Shaft-neighbour smoothing did not improve top-1 SOZ detection from interictal HFA in 117 SEEG patients

In archived per-pair interictal high-frequency activity (HFA) tables from 117 stereo-EEG patients across four public cohorts, we tested whether averaging each pair's HFA with that of adjacent pairs on the same shaft improved ranking of pairs in the clinical seizure-onset zone (SOZ). The top-1 SOZ hit rate was 0.308 for raw HFA and 0.291 after smoothing (difference -0.017, 95% CI -0.060 to +0.026; Holm-adjusted p = 0.69). Mean AUC was 0.687 raw and 0.681 smoothed (adjusted p = 0.58); precision at 3 and 5 also did not change. The top pair changed in 54 of 117 patients, but the replacement was in the SOZ about as often. In 65 of 74 patients with a raw top non-SOZ pair, the archived summary reports neighbour mean HFA above the patient median; the publication body reports that 9 of 74 were below it. Median within-patient percentiles were 0.83 for non-SOZ top pairs and 0.82 for SOZ top pairs. Thus, most incorrect top pairs were in spatially extended high-HFA segments rather than isolated peaks. A random-partner control lowered AUC to 0.611, while true-neighbour smoothing retained spatially coherent HFA. These results do not identify why high-HFA segments occur outside clinical SOZ labels. Limits include clinical labels not validated by outcome, 90–120 s clips, small neighbour sets, adjacency inferred from contact names, and descriptive post-hoc analyses.

stereo-EEG · interictal high-frequency activity · seizure-onset zone · epilepsy · spatial smoothing · AUC