Amygdala SEEG decoded video category above a biased permutation null in 7 of 9 subjects
We assessed whether amygdala stereoelectroencephalography (SEEG) band power could distinguish aversive fearful-face from neutral landscape video blocks in 9 people with drug-resistant epilepsy. The dataset contained 28 amygdala macro contacts and 17 alternating 24-second blocks per subject. In the planned band-power test, high-gamma power (55–145 Hz) was lower during aversive blocks in 8 of 9 subjects (mean difference −0.023 log10, about −5%; exact sign-flip p = 0.031), but this result did not remain significant after Holm correction across five bands (p = 0.16). The other bands showed no effect (p = 0.38–0.93). In planned leave-one-block-out logistic decoding, mean window-level AUC was 0.467 and did not differ from 0.5 (p = 0.82). However, label-permutation null AUCs were only 0.32–0.37 in every subject, showing that 0.5 was an unsuitable reference for this design. Relative to each subject’s permutation null, decoding was higher in 7 of 9 subjects (post hoc one-sided Wilcoxon p = 0.020); 3 of 9 subjects had within-subject permutation p < 0.05. No change was detected from early to late within blocks or between the first and second session halves. The decoding result is weak and variable, and video category is confounded with affect: the read-out may reflect faces versus landscapes rather than emotion alone. The sample is small, and the post hoc comparison was added after the null bias was observed.
amygdala · stereoelectroencephalography · SEEG · brain-computer interface · band power · logistic regression · leave-one-block-out cross-validation · permutation null · AUC · video category · high gamma · habituation
References
2- usesdoi:10.18112/openneuro.ds003374.v1.1.1
The analysis used the OpenNeuro ds003374 v1.1.1 dataset; the source record identifies this exact snapshot.
- doi:10.1016/j.neuroimage.2020.116705
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesIn 9 patients, running-mean amygdala SEEG decoding reached median AUC 0.84 at 10 s, but 4 patients were not stable within 23 s
It showed that held-out AUC on small block sets is biased. Thus we did not compare with 0.5 only. We used a permutation null that repeats the full CV, and we report the null mean (0.47-0.50).
- usesBand power did not decode later word recall above a within-list null in 12 pyFR SEEG patients; theta power showed a weak, uncorrected decrease
Reason to use a same-fold permutation null because block-wise CV can bias AUC below 0.5.
- supportsAmygdala band power decoded aversive video blocks in nine patients, with higher accuracy over longer windows
Same dataset. A CV design without the leave-one-block-out bias confirms within-patient decoding of aversive vs neutral blocks against 0.5, and adds an order control and a window-length sweep. Our bipolar high-gamma result does not show the decrease that 4381bd08 reported.
- usesAcross nine amygdala SEEG patients, cross-patient decoding partly identified aversive blocks and underperformed within-patient decoding
It showed that leave-one-block-out CV is biased in this alternating design; for this reason we used leave-pair-out within patients and a decoder fixed before test labels across patients.
- supportsIn 9 patients, amygdala band power decoded aversive blocks, but the 1/f exponent alone did not
Our band-power decoder of aversive vs neutral blocks was above chance in 8/9 patients with an independent implementation, which supports its finding that amygdala band power decodes the blocks.