between60and160Hz(High-Gamma).This
range of frequencies was used as the key electro-
physiological marker of local neural population
activity ( 45 – 47 , 73 ). For analyses in which spectro-
grams were computed, HFB power was calculated
as the average of frequency rows between 60 and
160 Hz. In all other cases, HFB power was com-
puted by filtering the signal in 20-Hz bands
between 60 and 160 Hz (using zero-lag linear-
phase Hamming-windowed FIR filters) and cal-
culating the normalized, 1/f corrected, analytic
amplitude using a Hilbert transform ( 44 ). The
latter method was mainly used for detecting
visually responsive/category-selective recording
sites and estimating their response latency [see
( 44 ) for details].
HFB data were inspected for transient electrical
artifacts, defined as peaks above 5sthat appear
in the HFB time series of the common average
signal (i.e., the average LFP across all iEEG chan-
nels). Time windows of 200 ms around these
peaks were logged for exclusion in subsequent
analyses.
Spectral decomposition of SWR events in
hippocampal recording sites was done using a
Morlet-wavelet time-frequency method, im-
plemented in EEGLAB. We used a window of
1 cycle at the lowest frequency (4 Hz) and up
to 20 cycles at the highest plotted frequency
(220 Hz), with a step size of 4 ms. Ripple-
triggered spectrograms were normalized by the
geometric mean power in each frequency, com-
puted over the entire epoch length (–750 to
750 ms) and across all epochs belonging to the
same condition (i.e., rest, picture viewing, re-
calling faces, recalling places) in a given run.
Spectral decomposition of HFB activation in
cortical recording sites was done using the mul-
titaper method ( 74 ) implemented in Chronux
(http://chronux.org/)( 75 ). For analysis of fre-
quencies above 30 Hz, we used a combination
of five tapers and a 200-ms-wide time window
(advanced in 6-ms steps), resulting in frequency
resolution of 20 Hz. For analysis of frequencies
below 30 Hz, we used a combination of three
tapers and a 500-ms-wide time window (advanced
in 10-ms steps), resulting in frequency resolu-
tion of 5 Hz. Here again, the ripple-triggered
spectrograms were normalized by the geomet-
ric mean power in each frequency, computed
over the entire epoch length (–750 to 750 ms)
and across all epochs belonging to the same
conditioninagivenrun.Stimulus-triggered
spectrograms in the picture-viewing stage were
normalized relative to a baseline period of– 400
to–100 ms prestimulus.
Visually responsive electrodes
We identified visually responsive sites by com-
paring, in each bipolar electrode pair, the post-
stimulus HFB response (averaged over a time
window of 100 to 500 ms) to the prestimulus
baseline (–400 to–100 ms) using a two-tailed
Wilcoxon signed-rank test.Pvalues from all re-
cording sites (across all patients) were pooled
together to control the FDR ( 76 ). Bipoles that
showed a significant HFB response (PFDR<0.05)
were regarded as visually responsive. Visual
bipoles that were fully contained within Brodmann
areas 17/18 (V1/V2), and exhibited response la-
tency shorter than 180 ms were labeled“early
visual”. To define face-selective and place-selective
bipoles, we averaged the visual HFB responses
over a time window of 100 to 500 ms poststimu-
lus and compared faces versus places using a
Wilcoxon rank sum test. Significant bipoles (PFDR<
0.05) located beyond early visual areas (V1/V2)
were labeled either“face-selective”or“place-
selective,”correspondingly. The remaining visu-
ally responsive bipoles were grouped together
according to their anatomical/retinotopic loca-
tion ( 53 – 55 ). When assigning bipoles to a spe-
cific region, we only required that one of the
two contacts be located within that region, thus
allowing for the same bipole to be attributed to
two different regions (in cases where the bipole
was located on the border between regions).Offline ripple detection
Ripple detection was performed using a macro
electrode contact located in or adjacent to the
CA1/CA2 subfields, as identified anatomically in
each patient using FreeSurfer’shippocampal
subfields parcellation algorithm ( 77 )(theexact
anatomical location in each patient is depicted
in fig. S1). For technical reasons, ripple detec-
tion in two of the patients was performed using
a contact located in the subiculum [where SWR
events can also be clearly identified ( 12 )]. Prior
to ripple detection, a reference signal from a
nearby white-matter contact was subtracted to
eliminate common noise. LFPs were then filtered
between 70 and 180 Hz (zero-lag linear-phase
Hamming windowed FIR filter with a transition
bandwidth of 5 Hz) and instantaneous analytic
amplitude was computedusing a Hilbert trans-
form. Following the procedure of ( 78 ), extreme
values were clipped to 4 SD to minimize ripple
rate–induced biasing. The clipped signal was
then squared and smoothed (Kaiser-window
FIR low-pass filter with 40 Hz cutoff), and the
mean and SD were computed across the entire
experimental duration to define the threshold
for event detection. Events from the original
(squared but unclipped) signal that exceeded
4 SD above baseline were selected as candidate
SWR events. Event duration was expanded until
ripple power fell below 2 SD. Events shorter
than 20 ms or longer than 200 ms were ex-
cluded. Adjacent events with less than 30 ms
separation (peak-to-peak) were merged. Finally,
SWR peak was aligned to the trough (of the
nonrectified signal) closest to the peak power.
