Intracranial stimulation and EEG feature analysis reveal affective salience network specialization

Abstract

Emotion is represented in limbic and prefrontal brain areas, herein termed the affective salience network (ASN). Within the ASN, there are substantial unknowns about how valence and emotional intensity are processed—specifically, which nodes are associated with affective bias (a phenomenon in which participants interpret emotions in a manner consistent with their own mood). A recently developed feature detection approach (‘specparam’) was used to select dominant spectral features from human intracranial electrophysiological data, revealing affective specialization within specific nodes of the ASN. Spectral analysis of dominant features at the channel level suggests that dorsal anterior cingulate (dACC), anterior insula and ventral-medial prefrontal cortex (vmPFC) are sensitive to valence and intensity, while the amygdala is primarily sensitive to intensity. Akaike information criterion model comparisons corroborated the spectral analysis findings, suggesting all four nodes are more sensitive to intensity compared to valence. The data also revealed that activity in dACC and vmPFC were predictive of the extent of affective bias in the ratings of facial expressions—a proxy measure of instantaneous mood. To examine causality of the dACC in affective experience, 130 Hz continuous stimulation was applied to dACC while patients viewed and rated emotional faces. Faces were rated significantly happier during stimulation, even after accounting for differences in baseline ratings. Together the data suggest a causal role for dACC during the processing of external affective stimuli.

Metzger et al. report that activity in anterior cingulate (ACC) predicts performance on a task involving judging emotional expressions on faces, while ACC stimulation modulates performance on the same task. The data suggest that closed-loop stimulation targeting the ACC could ultimately help treat severe mood disorders.

Introduction

The ability to interpret emotional content is an important aspect of social interaction. Failures in this ability are often associated with severe consequences for social communication and are linked to neuropsychiatric disorders including depression. Following from Russell's circumplex model of emotion (i.e. Core Affect),^1,2 external affective stimuli can be examined orthogonally in terms of both valence (pleasure versus displeasure; positive versus negative) and arousal (activating versus deactivating). The resulting interpretation of a stimulus and its ecological importance or ‘salience’ are likely influenced by both. We examined the neural mechanisms underlying the interpretation of emotional facial expressions displaying happiness and sadness, which are orthogonal dimensions of valence and arousal in core affect space. Facial expressions, however, can vary in how expressive they are. That is, it is possible for a face to express subtle and overt happiness and sadness (Fig. 1A depicts faces that have been morphed from neutral to maximally expressive in steps of 10, 30, 50 and 100%). We quantify the magnitude of happiness and sadness in terms of ‘intensity’—neutral/subtle to overt, measured from 0 (neutral) to 100 (overt).

Figure 1.

Emotion rating task. (A) Faces were morphed from neutral to happy, and from neutral to sad in steps of 10, 30, 50 and 100%. (B) Faces were presented on a black background. Participants used a slider bar to rate the emotion of the face.

Distinct functional networks have been defined for processing affective³ and salient^4,5 stimuli; however, there is overlap in terms of brain anatomy and function of socio-emotional processes in these networks. Brain areas that are known to respond to such stimuli can be broadly categorized as the affective salience network (ASN). Regions of the ASN include the amygdala,^6-8 anterior insula (aINS),^5,8-10 ventral-medial and medial-orbital prefrontal cortices (vmPFC),¹¹ dorsal anterior cingulate cortices (dACC),^5,12-14 ventral striatum, thalamus and hypothalamus.^4,5 Dysfunctional activity and structural irregularities in regions of the ASN have been linked to social-emotional dysregulation in disorders such as depression,^15,16 and the ASN has been implicated in neural responses modulated by deep brain stimulation of the subcallosal cingulate (SCC) and ventral capsule/ventral striatum (VCVS), white matter fibre pathways innervating the ASN.^17-19 The field of affective neuroscience is now critically involved in building our understanding of how affective brain networks dissociably respond to emotional stimuli in terms of valence and intensity. Understanding these means of encoding and the abnormalities of perception of emotional valence and intensity in evaluating external stimuli in neuropsychiatric disorders such as treatment-resistant depression could help inform the network basis of mood disorders and help produce better-targeted therapies.

