Early high-frequency activity did not predict response time in FR1 math trials after covariate adjustment

This analysis tested whether 70–150 Hz high-frequency activity (HFA) in bipolar SEEG pairs, 0.2–1.2 s after arithmetic-problem onset, predicted response time (RT) beyond problem difficulty, position, and list drift. It used session 0 of OpenNeuro ds004789 (RAM FR1), with event time calculated as sample/fs, and a pre-registered plan written and hashed before signal data were read. Correct PROB trials with RT from 2 to 30 s were included; subjects required at least 30 valid trials. For each trial and pair, the feature was mean log Morlet power (7 cycles; 70, 80, 90, 100, 110, 130, 140, and 150 Hz) over 0.2–1.2 s. A ridge decoder predicted log RT after training-fold covariate adjustment, using five folds grouped by list. The effect was test-fold Spearman rho minus the mean rho from 200 within-list RT permutations; the group test was a two-sided sign-flip test with 20,000 flips. Of 19 attempted subjects, 18 were analysed (36–151 trials, median 77.5; 14–141 pairs). Mean corrected effect d was +0.043 (median +0.052, SD 0.113; 14/18 positive), with p = 0.127. Mean raw rho was −0.0002 (p = 0.996), and no subject had permutation p < 0.05 (smallest p = 0.060). The positive mean d therefore arose mostly from the negative bias in the cross-validated null, not a positive raw correlation. Covariates alone predicted log RT (mean cross-validated rho = +0.52, p < 0.0001). On average, 11% of pairs showed HFA above the −0.5–0 s baseline. In these 18 patients, early HFA did not predict which problems would be slow after difficulty and position effects were removed. The test detects only a within-patient pattern that is consistent and linear in log power; small effects (|rho| below about 0.1) are not excluded. Thirty-five of 54 subjects were not attempted because of the time limit. No anatomical labels were used, RT includes typing time, and the pre-onset baseline can overlap the previous answer. The analysis also removed an epoch mean, linear trend, and 60/120/180 Hz sine and cosine terms before wavelet calculation, a step absent from the plan. The package preserves the reported data hashes, code, results, logs, and plan. The exact dataset release is OpenNeuro ds004789 version 3.1.0 (DOI: 10.18112/openneuro.ds004789.v3.1.0; CC0).

stereo-EEG · high-frequency activity · response time · free recall · arithmetic distractor · cross-validation · ridge regression · OpenNeuro