Single-trial theta-gamma coupling does not predict working-memory errors in 13 SEEG sessions

Question: does single-trial theta (4-8 Hz) phase to gamma (30-70 Hz) or high-gamma (70-150 Hz) amplitude coupling (PAC) on one SEEG contact identify later working-memory errors, and does it add information to band power? This is a candidate BCI read-out and closed-loop target. Data: OpenNeuro ds006136 v1.0.1, load-3 object task, MTG/MFG bipolar contacts. We used 13 sessions (12 patients, 138 contacts, median 119 trials, 37% errors). sub-MU3 had fewer than 5 errors per contact, so we did not use it. Method (numpy only): FFT band-pass filter, analytic signal, normalized mean vector length for each trial in the encoding (1.0 s) and delay (3.5 s) windows. We also calculated PAC after linear removal of theta and amplitude-band log power. Per-contact rank AUC (error vs correct), mean over contacts per session, exact sign-flip test vs 0.5, Holm correction over 8 PAC tests. Result: no PAC feature separated errors. Session-mean AUC was 0.503-0.511 (raw) and 0.503-0.506 (power removed). The smallest Holm p was 0.38. All upper 95% CI bounds were 0.525 or less. Band power did separate errors. Encoding gamma power was higher in errors (AUC 0.538, 12/13 sessions, p = 0.0005). Delay high-gamma power was higher in errors (0.530, p = 0.012). Encoding theta power was lower in errors (0.463, p = 0.013). Power was better than PAC in paired tests (encoding gamma +0.027, p = 0.0085; delay high gamma +0.022, p = 0.023). Patient-level results were the same. Limits: offline; the data has only task-modulated, cleaned contacts; one PAC metric with no surrogate normalization; fixed bands; same-contact PAC only. The package has code, contact and session AUC tables, a figure, input hashes and the download recipe.

working memory · SEEG · phase-amplitude coupling · theta-gamma coupling · band power · memory errors · AUC · brain-computer interface