Reviewed by Guillory and Bujarski,²⁰ much is known about how valence is processed in the ASN, particularly in the amygdala, vmPFC and the insula. However, less is known about emotional intensity, or the role of the dACC for processing emotion. In a pivotal study, Caruana and colleagues²¹ reported affect-related changes in behaviour when focal stimulation was applied to the dACC. In a case report, Bijanki and colleagues²² reported increased positive affect and anxiolysis for dACC stimulation. Together, these findings suggest that dACC plays a critical role in emotion-related processing. In the current study, we apply a relatively novel procedure²³ to identify and analyse neural features of interest from human intracranial electrophysiological recordings (iEEG) in the ASN. Activity in the dACC is compared to other nodes of the ASN (amygdala, aINS and vmPFC). We use an affective bias emotional evaluation task (ABT) to examine the relationship between brain activity and behaviour as a function of two constructs: valence and intensity, plus an additional construct known as affective bias, which is the phenomenon whereby external emotional stimuli are interpreted in a manner consistent with one's own emotional state. Negative affective bias (emotional stimuli interpreted as more negative) has been associated with depression,^24-28 reliably dissociates mood groups, predicts depression treatment responses based on behavioural rating of happy and sad faces,^6,7,22,24,27 and when research paradigms permit, allows us to dissociate emotional valence and intensity.

These advances and improvements in experimental design allow us to investigate the following questions: (i) which ASN nodes are selective for valence (happy versus sad faces) or intensity (subtle versus overt emotional expressions); (ii) can we use neural activity to predict subjective ratings during the affective bias task; and (iii) how does brain stimulation to dACC change the perception of emotional faces? Answers to these questions provide further insight into affective neural processes and may identify candidate areas for closed-loop brain stimulation to treat affective mood disorders.

Materials and methods

Participants

The data reported herein come from three cohorts of participants: an iEEG cohort collected at Baylor College of Medicine (BCM), a stimulation cohort that consisted of patients from both BCM and Emory University School of Medicine, and a behavioural cohort from Amazon's Mechanical Turk (MTurk), which was used as a normative sample. The BCM iEEG and MTurk studies were approved by the Institutional Review Board at Baylor College of Medicine (IRB: H-18112, H-48155), and the Emory iEEG study was approved by the Institutional Review Board at Emory University (IRB: 00067252). All participants provided written informed consent obtained according to the principles of the Declaration of Helsinki. Participants in the BCM iEEG cohort consisted of 18 subjects with medically refractory epilepsy undergoing intracranial stereo EEG (sEEG) placement for intracranial epilepsy monitoring (Table 1). At the time of recording, the neural epileptic source was unknown. The iEEG signals were inspected by a neurologist for evidence of epileptic activity. Electrodes showing any epileptiform activity were excluded from analysis. The sEEG implantation scheme was solely based on clinical criteria, with no influence from research considerations. Two BCM iEEG study participants were excluded due to an insufficient number of trials remaining after outlier exclusion (fewer than 10 trials in any condition). Of the remaining 16 participants, six were female. The mean age of all BCM iEEG study participants was 38.25 years. Participants in the Amazon MTurk study consisted of 86 subjects (39 female, mean age of all MTurk particiapants was 38.5). The BCM stimulation cohort consisted of four patients, three of whom provided data for the iEEG study. Not all BCM iEEG study participants were able to participate in the stimulation experiment because stimulation was carried out only when patients met the following criteria: (i) they have an electrode placed in the stimulation target of interest and the target passes safety sweeps to demonstrate an absence of epileptiform activity arising following ultra-low-intensity trials; and (ii) time permits stimulation during a period in which anti-epileptic medications have been restarted prior to the explant of electrodes. The Emory stimulation cohort consisted of 10 patients. All Emory patients received bipolar stimulation. One BCM patient received bipolar stimulation (Patient B004) while the remaining three BCM patients received monopolar stimulation. Data from one of the Emory stimulation patients was excluded as a statistical outlier [>2 standard deviations (SD) from the mean].

Table 1.

Patient Beck Depression Inventory and channel inclusion summary by ROI

Patient	BDI	dACC	aINS	AMY	vmPFC
YCQ	NA	3	12	1	No coverage
YCW	21	2	9	No coverage	3
YCY	NA	1	10	1	4
YCZ	10	7	No coverage	2	No coverage
YDB	4	No coverage	8	3	No coverage
YDF	NA	7	14	No coverage	No coverage
YDG	0	1	6	No coverage	5
YDH	9	6	15	No coverage	No coverage
YDI	4	12	21	No coverage	6
YDK	18	1	No coverage	3	5
YDP	35	No coverage	2	11	No coverage
YDQ	38	6	11	6	13
YDR	21	6	9	11	No coverage
YDS	3	No coverage	6	No coverage	No coverage
YDU	6	3	5	5	28
YDV	15	18	9	No coverage	22
Total in ROI	–	73	137	43	86

iEEG analysis cohort (n = 16) patient summary. Depression scores (clinician-administered Beck Depression Inventory, BDI) and channel inclusion summary for each patient by region of interest (ROI). Channel inclusion information is represented as the total number of valid grey matter channels that were included in data analysis. ‘No coverage’ indicates patient did not have electrodes implanted in the ROI. AMY = amygdala; aINS = anterior insula; dACC = dorsal anterior cingulate cortex; NA = not assessed; vmPFC = ventral-medial prefrontal cortex.

