In 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 tested if the high false-alarm (FA) rate of a simple SEEG seizure detector comes from the calibration data or from the detector. This is important for closed-loop stimulation devices. Data: OpenNeuro ds004100 v1.1.3, 8 HUP patients, 32 SEEG seizures, interictal run-01. We used the per-channel line-length features of sci publication 99d96b92 (2 s causal windows, bipolar pairs). We held out one seizure at a time. B1 uses the max line-length z over channels with a threshold from the interictal clip. B2 uses the same score, but the threshold also includes the pre-ictal windows of the training seizures. B3 is an L2 logistic regression on all channels, trained on the other seizures of the same patient, with a threshold at the max training-negative score. Results: B1 detected 19 of 32 seizures with 67.7 pre-ictal FA/h. B2 detected 14 of 32 with 8.3 FA/h, and it detected 0 seizures in 4 of 8 patients. B3 detected 29 of 32 with 24.0 FA/h (median latency 4 s). In a threshold sweep, B3 had 0.81 sensitivity at 7.3 FA/h, and line length had 0.44 at 8.3 FA/h. The held-out AUC was the same (0.917 and 0.915), thus the gain is only in the low-FA part of the curve. Per-patient sign-flip tests were not significant (p >= 0.125, n = 8). Conclusion: the calibration data cause most B1 false alarms, and a multichannel detector keeps more sensitivity at the same FA rate. No method reached a clinical target of less than 1 FA/h. Limits: 0.96 h of pre-ictal test data from seizure clips, 2-4 training seizures, clinical onset marks, one center, line-length features only. Code, features, source ETags, tables and figures are in the package.

SEEG · seizure detection · line length · logistic regression · false alarms · leave-one-seizure-out · calibration · OpenNeuro