Four-pair seizure detectors showed similar detection for seizure-1 onset and clinical SOZ pairs

We compared two rules for selecting four bipolar pairs for an RNS-like detector in 20 patients with stereo-EEG (SEEG) from one center and 57 later seizures in OpenNeuro ds004100 v1.1.3. Runs were treated as seizures in recorded order. The seizure-1 rule (S1) selected the four pairs with the highest 20–100 Hz z-score in the first 10 s of the first seizure; SOZ4 selected four clinical seizure-onset-zone pairs. Each pair was z-scored against its own [-40, -25) s baseline. The detector used the maximum z-score across pairs and alarmed after two consecutive windows at or above 5 SD. The primary endpoint was the fraction of a patient's later seizures detected within 10 s (D10). S1 and SOZ4 had similar HF D10 (0.64 vs 0.66; difference +0.01, 95% CI -0.23 to +0.24; p = 0.66; 19 patients). Line-length results were also similar (0.71 vs 0.78; p = 0.94), as were pre-onset false-alarm fractions (0.25 vs 0.26). Both rules detected seizures about 3 s earlier than random four-pair sets (S1 vs random: HF p = 0.008; line length p = 0.019). The top four pairs from each test seizure's own onset map, an oracle and not a usable selection rule, reached D10 0.93 (p = 0.004 vs S1). This pattern indicates that onset pairs varied across seizures and limited fixed four-pair rules in this sample. All pairs together reached D10 0.94 but had 0.69 pre-onset false alarms. Limits include one center, clinical onset marks, run order used as seizure order, per-seizure baseline normalization, only 25 s of pre-onset data for false-alarm assessment, few seizures per patient, and wide confidence intervals. The primary comparison was null and does not establish equivalence.

stereo-EEG · seizure detection · responsive neurostimulation · seizure onset zone · bipolar channels · line length · high-frequency power