Affective bias task

Baylor College of Medicine iEEG study

Participants rated the emotional content of static, colorized photographs of adult human faces presented to a display monitor (Viewsonic VP150, 1920 × 1080) positioned at a distance of 57 cm. Faces consisted of emotional and neutral faces adapted from the NimStim Face Stimulus Set.²⁹ Happy, sad and neutral face exemplars (six identities each; three male, three female) were morphed using a Delaunay tessellation matrix to generate subtle facial expressions ranging in emotional intensity from neutral to maximally expressive in steps of 10, 30, 50 and 100% for happy and sad faces alike (Fig. 1A). The final stimulus set consisted of 54 stimulus exemplars [six identities × nine levels of intensity (100% sad, 50% sad, 30% sad, 10% sad, neutral, 10% happy, 30% happy, 50% happy and 100% happy)]. All stimuli were presented using the Psychtoolbox extension for MATLAB.³⁰

Trials began with the presentation of a white fixation cross presented on a black background for 1000 ms (jittered ± 100 ms) followed by the simultaneous appearance of a face and the rating prompt (appearing on the left and right sides of the display, respectively; Fig. 1B). The rating prompt consisted of an active analogue slider bar placed below text instructing patients to ‘Please rate the emotion’. Participants used a computer mouse to indicate their rating by clicking a location on the slider bar. Ratings were recorded using a continuous scale ranging from 0 (‘Very Sad’) to 0.5 (‘Neutral’) to 1 (‘Very Happy’). Stimuli were presented in a blocked design in which all happy faces (plus neutral) appeared in one block while all sad faces (plus neutral) appeared in a separate block. Blocks consisted of one repetition of each image for a total of 30 randomized trials per block (six identities × five levels of intensity). A recording session consisted of three blocks each of happy and sad faces, alternating between happy and sad (order counterbalanced across participants).

Amazon Mechanical Turk study

Performance on the affective bias task was normalized by subtracting an expected score from an observed score. The expected scores came from a separate study that was conducted using Amazon's MTurk. Procedures, stimuli and experimental design were like those described above, with a few exceptions. Participants completed the task on their personal device, which could be laptop, desktop, tablet or phone in an environment of their choosing. Hence, we were unable to control certain elements of experimental design, such as the size of the display or stimuli. Emotional faces were the same as described above and were presented in random order in blocked design (i.e. neutral plus sad faces in one block, neutral plus happy faces in a separate block). Four hundred MTurk participants were initially screened for depression symptoms, and the face rating task was administered to the participants exhibiting no initial symptoms via Qualtrics. Participants had to respond ‘no’ to both of the following questions: ‘In the last month, has there been a period of time when you were feeling depressed or down most of the day or nearly every day?’ and ‘Have you lost interest of pleasure in things you usually enjoyed nearly every day?'. Averaged ‘norm’ or expected values were calculated for each face after administration of the task to the healthy controls through MTurk (the ‘expected score’). Expected scores were calculated for each image in the stimulus set, which was the average of the ratings across MTurk participants. These expected scores were then subtracted from the ratings given by participants who participated in the iEEG and stimulation studies. We term these differences between observed and expected (observed minus expected) the affective bias score.

Stimulation experiment

Stimuli were identical to those used in the main experiment. Experimental sessions began by applying brief 1-s pulses of stimulation during neurophysiological monitoring of the iEEG signal to ensure no presence of epileptic after-discharge activity. Stimulation amplitude started at 0.5 mA and increased in 0.5 mA steps to 4 mA, and after not detecting any after-discharge activity, the experiment continued into the long-form stimulation phase. This procedure also serves as a sweep for stimulation-induced paraesthesia (abnormal sensations). Biphasic symmetric rectangular-wave continuous stimulation was applied using a current-regulated device (CereStim R96, Blackrock Microsystems) at 130 Hz with a pulse width of 100 µs while patients completed the affective bias task. Stimulation data were compared to non-stimulation sham runs of the task, which were collected just prior to the stimulation runs. Non-stimulation sham runs served as a baseline comparison. Patients were blinded from stimulation condition. Patients were instructed to indicate if they felt any sensation whatsoever, and if so, the study protocol called for the experiment to be halted. No participant reported any sensation.

