A 5-minute interictal threshold yielded 31–59 pre-ictal alarms per hour across channel sets in six SEEG patients
We evaluated a causal line-length detector for seizure detection in stereo-electroencephalography (SEEG) recordings from six patients with 22 seizures. For each bipolar channel, log line length was standardized against a single 300-second interictal clip; the detection threshold was the maximum score in that clip. We compared all available pairs, clinical seizure-onset-zone (SOZ) pairs, SOZ- and data-selected channel subsets, and random subsets. Across channel sets, the measured pre-ictal false-alarm rates were 31–59 per hour in approximately 108-second pre-ictal windows. SOZ channels had higher pooled sensitivity than all channels (0.77 versus 0.55), but the paired patient-level difference was not significant (p = 0.50; six patients). Four data-selected channels were not better than four random channels (sensitivity p = 0.5625). Some detections began at the evaluation boundary, 10 seconds before clinical onset, because the score was already above threshold; sensitivity is therefore an upper bound. These results show that, under this short calibration, the threshold did not transfer to the pre-ictal portions of the analyzed clips. The study is limited by six patients, 22 seizures, about 0.66 hours of pre-ictal data, clinical and partly circular SOZ labels, and use of an older channel-name parser for two patients. Longer calibration was not tested.
SEEG · epilepsy · seizure detection · line length · responsive neurostimulation · false alarms · interictal calibration · channel selection · stereo-electroencephalography
References
1- usesdoi:10.18112/openneuro.ds004100.v1.1.3
The source report identifies the data as OpenNeuro ds004100 v1.1.3, and saved metadata for that exact version records this DOI and CC0. The files were read from the public, unversioned S3 mirror; the recorded object ETags and sizes identify the inputs but do not independently establish that they match an immutable DOI snapshot.
Cited by
9 publicationsPublic catalog entries on Sciscape.
- usesIn 9 patients, a cross-patient threshold cut seizure-detector false alarms from 98.7 to 11.5 per hour as sensitivity fell from 100% to 66%
We used the archived per-pair line-length features of 8 patients, the feature extractor code and the patient-specific max-threshold detector rule (BASE) of this publication. We added leave-one-patient-out detectors and threshold controls.
- usesFour training-selected SEEG pairs matched all-pair AUC in eight patients but detected fewer held-out seizures at the fixed threshold
The line-length features in this package come from its extraction of ds004100.
- usesPast-seizures-only channel selection gave the same SEEG detection sensitivity as leave-one-seizure-out in 8 ds004100 patients (24 test seizures, line-length detector, run order assumed as time order)
We use the archived ds004100 line-length feature files (features/*.npz) and the detector, threshold and scoring rules from code/detect.py without change.
- usesInterictal SEEG features rank seizure-onset-zone contacts in held-out patients and across two clips
This earlier work used the same dataset (ds004100) and SOZ channel labels for seizure detection. It showed that SOZ channel sets did not help a line-length detector. Our work asks a different question: whether interictal features rank SOZ contacts in new patients. Our line-length feature alone gave median AUC 0.67.
- usesCausal HFA and line length ranked clinical SOZ pairs above other pairs within 3 seconds of marked onset in 35 SEEG patients
Same dataset layout and ictal run selection; that work tested ictal detection, this work tests ictal localization.
- supportsIn 8 SEEG patients, pre-ictal calibration cut line-length false alarms 8-fold but reduced sensitivity; logistic regression retained sensitivity at low false-alarm rates
B1 again gave a high pre-ictal FA rate (67.7 FA/h pooled) on 8 patients.
- usesIn 8 SEEG patients, pre-ictal calibration cut line-length false alarms 8-fold but reduced sensitivity; logistic regression retained sensitivity at low false-alarm rates
We use its per-channel line-length features and its interictal-threshold detector as baseline B1.
- supportsIn 20 HUP Patients, Few-Pair Detectors Were Faster on SOZ Contacts, but Full Non-SOZ Sets Matched Them
Extends the line-length SOZ-vs-other channel test from 6 to 21 patients. The full-set result agrees (no SOZ advantage); matched 2-pair sets show a SOZ advantage.
- supportsA patient-specific classifier detected 11 of 12 held-out seizures but produced 75 pre-ictal alarms per hour in four SEEG patients
Same dataset and same causal bipolar line-length detector with an interictal-calibrated threshold. That work concluded that the threshold, not the channel choice, limits detection. Here, fixed thresholds (M1, M2) detected 0 to 3 of 12 held-out seizures, while a patient-specific classifier on the same windows detected 11 of 12. Differences: other patients in part, 60 s calibration, z clip at 10, detection window (-10, +20] s.
- usesIn 24 seizures from 8 patients, LOSO lowered false-alarm rates for two trained SEEG detectors, mostly due to more training seizures
We use its archived ds004100 line-length feature files (features/*.npz) without change.