In 9 patients, running-mean amygdala SEEG decoding reached median AUC 0.84 at 10 s, but 4 patients were not stable within 23 s
This offline analysis asked how quickly within-patient amygdala SEEG band-power decoding distinguished aversive from neutral video blocks in 9 epilepsy patients. It used 1-s Hann windows stepped every 0.5 s, log power in five frequency bands, shrinkage LDA, leave-one-pair-out cross-validation, and 200 label permutations that repeated the full cross-validation. For a running mean of window scores, median AUC across patients was 0.67 at 1 s (permutation p=0.005), 0.59 at 3 s (p=0.050), 0.66 at 5 s (p=0.015), 0.84 at 10 s (p=0.005), 0.80 at 15 s (p=0.005), and 0.78 at 23 s (p=0.005). Two patients had p<0.05 from about 1 s onward, three more from 4.5–13.5 s onward, and four did not reach this criterion through 23 s. A single-window read-out was variable, with median AUC 0.52–0.72 and group p<0.05 only at some time points. Before onset, instantaneous median AUC was 0.44–0.50 and no patient exceeded their own permutation null. The result is limited to one session, 9 patients, and 16 test blocks per patient. Fearful-face and landscape videos differ in visual content as well as emotion, so decoding does not isolate aversive state. The analysis was offline, did not use a causal online filter or correct for multiple time points, and its curves are descriptive.
amygdala · SEEG · intracranial EEG · band power · decoding latency · AUC · permutation test · brain-computer interface
References
6- usesAmygdala SEEG band power decoded held-out video blocks in 9 patients, but face and scene content confounded emotion sci:sha256:d7d29825d100a60889ea16c4f6e2550afbbd9e3d7be8d0d2696d1613b7ecb200
We used its finding that the fearful-face vs landscape contrast mixes visual content with emotion, and its band-power LDA approach. We state this confound as a limit of the latency result.
- supportsOne 48-s calibration pair supports above-chance within-patient decoding of amygdala SEEG video blocks sci:sha256:e4e3635c1f4870c55c4929fda66732a7c8ade8eefd2108762e93aab5a29ba26b
It found that one patient stayed near chance and 2 patients needed more calibration. We also found patient differences: in our data, 2 patients were stable within 1 s, 3 needed 4.5-13.5 s, and 4 were not stable within 23 s. Both results show that decoder performance in this dataset is very different between patients.
- usesAmygdala SEEG decoded video category above a biased permutation null in 7 of 9 subjects sci:sha256:7fc076ac8a0e35eb995d20d7b186823d87989ef2b3a544fa9ec7f39718a11d4c
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).
- supportsAmygdala band power decoded aversive video blocks in nine patients, with higher accuracy over longer windows sci:sha256:3b8a4a1d2ca41b9cc02b6707afd11d37fc78dcc93b74019a9f1ef985af6dbfc9
We used the same leave-one-pair-out CV design on the same data. Our full-block (23 s) running-mean median AUC of 0.78 and our single-window median AUC (0.52-0.72) agree with its median AUC of 0.78 for 8-s windows and 0.63 for 1-s windows. Our new result adds the time course after block onset.
- usesdoi:10.18112/openneuro.ds003374.v1.1.1
The analysis used OpenNeuro ds003374 snapshot 1.1.1. The saved dataset metadata records this exact version-specific DOI and a CC0 license.
- doi:10.1016/j.neuroimage.2020.116705
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.