Intracranial EEG data

Neural signals were recorded from stereo EEG probes (Ad-Tech Medical Instrument Corporation, PMT Corporation) connected to a Cerebus data acquisition system (Blackrock Neurotech). All recording signals were amplified, filtered (high-pass 0.3 Hz first-order Butterworth, low-pass 500 Hz fourth-order Butterworth), and digitized at 2000 Hz. Trial onsets were marked using a photodiode, which was placed in the lower right-hand corner of the visual display. Simultaneous with trial onsets, a white square (hidden from the participant's view) appeared at the location of the photodiode. The analogue voltage response of the photodiode was recorded by the data acquisition system to ensure precise synchronization.

Intracranial EEG data preprocessing

Based on visual inspection, channels with excessive noise artefacts, as well as channels containing ictal and interictal activity, were excluded from data processing. Valid channels were then decimated to 500 Hz using a low-pass Chebyshev Type 1 IIR filter of order 8. Data were notch filtered to remove line noise (60 Hz, first harmonic, second harmonic) and then referenced according to a modified, single-dimension Laplacian approach, in which each electrode was referenced to the average of its adjacent neighbours along an sEEG probe. Contacts at either end of the probe were bipolar referenced to its immediate neighbour along the probe.

Neural feature selection

Selection of frequency bands for analysis has shifted away from using canonical frequency bands (i.e. theta, alpha, etc.) in favour of data-driven approaches, in which a prominent feature is identified from the frequency domain (i.e. power spectral density information, PSD), and subsequent analyses focus on differences in activity at the identified feature.^23,31-34 Accurate estimation of the central frequency and magnitude of the feature requires removal of the 1/f aperiodic signal.²³ Dominant spectral features were identified using ‘specparam’ (also know as ‘FOOOF’).²³ After referencing, channel signal data were transformed from the time domain into the frequency domain using a wavelet transformation (Morlet, seven cycles per wavelet; frequencies were equally spaced on a logarithmic scale from 1 to 200 Hz). This provides a channel × time matrix of power values for each data run. Data runs were concatenated to form a final matrix of channel × time power values. Our choice to perform wavelet transformation as opposed to Fourier-based transformation (FFT) is based on the fact that the properties of the EEG signal change over time, thus violating the statistical assumption known as ‘stationarity’.³⁵ This violation can result in distortion of spectral features present in a non-stationary signal under an FFT analysis regime.³⁶ On the other hand, wavelet methods only require stationarity in a relatively small window—the duration of the wavelet, generally less than a few hundred milliseconds.³⁷ Given that EEG data are stationary at a shorter duration,³⁷ the current study adopts the wavelet approach for calculating PSD.

At each wavelet frequency we calculated the mean power across all data runs, which provides an estimate of PSD for each channel. Once generated, the PSD is passed into ‘specparam’, which provides a list of spectral peaks (spectral features) and peak/feature magnitude. The largest magnitude spectral feature within a band of 4–14 Hz was selected from each channel as the dominant spectral feature (DSF; see Supplementary Figs 1–4 for a comparison of DSFs by run). Frequency band analysis windows were created for each channel using a 4-Hz window centred around the peak frequency of the dominant detected feature, such that the 4-Hz frequency band used for analysis was chosen for each channel independently based on each channel's detected dominant spectral feature.

Intracranial EEG data analysis

After identifying a dominant feature for each channel, data were epoched using a time window beginning at stimulus onset and ending at the behavioural response. Trial averages were calculated as mean power across all time samples within the trial window, as well as across wavelet frequencies within the dominant frequency band. Condition averages were calculated separately for each level of intensity within each level of valence.

Data were analysed using three statistical approaches. For the first approach, trial data were entered into a linear mixed effects model (LME, lme4).³⁸ The dependent variable was power in the dominant feature frequency band, fixed effects of intensity (neutral, 10%, 30%, 50%, 100%) and valence (happy and sad) and affective bias were included along with random effects of identity (i.e. actor, of which there were three males and three females). Separate models were run for each channel within each of four predefined regions of interest (ROIs): amygdala (AMY), dACC, aINS and vmPFC. This approach yielded a beta-weight for each channel and each fixed effect. We first tested for violation of normality using the Shapiro-Wilk test, in which P-values < 0.05 indicate data are not normally distributed. Two violations of normality were identified: dACC affective bias (P < 0.05) and vmPFC valence (P < 0.001). Hence, significance at the ROI level was determined by entering channel beta weights into a two-sided Wilcoxon signed-rank test. The second approach was identical to the first with the exception that all channels within an ROI were entered into separate LME for each ROI.

The goal of the third approach was to determine whether each ROI was equally sensitive to valence or intensity information. Data were averaged across trials at each level of intensity for each level of valence (total of 10 conditions per channel). The aggregated data were then entered into separate LMEs for each ROI. Three LME models were run for each ROI. The full model included fixed effects of intensity and valence, and random effects of patient and channel nested within patient. Two additional reduced models were also run, identical to the full model except that one reduced model did not include valence (‘excluding valence’) and the other reduced model did not include intensity (‘excluding intensity’). Prediction error (Akaike information criterion, AIC) was calculated for each model and then compared according to the best fitting model, which for every ROI was the full model. ΔAIC was calculated as the difference in AIC from the full model and both reduced models, as well as for the difference between the reduced models.

