windows. Small reconstruction differences did not change ranks, commands, hits, or group results after refitting. A mathematical example shows why a rank threshold below one cannot alone limit the future command rate. This reused cohort supplies exploratory evidence. Key presses are event labels. No live control or brain stimulation was tested.

stereoelectroencephalography · brain-computer interface · causal score ranking · command rate · offline replay · OpenNeuro