306 channel whole head meg system

MEG data were acquired with a 306-channel whole-head MEG system (Elekta Neuromag Oy, Helsinki, Finland), while subjects were in the supine position in a magnetically shielded room (VacuumSchmelze GmbH, Hanua, Germany). For each subject, two 5-min mainly eyes-closed resting-state recordings were made. Subjects were instructed to open and close their eyes several times on cue during the recordings for clinical purposes (i.e. to assess the reactivity of the alpha rhythm). Magnetic fields were recorded at a sample frequency of 1250 Hz, with an anti-aliasing filter of 410 Hz and a high-pass filter of 0.1 Hz. The subjects’ head position in relation to the MEG sensors was recorded using signals from four or five head localization coils.

The continuous neuromagnetic signal was recorded using a 306-channel, whole-head MEG system (Elekta-Neuromag TRIUX, Helsinki, Finland) at Peking University. Before each block started, the head position of each subject was determined by four head-position indicator (HPI) coils. The electrooculogram (EOG) signal was simultaneously captured by two electrodes placed near the eyes as follows: one electrode was placed above the right eye, and one electrode was placed below the left eye. The continuous MEG data were on-line bandpass filtered at 0.1–330 Hz. The sampling rate was 1000 Hz.
The subjects’ structural MRI images were obtained with a 3T GE Discovery MR750 MRI scanner (GE Healthcare, Milwaukee, WI, USA). A three-dimensional (3D) fast gradient-recalled acquisition in the steady state with a radiofrequency spoiling (FSPGR) sequence was used to obtain 1 × 1 × 1 mm³ resolution T1-weighted anatomical images. We co-registered the MEG data with the MRI data based on the location of three fiducial marks (the nasion and two pre-auricular points) and approximately 150 digitalization points on each subject’s scalp.

Magnetoencephalography (MEG) measurements were recorded using a 306-channel, whole-head MEG system (ElektaNeuromag Oy, Helsinki, Finland) in a magnetically shielded room (Vacuumschmelze GmbH, Hanau, Germany). Participants were instructed to lie on a bed with their eyes closed but to stay awake and reduce eye movements in order to minimize artifacts. Participants were scanned for 5 min with eyes closed, 2 min with eyes open, and another 5 min with eyes closed. On MEG we used source-reconstructed time series (10.1016/j.neuroimage.2011.11.005) to extract both frequency spectrum properties (relative band power and peak frequency) and functional connectivity between regions, as well as network topology using modern network theory (synchronization likelihood, modularity, path length, phase lag index) [79 (link), 80 (link)]. These analysis techniques were applied using BrainWave software (http://home.kpn.nl/stam7883/brainwave.html) [81 (link)] and inhouse MATLAB scripts (MATLAB Release 2012a; The MathWorks, Inc., Natick, MA, USA).

We looked at the scene viewing data^{33 –35} . The dataset includes both MEG and EEG recordings of 15 participants viewing 362 scene images (trials), and the MEG and EEG sessions were recorded separately on different days. MEG data were recorded using the 306-channel whole-head MEG system (Elekta Neuromag, Helsinki, Finland) at the Brain Mapping Center at the University of Pittsburgh. EEG data were recorded using a 128-channel whole-head system (ActiveTwo, Biosemi, Amsterdam, Netherlands) at the EEG laboratory of the Psychology Department at Carnegie Mellon University. Each image was shown for 3−6 repetitions and the signal was averaged across these repetitions for each image. MEG data were recorded at 1000 Hz and were downsampled to 110 Hz. EEG data were recorded at 512 Hz and was downsampled to 110 Hz. Each trial corresponds to the 1 second of the stimulus presentation. The shape of one trial is 102 channels by 110 time points for MEG (averaged across three sensors for the 102 sensor locations) and 128 channels by 110 for EEG. We used 200 trials to create features for each of the 100 identification runs. For the within-session identification (diagonal entries of Fig. 6a), we split the recording for each individual into non-overlapping source and target set before featurization.

We considered the following two datasets which have recordings on multiple days:
1- FST data^{28 (link),29} , shared online: individuals saw faces with each face appearing on the screen. There were 1464 trials for each individual. Each trial lasted 0.5 s. There were 4 individuals and 4 sessions. The sampling frequency was 1000 Hz and was downsampled to 200 Hz. Intervals between consecutive sessions were several days.
2- SEN data (shared with us by Tom Mitchell’s lab at Carnegie Mellon University): individuals read sentences. Each trial lasted 0.5 s. There were 4 individuals and 3 sessions. There were 3575 trials for each individual. The sampling frequency was 1000 Hz and was downsampled to 200 Hz. Intervals between consecutive sessions ranged from days to weeks. In this dataset, two sessions for two individuals were recorded at the same day.
The 306-channel whole-head MEG system (Elekta Neuromag, Helsinki, Finland) at the Brain Mapping Center at the University of Pittsburgh was used to obtain both recordings. The shape of one trial of the two datasets is 102 channels by 100 time points, the same as the Harry Potter data. We used 300 trials to create features for each run of identification. For the within-session identification (diagonal entries of Fig. 4b, c), we split the recording for each individual into non-overlapping source and target set before featurization.

The Harry Potter dataset was recorded using the 306-channel whole-head MEG system (Elekta Neuromag, Helsinki, Finland) at the Brain Mapping Center at the University of Pittsburgh. The study was approved by the Carnegie Mellon University and the University of Pittsburgh IRBs and informed consent was obtained from all participants. Individuals were asked to read a chapter of Harry Potter²³ while each word was presented for 0.5 s on a screen. There were 5176 trials (words) for each one of the eight individuals. There were 306 sensors at 102 locations where each location has one magnetometer and two planar gradiometers whose signal was averaged. The sampling frequency of the data was 1000 Hz which was further downsampled to 200 Hz. Details about the preprocessing of all the datasets in this paper can be found in Supplementary Information A and Supplementary Table 1. The data was parsed into trials where each trial corresponds to the MEG recording when an individual was reading a word. Specifically, the trials of individual k is ${\left\{X_i^k\in {\mathbb{R}}^{102\times 100}\right\}}_{i=1}^{I_k}$ where I_k is the number of trials for individual k, 102 represents the number of spatial channels, and 100 represents the number of temporal points in the trial. Since the recording of each individual was collected in one session, we simply split the data into a target and source dataset for the within-session identification task.