Our initial LME models included an interaction of valence and intensity; however, no interaction was significant (all P-values >0.05 uncorrected). Hence, interaction terms were left out to reduce model complexity. For the final models, all P-values for every ROI and every approach were corrected for multiple corrections according to the Benjamini-Yekutieli³⁹ discovery rate approach (total of 32 reported P-values). The critical P-value after correction was 0.0104.

Electrode localization and region of interest inclusion criteria

Freesurfer was used to align post-implantation CT brain scan from each patient, showing the location of the intracranial electrodes, to their pre-operative structural T₁ MRI scans. Electrode positions were manually marked using BioImage Suite 35. iELVis⁴⁰ was used to overlay electrode location into the MRI. Electrodes were then assigned to an ROI according to Freesurfer segmentation, which was then confirmed by independent expert visual inspection. Electrodes were included if they were located within 5 mm of the grey matter boundary of an ROI.

All ROIs were manually parcelled and verified by two experts. The dACC extended from a vertical boundary placed orthogonal to the anterior end of the corpus callosum extending to the end boundary between the paracentral lobule and superior frontal gyrus.^24,41 The anterior insula included gyri anterior of the central sulcus of the insula and was split into ventral and dorsal ROIs. All electrodes traversing the orbitofrontal and ventromedial cortex were assigned to the vmPFC ROI. For group-level visualizations, electrode coordinates were linearly transformed into standard space (MNI305).⁴² From a total of 2472 channels, 339 grey matter channels from four ROIs were included in the study: 43 in the amygdala (left hemisphere = 17), 73 in dACC (left = 34), 137 in aINS (left = 33) and 86 in vmPFC (left = 17). Channel locations are displayed in Fig. 2A, coloured by ROI. Figure 2B summarizes the distribution of dominant spectral features across participants by ROI.

Figure 2.

Channel location information. (A) Recording electrodes are coloured according to one of four regions of interest (ROI). (B) Distribution of dominant detected spectral features across channels for each ROI as estimated using ‘specparam’ (i.e. FOOOF). AMY = amygdala; aINS = anterior insula; dACC = dorsal anterior cingulate cortex; vmPFC = ventral-medial prefrontal cortex.

Results

Baylor College of Medicine intracranial EEG study

We collected iEEG data from patients with medication-refractory epilepsy undergoing intracranial EEG monitoring while they participated in the affective bias task in which happy and sad faces are rated (Fig. 1A and B). The feature detection approach selected a single DSF in each of the 339 channels across four regions of interest (AMY, dACC, aINS and vmPFC; Fig. 1C).

Emotional valence versus intensity

The first goal was to determine whether DSF activity varied as a function of valence (i.e. happy versus sad) or intensity (i.e. neutral to fully expressive). For each ROI, trial data from each channel were first entered into an LME (one LME per channel) that included fixed-effects of intensity, valence and affective bias score, as well as random effects of participant and channel nested within participant (model notation and summary in Table 2). This process yielded beta weights for each predictor (i.e. fixed-effects: valence, intensity and bias) for each channel. Statistical significance was determined by submitting fixed-effect beta-weights from each channel within an ROI to a two-sided Wilcoxon signed-rank test. We corrected for multiple comparisons using Benjamini-Yekutieli procedure (critical P-value after correction was 0.0104).

Table 2.

Linear mixed-effects summary by region of interest

ROI	Analysis level	Model coefficients	Fixed effect	Estimate	P-value
AMY	Trial-level	B + V + I + (1|Face identity)	Bias	0.0239	1.0000
	(channel beta-weights)	–	Valence	−0.0086	0.4407
	–	–	Intensity	−0.0006	0.0489
dACC	Trial-level	B + V + I + (1|Face identity)	Bias	0.2579	0.0021**
	(channel beta-weights)	–	Valence	0.0401	0.0023**
	–	–	Intensity	−0.0007	<0.0001****
aINS	Trial-level	B + V + I + (1|Face identity)	Bias	0.001	0.5323
	(channel beta-weights)	–	Valence	0.0144	0.0044**
	–	–	Intensity	−0.0003	0.0013**
vmPFC	Trial-level	B + V + I + (1|Face identity)	Bias	0.1662	0.0104*
	(channel beta-weights)	–	Valence	0.0794	<0.0001****
	–	–	Intensity	−0.0013	<0.0001****

Statistical significance (before correction for multiple comparisons): *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Critical P-value after FDR correction for multiple correction (Benjamini-Yekutieli) was 0.0104. Asterisks have been removed if no longer significant after correction. Estimates reflect the modelled difference between factors within each fixed effect: B = bias score; V = valence (positive values reflect greater DSF power for sad faces); I = intensity; ROI = region of interest; AMY = amygdala; aINS = anterior insula; dACC = dorsal anterior cingulate cortex; vmPFC = ventral-medial prefrontal cortex.

