Split sign selection reduced mean beta sign agreement by 1.91 percentage points at 10% quota in paired 60 second records from 36 SEEG patients

We tested a sign rule in public SEEG data from 36 patients with epilepsy. Each patient had two 60 second records. The baseline selected the strongest absolute beta imaginary coherency scores. The split rule used the smaller absolute score from two training halves, if their signs matched. Both rules had a 10% link quota. Both predicted the full training record sign. Mean agreement between records was 65.15% for the baseline and 63.24% for the split rule. The mean difference was -1.913 percentage points. Its 95% patient bootstrap interval was [-2.957, -1.158]. A two-sided patient sign randomization test gave p = 0.00001, the minimum value for the fixed draws. The interval supports a negative mean gain in this fixed comparison. The within-shaft-pair check included zero. All seven planned checks are reported. The rule changes both sign and magnitude selection; this test does not isolate their effects. Inference assumes independent patients and null sign symmetry. File separation and recording state are unknown. The sign does not establish causal flow. This is a new analysis of a cohort used in prior work. The package preserves compact data, code, source versions, range hashes, all results, and repeat instructions.

imaginary coherency · phase sign · SEEG · intracranial EEG · link selection · sign agreement · epilepsy · signal processing