Local calibration lowered mean pre-onset window alerts from 25.6% to 8.8% for a median monitor in 28 SEEG record pairs after onset shafts were excluded
In 28 matched pairs of SEEG seizure records from patients with epilepsy, a median shaft monitor calibrated in the second record had a lower mean pre-onset alert fraction than a monitor transferred from the first record: 8.8% versus 25.6%, a local-minus-transfer difference of −16.9 percentage points (descriptive patient bootstrap 95% interval −25.6 to −8.9; 20 negative, 3 positive, 5 tied patient differences; exact two-sided sign-test p=0.000488281). During the first 20 seizure seconds, mean alert fractions were 47.1% local and 53.2% transferred (difference −6.1 points; interval −15.0 to 2.5). This exploratory test used one record pair per patient, excluded onset shafts, reused an earlier cohort and methods, and did not establish preservation of seizure detection or a general alert-rate guarantee. It tested seizure-state monitoring, not BCI command accuracy, stimulation benefit, or clinical utility.
SEEG · brain-computer interface · calibration transfer · seizure-state monitoring · signal processing
References
3- usesMedian and maximum monitors had mean AUC 0.763 and 0.774 for the first 20 seizure seconds in 30 SEEG records after onset shafts were excluded sci:sha256:c170342b2e116d03b434dd1a1bf0b84f39fccd140306db39a2b40a7c00ffce27
We used its median shaft monitor, separate 30-second scale and threshold intervals, and matched first-20-second test. This study adds 28 second records and compares transferred and local fits.
- usesPower-ratio deviations exceed early limits more often during seizures outside onset-labeled contact groups in 30 SEEG records sci:sha256:46ab84bca2ba1c6a358b6eaf9312db59b5ed68f3f01bca4eb378539fd4ff1a65
We reused its 30 SEEG patients, first-record powers, raw audit samples, metadata, EDF extraction code, and frozen source checks. We checked all 280 publication files before analysis.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
Source data: OpenNeuro ds004100 version 1.1.3; the fixed Git commit and full-file checks are documented in the publication.
Cited by
11 publicationsPublic catalog entries on Sciscape.
- usesA 10-second score offset gave lower mean balanced accuracy for a four-pair seizure reader in 28 SEEG subjects
Its local power-monitor result informed the question. We tested a different supervised reader and isolated a score offset that cannot change per-record score order.
- usesRank bounds certified 79.7% of equally weighted state decisions in 28 SEEG patients against any finite change to one nonnegative shaft score with fixed calibration
We used its matched 28-patient SEEG record pairs, frozen powers and voltage audits, calibration rule, pair selection, source metadata and hashes, and ordinary median reader outputs. We added exact bounds for one changed shaft score and measured decision stability and retained alerts.
- usesMixed windows gave 9.1 percentage points higher mean balanced accuracy than separate records for a four-pair seizure reader in 28 SEEG subjects
Its result on calibration changes between seizure records informed the record-transfer question. This study retrieved source data again and used new code and features.
- usesThree pairs selected in record A had lower mean AUC than all eligible shafts for the first 20 seizure seconds in record B from 28 SEEG patients
Supplied patient and record selection, matched pair names, source channel and event tables, EDF identities and full hashes, and the all-shaft reader with separate fit intervals.
- usesEight high-gamma SEEG pairs selected with ten seconds had 76.7% mean retention across two interictal records in 36 patients
The calibration-transfer question and the rule for removal of a whole seizure-onset contact prefix informed this fixed test. New features came from separate interictal source retrieval.
- usesEight high-gamma SEEG pairs selected with ten seconds had 76.7% mean retention across two interictal records in 36 patients
The calibration-transfer question and the rule for removal of a whole seizure-onset contact prefix informed this fixed test. New features came from separate interictal source retrieval.
- usesTen baseline seconds reduced mean baseline alerts and seizure sensitivity for a fixed four-pair reader in 28 SEEG subjects
The local-calibration result informed the separate-record question. This study tests a different fixed supervised reader with ten baseline seconds and reports both baseline alerts and seizure sensitivity.
- usesThree-second averaging raises AUC for seizure windows 2 to 20 seconds after onset in 28 patients with SEEG recordings
Uses the fixed 28-patient cohort, matched remote bipolar pairs, frozen source provenance, verified band powers, and MIT EDF reader.
- usesA 120-second median calibration left a mean high gamma rate error of 5.1 percentage points for a 10% target in eight SEEG session pairs
Its prior summary helped to motivate a separate test of local calibration for an unlabeled SEEG power threshold. This study uses a different dataset and a different trigger. No code, recordings, or derived features from the prior publication were reused.
- usesEarly baseline correction lowered mean balanced accuracy for a fixed SEEG reader in 28 people
Its local calibration result for a different unsupervised state monitor informed the question. We tested a fixed supervised reader and an additive baseline correction. We did not reuse its monitor or claim the two tests were direct replications.
- usesCommon power and band contrast had mean AUCs of 0.750 and 0.753 in 28 locally calibrated second SEEG records outside onset shafts
This report supplied the 28 matched patient record pairs, band powers, source metadata, digital audit samples, full EDF hashes, and ratio endpoints. We checked all parent file hashes and used these inputs for the new common-versus-contrast comparison.