In 256 RAM SEEG patients, known covariates explain at most 8% of the reliable patient-level spectral component

Bipolar SEEG pairs of one patient share a large patient-level component of spectral features (AutoSci 46c98a3e). We tested which part of this component known covariates explain. We used the archived pair table of 46c98a3e (256 patients, 27974 pairs, 5 features). We removed location with the mean of the 50 nearest pairs of other patients (MNI). Then we calculated the patient mean of the residuals. Random half-splits of pairs gave a reliability of 0.97-0.99. A model with site, sampling-rate class, task dataset, age and sex gave a 10-fold cross-validated R2 from -0.08 to +0.08. The explained fraction of the reliable component was below 0.5 for all 5 features (upper 95% bound 0.15 or less). The model was better than permuted labels for broadband power, the aperiodic exponent and relative beta (Holm p 0.010-0.036), mostly from site and sampling rate. Age and sex added almost nothing. Thus more than 85% of the patient-level component has other causes, for example state, medication, disease or implant details. Studies must model the patient as a unit. The plan was hashed before the analysis. Limits: one 60 s segment, the random-split reliability includes shaft and session effects, and MNI location error remains.

stereo-EEG · SEEG · spectral features · patient-level effects · reliability · cross-validation · free recall