Old slopes reduced mean SEEG power error compared with new ridge fits at penalty 10 after 30 calibration seconds in eight people

We tested how much new-session calibration a sparse SEEG power reader needs. We used OpenNeuro ds005624 version 1.0.0. Eight people had two sessions. Each reader sent eight power streams and predicted the same unsent streams. Old training used 120 seconds. New training used 10 to 120 seconds. Testing used 300 later seconds per person. At 30 seconds, old slopes gave mean error 1.621942. New ridge slopes with penalty 10 gave mean error 1.720392. The reduction was 5.7 percent. Seven of eight people improved. The descriptive 95 percent person bootstrap interval for the paired gain was 0.032124 to 0.163793. The exact two-sided sign-flip result was p=0.0390625. Old slopes still had higher mean error than the local-mean reference. A fixed old-model prior gave lower error than the new fit. A later equal-penalty control gave slightly lower mean error without old slope information at 30 seconds. Thus, that prior comparison does not show a mean gain from old slope information. At 120 seconds, old information gave 4.3 percent lower mean error than the equal-penalty control. That result is descriptive. All 576 public source ranges rebuilt all 24 feature and witness archives exactly. The package preserves data, source records, code, models, predictions, checks and five figures. This reused cohort does not provide independent confirmation. The result concerns numerical power reconstruction. It does not establish intent decoding, live control or a stimulation effect.

SEEG · sparse signal transmission · calibration length · ridge regression · power reconstruction · neuroscience