Data visualizations are depicted in Fig. 3A and B. Shown in red, positive beta-weights for valence indicate greater DSF activity for happy faces relative to sad faces, while negative beta-weights (shown in blue) indicate the opposite. For intensity, negative beta-weights indicate reduced DSF activity for higher levels of intensity, while positive beta-weights indicate the opposite. For example, the higher the self-reported intensity, the lower the power spectrum between 11–15 Hz in the vmPFC.

Figure 3.

Affective salience network: intensity versus valence. Spectral power predictive beta-weights by channel shown separately as a function of valence (A) and intensity (B). Shown in red, positive beta-weights for valence indicate greater dominant spectral feature (DSF) activity for happy faces relative to sad faces, while negative beta-weights (shown in blue) indicate the opposite. For intensity, negative beta-weights indicate reduced DSF activity for higher levels of intensity, while positive beta-weights indicate the opposite. n.s. = non-significant. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. For all box plots, boxes represent the 25th to the 75th percentiles (i.e. interquartile interval), whiskers represent the full extent of the data. Horizontal red lines with each box plot indicate median values. Individual dots represent individual channels, which are coloured according to beta-weight magnitude. (C) Model comparison using Akaike information criterion (AIC) reported as the difference (ΔAIC) from a full model that contains both valence and intensity. Larger bars indicate greater model prediction error. AMY = amygdala; aINS = anterior insula; dACC = dorsal anterior cingulate cortex; vmPFC = ventral-medial prefrontal cortex.

Intensity was predictive of DSF activity in dACC (P < 0.0001), aINS (P = 0.0013), vmPFC (P < 0.0001) and marginally significant for the amygdala (P = 0.0489). Valence was predictive of DSF activity in dACC (P = 0.0023), aINS (P = 0.0044) and vmPFC (P < 0.0001) but not in the amygdala (P = 0.4407). Several studies have revealed a negative correlation between cortical activation and EEG spectral power in lower frequency bands (i.e. 4–8 Hz ‘theta’ and 8–14 Hz ‘alpha’) such that a decrease in EEG spectral power in theta and alpha frequency bands is associated with an increase in cortical activation.^43-45 Consistent with these findings, our data, which show a decrease in iEEG spectral power in theta and alpha frequency bands as emotion intensity increases from neutral to overt facial emotional expressions, suggest that neural populations in AMY, dACC, aINS and vmPFC become more active as facial emotional expressions increase from neutral to overt.

To corroborate these findings, two additional analyses were run. One analysis included trial-level data from all channels, which were entered into a separate LME for each ROI (trial-level LME). The other analysis averaged data across trials before being entered into separate LMEs for each ROI (channel-level LME). Fixed and random effects were included, as described above. Critical P-value was 0.0104 (same for all DSF analyses as P-values from all DSF stats/modelling were entered into the same correction). A summary of all statistical analyses can be found in Supplementary Table 1. Consistent with our primary analysis, the supplementary analyses suggest that activity in dACC is predictive of valence (trial level P = 0.0003, channel level P < 0.0001) and intensity (trial level P < 0.0001, channel level P < 0.0001). Activity in aINS was marginally predictive of valence (trial level P = 0.0426, channel level P = 0.0366) and highly predictive of intensity (trial level P < 0.0001, channel level P = 0.0002). Activity in AMY was not predictive of valence (trial level P = 0.0684, channel level P = 0.0667) but was predictive of intensity (trial level P < 0.0001, channel level P < 0.0001). Activity in vmPFC was predictive of valence (trial level P < 0.0001, channel level P < 0.0001) and intensity (trial level P < 0.0001, channel level P < 0.0001).

Our next goal was to determine whether DSF activity was equally sensitive to valence and intensity information separately for each region of interest in the ASN. Trial data were averaged by condition for each channel (two levels of valence × five levels of intensity). Three models were compared for each ROI: a full model that included both intensity and valence, a reduced model that only included valence, and a second reduced model that only included intensity. AIC, as a measure of model prediction error, was calculated for each model and reported as the change from the full model (i.e. ΔAIC for full model is always 0; Fig. 3C). We first compared the reduced models to the full model to determine which variables were important for predicting DSF activity. Next, we compared the ΔAIC between the reduced models to assess the relative importance of each variable. Following from Symonds and Moussalli⁴⁶ (see also Burnham and Anderson⁴⁷), ΔAIC values >10 indicate sufficiently worse fit for the reduced versus the full model—the excluded variable explained a significant percentage of variation in DSF activity.

