Across 60 patients at four centres, a three-feature sleep SEEG model transferred but did not outperform HFA alone

We tested whether a three-feature model of clinical seizure-onset-zone (SOZ) pairs transfers to recording centres that were not used for training. We reused archived feature tables for 60 eligible patients from four centres (2010 bipolar pairs; 340 clinical SOZ pairs). Features were patient-wise z-scores of log 80–300 Hz power (high-frequency activity, HFA), log 0.5–4 Hz power (delta), and log Tort modulation index of 0.5–2 Hz phase to 80–300 Hz amplitude. A weighted logistic model trained on three centres had a mean per-patient AUC of 0.734 in the held-out centre (95% CI 0.680–0.785); 52/60 patients had AUC above 0.5 (Holm p = 3e-5). It was not significantly better than HFA alone (+0.026 AUC, 95% CI −0.012 to 0.063, Holm p = 0.19). Its difference from same-centre leave-one-patient-out training was +0.022 (95% CI −0.004 to 0.050), with no transfer loss larger than 0.05 AUC. An untrained equal-weight sum of the three z-scores gave AUC 0.739, similar to the trained model. Detroit was the weakest centre (mean AUC 0.644, p = 0.057). SOZ labels were clinical rather than surgical outcome labels; each patient contributed 90 s of NREM sleep, centre sizes were 13–18 patients, and patient-wise z-scoring used all pairs from the test patient. This is a re-analysis of archived features and includes no new signal extraction.

stereo-EEG · seizure-onset zone · epilepsy · interictal sleep · high-frequency activity · delta power · modulation index · cross-centre transfer · leave-one-centre-out · AUC