Figure 1.
Illustration of the experimental procedure.
(a) The experimental protocol included one training session on day one, followed by one night of sleep, and a test session on day 2. Each session lasted about 30 min and comprised 60 S-sequence (3 blocks) and 20 C-sequences (1 block). The task was performed with the non-dominant hand and the trained 8-item sequence is shown on the bottom: finger 1 corresponds to the index and finger 4 to the little finger. (b) On each trial in the sequence, one of the four grey rectangles briefly flashed (yellow during 200 ms), indicating which button to press next. One sequence in this serial reaction time task required 8 keypresses in a specific order.
Figure 2.
(a) Representation of the implanted areas in patient C.S, lateral view of both hemispheres; (b) intracranial EEG (iEEG) recorded over the N contacts at each time-frame form an N-dimensional vector; (c) iEEG measures are pooled together within an N-dimensional space irrespective of their temporal order and a GMM is used to estimate Q Gaussian parameters (here Q = 3); (d) a Hidden Markov Model with Q hidden states, initialized by the GMM parameters, is used to update the Gaussian estimations as well as the transition matrix between the hidden states.
Figure 3.
ROC curves for M.R. (a) and C.S. (b) representing the classification results when taking into account all contacts, comparing the iEEG data from S-sequence (plain red line) and those from the C-sequence (plain blue line). The same model was used to classify reaction times from the S-sequence (dotted red line) and from the C-sequence (dotted blue line). Accuracy for classifying single-trials belonging to S-sequence was always significantly higher than that for the C-sequence and for the reaction times, suggesting that the classification algorithm captured information in the signal that was specific for the trained visuomotor sequence.
Figure 4.
ROC areas for M.R. (a) and C.S. (c), comparing classification performance when taking into account all implanted areas versus after dropping in turn one area. Red bars refer to the S-sequence, blue bars to the C-sequence (b and d). Significant differences between keeping all the contacts and dropping in turn couples of electrodes were observed for the S-sequence only for the string including the hippocampus and right frontal caudate contact for C.S. By contrast for the C-sequence, a significant drop in the classification accuracy was evident for many areas. Importantly single-trial classification remains always lower in the C-sequence than in the S-sequence. Abbreviations: LA = left amygdala; RA = right amygdala; LAH = left anterior hippocampus; LPH = left posterior hippocampus; RAH = right anterior hippocampus; RPH = right posterior hippocampus; LFO = left fronto-orbital area; RFO = right fronto-orbital area; LO = left occipital; RO = right occipital; RFC = right frontal-caudate. Standard errors are shown.
Figure 5.
Visualization of the areas most contributing to the classification.
(a) Left posterior hippocampus for patient M.R. (b) Right anterior hippocampus and (c) right frontal-caudate region for patient C.S. Electrodes localized by the CT scan were coregistered to the MRI T1-weighted brain volumes.
Figure 6.
Segmentation of single-trials belonging to the structured sequence into voltage configurations (‘momentary states’).
Each of these states is labeled in a gray scale gradation. The trials length (right-padded in black) is marked by a vertical line. For each subject (panel a and panel b), and each day (from left to right, day 1 and day 2 respectively), we show a series of exemplar single trials and the segmentation provided by the HMM algorithm. It should be noted that during day2 the average duration of these states is comparably longer that that during day 1 in both subjects. This observation was confirmed statistically in both subjects. The difference in duration length between day 1 and day 2 was highly significant and only during the structured sequence. This evidence provides new insight about spatio-temporal properties of neural activity underlying sequence-learning.