For dACC, both variables were predictive of DSF activity. Removing intensity from the model resulted in a ΔAIC of 29.01, while removing valence from the model yielded a ΔAIC of 16.19. The ΔAIC between the two reduced models was 12.82. These comparisons suggest that dACC is most predictive of DSF activity when valence and intensity are included as predictors, however, the data also suggest that dACC is sufficiently more sensitive to intensity. A similar pattern was observed in the vmPFC, such that both variables were predictive of DSF activity (no intensity model ΔAIC = 87.88; no valence model = 64.71). Though the models perform best when both variables are included, the ΔAIC between the two reduced models was 23.17, suggesting that vmPFC is more sensitive to intensity.

A different pattern was observed in aINS and AMY. For aINS, AIC model comparison suggested that aINS was primarily sensitive to intensity (no intensity model ΔAIC = 11.63), but not to valence. In fact, AIC comparison suggests that removing valence from the model makes little difference (no valence model ΔAIC = 2.35). Furthermore, the ΔAIC between the two reduced models was 9.28, suggesting aINS is sensitive to intensity but not valence. This pattern was also observed in the AMY, where intensity was predictive of DSF activity (no intensity ΔAIC = 8.92), but not valence (no valence ΔAIC = 1.29).

Taken together, LME modelling and AIC model comparison suggest the following: (i) dACC and vmPFC are sensitive to changes in intensity and valence; (ii) AMY and aINS are mainly sensitive to intensity; and (iii) the entire ASN is more sensitive to emotional intensity compared to emotional valence.

Predicting affective bias

To better understand the relationship between affective bias and ASN iEEG activity, we transformed the emotional ratings from the behavioural task into affective bias scores by subtracting normative expected ratings from participants’ observed ratings. Following from this transformation, negative scores indicate ratings that are sadder compared to the normative sample. Expected ratings came from a separate study consisting of participants recruited via Amazon's MTurk. The face rating task was only administered to participants answering ‘no’ to two stringent exclusion questions regarding recent depressive symptomatology (n = 86, refer to the ‘Materials and methods’ section). Averaged ‘expected’ values were calculated for each stimulus (Table 3).

Table 3.

Affective bias normative data (from the Amazon mTurk cohort)

Intensity	Negative faces		Positive faces
	Mean	SEM	Mean	SEM
0	0.458	0.004	0.474	0.004
10	0.453	0.004	0.493	0.003
30	0.409	0.005	0.562	0.004
50	0.31	0.006	0.662	0.007
100	0.131	0.007	0.832	0.008

Emotional rating averages and standard error of the mean (SEM) (MTurk cohort, n = 86) for each level of valence (negative versus positive) and for each level of intensity.

Affective bias scores were then entered into two LME models, both included trial level data (described above). The first approach calculated the predictive beta-weight between affective bias and DSF activity from an LME run for each channel. Statistical significance was calculated by submitting channel beta-weights for each ROI to a Wilcoxon signed-rank test (corrected for multiple comparisons). Channel beta-weights are plotted at their anatomical location and summarized by ROI (Fig. 4). The dACC was most predictive of affective bias (P = 0.0021), followed by vmPFC (P = 0.0104) (all other P-values > 0.05, Table 2). The second approach consisted of trial-level data from all channels that were entered into an LME, separate for each ROI. Consistent with the first approach, affective bias was predictive of DSF activity in dACC (P < 0.0001) and vmPFC (P = 0.0056).

Figure 4.

Predicting DSF spectral power from affective bias (observed − expected emotional ratings). (A) Predictive beta-weights for affective bias for each channel across all four regions of interest (ROI). (B) ROI-level comparisons of the trial-level predictive relationship between dominant spectral feature (DSF) spectral power and affective bias; n.s. = non-significant. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. For all box plots, boxes represent the 25th to the 75th percentiles (i.e. interquartile interval), whiskers represent the full extent of the data. Horizontal red lines with each box plot indicate median values. Individual dots represent individual channels, which are coloured according to beta-weight magnitude. AMY = amygdala; aINS = anterior insula; dACC = dorsal anterior cingulate cortex; vmPFC = ventral-medial prefrontal cortex.

