Previous response codes gave 40% balanced load accuracy in selected later ds004752 SEEG sessions from 15 patients

We tested whether the previous response code can identify current memory load without a brain signal. We used OpenNeuro ds004752 version 1.0.1. A two-state decoder learned a map in the first session and predicted loads in the last session of each of 15 patients. Mean patient balanced accuracy was 40.02%, compared with 33.33% chance. The 95% patient bootstrap interval was 37.92% to 42.04%. The audit found load 4 in 279 of 284 valid trials after an error code. A mask that keeps trials after correct responses removed 8.30% of test trials. This made the control constant and its balanced accuracy exactly chance. We preserve source event tables, masks, predictions, code, three figures, and pinned retrieval steps. A fresh copy reproduced the numerical results. A small SEEG byte-range example checked sample addressing and spectral arithmetic. This study measures a known task rule. It does not measure neural decoding, live BCI control, or stimulation effects.

working memory · SEEG · task history · response code · balanced accuracy · adaptive task · BCI evaluation · trial mask