map of the single voxel in the insula as a distribution for calcu-
lating the Z-score. The Z-scores were then averaged across the
24 patients, cluster threshold k > 10.
iEEG Data
Stimuli and Task
Stimuli consisted of emotional and nonemotional sounds. The
nonemotional sounds included noise, 2 pure tones (PT), 2 har-
monic tones (Harm), Chinese speech /yi/ with 3 different pitch
contours (yi(2), yi(1), an d yi(4)), /da/, and phase-scrambled /da/.
The emotional sounds were taken from the Montreal Affec tive
Voices (
Belin et al. 2008), in which actors or actresses were
instructed to produce the vowel /a/ with different emotional
interjections. Voices from 2 identities (1 male and 1 female)
expressing 6 emotions (anger, fear, disgust, happy, sad, and
neutral) were chosen in this study. All stimuli were normalized
to the same sound level, were controlled by MATLAB (The
Mathworks, Natick, MA, USA) using the Psychophysics Toolbox
3.0 extension (
Brainard 1997), and were delivered via insert
air-conduction earphones. The volume was adjusted to a com-
fortable level of approximately 65 dB sound pressure level. All
stimuli were presented in a randomized order, and each was
repeated 20 times. Subjects needed to identify the emotion
type of the sound they heard by pressing a button after the
stimulus was over using the ha nd ipsilateral to the side of the
electrodes coverage.
iEEG Data Acquisition
iEEG signals were recorded via a 256-channel Neurofax EEG-
1200c amplifier/digitizer system (Nihon Kohden, Japan) from
implanted depth electrodes with a high-pass filter of 0.01 Hz
cut-off frequency, a notch filter at 50 Hz, and a sampling rate of
2000 Hz. Electrodes placed on the inner surface of the skull
were used as the ground and reference.
Electrode Localization
We only focused on electrodes that had contacts in the insula,
HG, and the amygdala. The locations of the electrodes relative
to the cortical surface were determined using Freesurfer. We
reconstructed the individual brain’s 3D map with electrode
locations on the surface by aligning the presurgical high-
resolution, T1-weighted MRI obtained with a Philips Achieva
3.0 T TX scanner with the postsurgical computed tomography
(CT) images obtained using the Siemens SOMATOM Sensation
64 CT scanner. This registration was visually verified and man-
ually adjusted if necessary. To show all subjects’ implanted
electrodes on an average brain surface, we coregistered the
individual MRIs to the fsaverage brain through Freesurfer soft-
ware, and all electrodes were displayed on a 3D-constructed
cortical surface representing the average brain. Furthermore,
the electrode locations were also superimp osed onto the
inflated average brain for visualization. We analyzed electrodes
from the HG, the PI, the AI, and the amygdala separately. HG,
insula, and amygdala were de fined and dissociated according
to the anatomic labels of the average brain in Freesurfer soft-
ware. The PI and the AI were defined and dissociated accordin g
to the fMRI-clustering results.
iEEG Data Analysis
All analyses were performed in MATLAB. Each channel was
visually inspected for artifacts. Channels with epileptiform
activity were excluded from further analysis. We used notch
filters from Fieldtrip software (
http://www.fieldtriptoolbox.org/)
to remove 50-Hz noise and its second and third harmonics, and
the data were down-sampled at 500 Hz. The data were then
segmented into a 200-ms prestimulus baseline and an 800-ms
poststimulus interval. All analyses focused on high-gamma
response (70–140 Hz). We used the term “following/broadband
response” in order to separate the phase-locking components
from the typical high-gamma (70–140 Hz) responses. The
Z-score of the high-gamma band was estimated using the fol-
lowing steps: 1) a 100-ms moving window (20-ms step) was
used to perform short-time Fourier transform for the prepro-
cessed iEEG signal, 2) each frequency component time series
was then normalized to its own baseline mean and was divided
by its own baseline standard deviation to determine its own Z-
score time series, 3) finally, Z-score time series of the “following
response” were averaged inside the range between F
0
and 10 Hz
and F
0
+ 10 Hz (F
0
: fundamental frequency of the sound stimu-
lus), whereas the average of the remaining frequency compo-
nents in the high-gamma band range produced the Z-score
time series of the broadband response. These normalization
procedures sought to cancel the 1/frequency decay of power in
the spectrum.
For each electrode, we defined the broadband response
amplitude of each emotion type as the peak response of the
broadband response Z -score time series (200-ms prestimulus to
800-ms poststimulus). “Response amplitude” refers to the
mean broadband response amplitude of responses to the emo-
tional stimuli recording from an electrode, which reflects the
average response amplitude of single electrode to all emotional
stimuli. To quantify the emotion-selective capacity of single
electrodes, we defined the emotion selectivity index (SI) as:
=
−
+
+
−
+
−
rr
rr
I,
i
i
i
where r
−
is the mean broadband response amplitude to non-
emotional sounds, and
+
i
is the mean broadband response
amplitude to type i emotional sounds.
SI measured the response distance between the emotional
and nonemotional sounds. For visualization, the SI values were
all mapped onto the inflated average brain using MATLAB. We
conducted statistical analyses to compare the SI between dif-
ferent sites (HG, PI, AI, and amygdala) and different hemi-
spheres by 2-sample t-test with Bonferroni corrections.
We defined a responsive trial if the Z-score of that trial
exceeded the 95% confidence interval of the prestimulus base-
line mean and was maintained for at least 100 ms. The latency
was defined as the peak latency of each responsive trial (see
Supplementary Materials).
To quantify the dissimilarity of neuronal responses elicited
by different types of emotion, we used Euclidean distance to
measure the pair-wise dissimilarities across electrodes in the
HG, the PI, the AI, and the amygdala and obtained the confu-
sion matrices. In these confusion matrices, larger dista nces
between the emotion pairs indicated higher dissimilarities of
their neural responses. We then defined the normalized
response distance as the average distance across all emotion
pairs in HG, PI, AI, and the amygdala (see Supplementary
Materials).
ECS Data
In the clinical setting, the stimulation protocol sought to iden-
tify areas that caused typical seizure sensations and to map
Emotion and Auditory Processing by Human Insula Zhang et al.
|
3
Downloaded from https://academic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhx334/4802217
by East Carolina University Health Sciences Library user
on 12 January 2018