A control detection was performed on the
common average signal computed across all
iEEG channels. Hippocampal SWR events that
coincided with common average ripple-band
peaks were removed, thus avoiding erroneous
detection of transient electrical and muscular
artifacts that tend to appear simultaneously
on multiple channels ( 79 , 80 ).
Lastly, to avoid inclusion of possible patho-
logical events, we removed any SWR events that
occurred within 50 ms from inter-ictal epilepticdischarges (IEDs) ( 81 ). The latter were detected
by filtering the raw hippocampal LFP between
25 and 60 Hz (zero-lag linear-phase Hamming
windowed FIR filter), and similar to the above
procedure, rectifying, squaring, smoothing,
normalizing, and detecting events that ex-
ceeded 4 SD.
The frequency window used for ripple detec-
tion in the present study was based on previous
research in humans ( 7 , 8 , 11 , 82 ), pointing to a
typical ripple-band frequency range of 80 to
140 Hz, that might occasionally reach up to
170 Hz in individual events. Thus, to minimize
the possibility of filtering out genuine ripples,
we used a frequency range of 70 to 180 Hz
(taking into account the filter roll-off). Notably,
selecting a narrower filter (e.g., 70 to 130 Hz)
did not introduce any substantial changes to the
main results.SWR peristimulus time histogram
We used the following parameters to construct
PSTHs of SWR events across the different ex-
perimental conditions. For picture viewing re-
sponses (Fig. 2A), we used 50-ms time bins
starting from–0.5 to 2.25 s relative to picture
onset, smoothed by a 5-point triangular window.
To compare ripple rate between remembered
and forgotten items (Fig. 3), we used a bin width
of 120 ms, based on Scott’s optimization method
( 83 ), to accommodate the lower number of trials.
To construct PSTH during recall events, we used
a bin size of 200 ms, smoothed by a 5-point
triangular window.Multivariate pattern analysis (MVPA)
Multivariate HFB activation patterns were con-
structed by pooling visually responsive record-
ingsites from all subjects. For the analysis, we
first defined six regions of interest along the
ventral visual hierarchy, using the Desikan-
Killiany atlas ( 54 ), including the lateral occipital
cortex (LO), inferior temporal gyrus (ITG), lin-
gual gyrus, parahippocampal gyrus (PHG), fu-
siform gyrus, and entorhinal cortex. Visually
responsive electrodes that fell within these
anatomical regions and showed a substantial
content selectivity in their responses during
picture viewing [i.e., a difference of at least 3 SD
between preferred (top 10) and nonpreferred
(bottom 10) images] were included in the analysis
(n= 78 bipolar electrode pairs; electrodes’location
is depicted in fig. S9).
To construct HFB activation patterns asso-
ciated with the viewed images, we first com-
puted in each recording site the instantaneous
HFB power using multitaper spectrograms (as
described above). In the picture-viewing condi-
tion,spectrogramswerecomputedinatime
window of–250 to 2250 ms relative to picture
onset. In the free-recall condition, spectrograms
were time-locked to hippocampal SWR events
that occurred during the verbal report of recall
(from–500 to 500 ms relative to ripple onset).
Each SWR event was uniquely associated with
the picture the subject was describing at the time
of the event.Normanet al.,Science 365 , eaax1030 (2019) 16 August 2019 11 of 14
RESEARCH | RESEARCH ARTICLE