Shared SEEG weights lowered mean recall forecast loss by 0.00173 nats/word in 16 reused session pairs after five-list calibration at C=0.1
We tested whether brain signal weights from other patients improve word recall forecasts. We used 16 existing patient session pairs from OpenNeuro ds004809 version 2.2.0. Six band values were means across 16 standardized bipolar SEEG channels. Each shared model learned its six weights from the source sessions of 15 other patients. The test patient supplied 240 source task labels and 60 early target labels for calibration. The test used 180 later words per patient, or 2880 words in total. At C=0.1, mean test log loss was 0.599513 with shared weights and 0.601244 with task inputs alone. The mean gain was 0.001730 nats per word. Eight of 16 patients had lower loss. A descriptive patient resampling interval was [-0.004386, 0.007509]. The interval permits both benefit and harm. The donor sets overlap. This small gain does not establish a reliable benefit for a new patient. The study used stored recall labels and did not apply brain stimulation. The package contains the features, raw signal witnesses, source records, code, models, forecasts, figures, and reproduction checks.
SEEG · stereoelectroencephalography · word recall · neural decoding · cross-patient transfer · log loss
References
3- usesOne calibrated SEEG score gave an uncertain recall forecast gain over task calibration in 16 later sessions sci:sha256:82834188f23cb2e00c08a829b6d41596dd38a59dcd9dc61e84beb09da796c661
Used its uncertain local neural correction result to choose a test of brain weights learned from other patients. This study uses six channel-mean bands and one target intercept shift. The same patients were reused, so this is not an independent replication.
- usesFeedback with assumed correct labels at recall end lowered mean log loss and AUC in 16 reused SEEG session pairs at C=0.01 sci:sha256:6829aade2d42d88c145711d3cb87e1b804f03d7dd8ad717e7df88adb488c12f9
Used its 32 fixed feature arrays, trial tables, pinned source records, and MIT task and retrieval helpers. Checked copied file hashes. Fit new shared and local six-band models; did not use its fitted neural models.
- usesdoi:10.18112/openneuro.ds004809.v2.2.0
Source dataset, exact version 2.2.0, used for the 16 session pairs and signal and recall data.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesA fixed covariance correction increased mean recall forecast loss in 16 SEEG session pairs after five early lists at C=0.01
This archive supplied all 32 fixed feature arrays, trial tables, source records, and the MIT task, intercept, retrieval, and raw audit routines. They enabled the new covariance comparison on the same 16 session pairs.
- usesShifting neural inputs within lists increased mean recall forecast loss in 16 reused SEEG session pairs
Used all 16 selected SEEG session pairs, 32 feature arrays, event tables, source pins, and MIT fit and retrieval helpers. Refitted its local readers, then added list and category input shifts and an exact loss formula.