Early delay burst rate and occupancy did not lower mean recall forecast error beyond power from the same causal filters in 13 patients under source trial order

We tested whether early memory delay bursts improve later recall error forecasts. We used OpenNeuro ds006136 version 1.0.1: 13 patients, 14 sessions, and 143 released bipolar SEEG channels. We assumed that source trial IDs give task order. Models trained on the first 65 percent of matched trials and tested on 751 later trials. Burst rate, burst occupancy, and mean power used the same filters that use past samples and signal windows. Only training trials set thresholds and fit parameters. Mean patient Brier error was 0.179150 for context and power, and 0.183939 after adding bursts. The mean gain was -0.004788. Its descriptive 95 percent patient bootstrap interval was [-0.011080, 0.000837]. The exact two-sided patient sign-flip value was 0.168945. 4 of 13 patients had positive gain. This fixed test did not show a mean gain. The interval includes zero and does not establish equivalence. Source trial processing limits live use. The package has fixed filter coefficients, thresholds, channel and trial features, every test forecast, fitted models, four figures, source hashes, code, and repeat commands. A full raw extraction and model repeat matched all 13 numerical output files exactly.

working memory · SEEG · burst rate · burst occupancy · Brier error · memory BCI · forecasting