Adding six spectral features increased mean seizure onset label recall within a 10 percent contact budget in 31 SEEG patients

We tested small contact lists in public SEEG data from OpenNeuro ds004100, version 1.1.3. The test used 31 patients and 2,879 eligible contacts. Each test patient was absent from model training. The models trained on one recording and scored the other recording. At a budget of at most 10 percent of contacts, spectra plus contact number captured a mean 32.5 percent of seizure onset labels per patient. Contact number alone captured 20.9 percent. The mean increase was 11.5 percentage points. A 500-draw patient bootstrap refitted both models. Its approximate 95 percent interval for the increase was 3.8 to 18.5 percentage points. The spectral lists shared a mean 85.7 percent of contacts between two recordings. Four lists had no onset label in the primary test. The package includes spectra, metadata, raw retrieval records, code, all scores and lists, and independent checks. This is a new analysis of an existing hospital cohort. It does not establish surgical benefit or performance at another hospital.

stereo-EEG · seizure onset zone · spectral features · contact budget · label recall · patient-level evaluation · logistic regression