Stimulation of anterior cingulate cortex during affective bias

The association between activity in the dACC and negative affective bias demonstrates an important role for the dACC in the evaluation of external emotional stimuli. To test the causality of this role, we conducted an additional experiment in which 13 patients (three of whom were included in the intracranial EEG analyses described above) received 130 Hz continuous stimulation to the anterior cingulum bundle during the affective bias task. Stimulation runs were blinded from patients and were sham controlled. Stimulation locations are displayed in Fig. 5A. All patients received bipolar stimulation, except for Patients B001, B002 and B003.

Figure 5.

Task performance during stimulation to anterior cingulate cortex. (A) Anterior cingulate cortex (ACC) stimulation locations displayed for each subject (each dot represents an individual subject). Baylor College of Medicine (BCM) patients are labelled with a ‘B’, Emory patients are labelled with an ‘E’. Colours indicate patient ID. (B) Emotional ratings for each patient as a function of whether stimulation was off (x-axis) or on (y-axis). Dashed line represents a line of equivalence. Ratings above the dashed line indicate ratings were higher while stimulation was on.

Average ratings were higher during stimulation compared to no stimulation in 10 of 13 patients [Fig. 5B, M_diffs = 0.01 (standard error of the mean, SEM = 0.003), t(12) = 3.88, P = 0.002, Cohen's d_z = 1.076]. To account for the variation in stimulation effect size as a function of baseline ratings, we tested for a relationship between ratings during baseline (no stimulation) and ratings during stimulation using an LME model that predicted ratings during stimulation from baseline and included random effects of image (i.e. each actor provided images for six levels of intensity and two levels of valence), as well as subject. All trials were included. The goal is to interpret the intercept of the LME, which is the difference in ratings between baseline and stimulation after accounting for the baseline. The LME output suggests that baseline ratings were predictive of the ratings during stimulation (P = 10⁻¹⁶). More importantly the intercept of the model (the difference between baseline and stimulation after accounting from baseline) was significant (P = 10⁻¹⁶). While it remains possible that our findings are the result of regression to the mean, the LME analysis provides some evidence that this is not the case. Stimulation increased the ratings even after accounting for baseline variation.

Control analysis: task familiarity and habituation

Faces were presented in a blocked design: happy + neutral faces in one block, and sad + neutral faces in another block. Even though block order was counterbalanced across participants, it remains possible that the observed effects are due to habituation, or familiarity with the task. To test this possibility, the averaged behavioural ratings were entered into an LME, which included fixed effects of valence and intensity, an interaction of valence and intensity, and a random effect participant ID. The LME included data from the 16 participants included in the iEEG cohort, which included thtree runs each of happy faces and sad faces (alternating between happy and sad blocks, order counterbalanced across participants). There was no effect of run order on rating (P = 0.3203), response time (P = 0.8649), nor did response time vary as a function of valence (P = 0.1912), intensity (P = 0.8649), or the interaction of valence and intensity (P = 0.5985).

Discussion

In summary, we used a spectral feature approach to analyse iEEG data, in which spectral features of interest are objectively determined prior to analyses for each channel independently. The data suggest several brain areas are sensitive to emotional valence (dACC, aINS and vmPFC, but not AMY) and intensity (dACC, AMY, aINS and vmPFC). The data also suggest that DSF activity in dACC and vmPFC are predictive of task performance (i.e. affective bias). AIC comparisons of models that only contain fixed effects of valence to those that only contain fixed effects of intensity suggest that dACC and vmPFC are sensitive to valence and intensity, that both areas are more sensitive to intensity, and that modelling activity in both ROIs performs best when both valence and intensity are included. AIC model comparisons yielded a different pattern in AMY and aINS. For these ROIs, removing valence from the model had little impact on AIC. However, removing intensity had a large impact, which suggests that aINS and AMY are primarily sensitive to intensity. We further applied 130 Hz continuous stimulation to the anterior cingulum bundle and observed robust increases in rating during stimulation (i.e. happier during stimulation) and found that variability in stimulation effects weren't merely a byproduct of baseline or change from baseline. These data strongly suggest a causal role for dACC when processing external emotional stimuli.

Targeting positive versus negative affect

Depression is a highly variable disorder characterized by several potential biotypes, each of which is associated with dysfunction in distinct brain networks.¹⁶ For instance, anhedonia is associated with hyperactivation of vmPFC,^48,49 while rumination is associated with hyperactivation of the default mode network, which includes nodes within dACC. Our LME analysis of activity in dACC revealed strong effects for intensity and valence. We were also able to reliably predict emotional ratings and bias scores from DSF trial-level activity in dACC. Together, these data suggested that dACC is critical for processing the emotional content of faces. Previous research in the dACC suggests that this area is important for integrating different dimensions of motivationally significant stimuli to encode value or emotion.^50,51 Our findings further support this view, suggesting the dACC may play a role in integrating valence and intensity components of emotion during affective and salience processing. Deficits in affective processing seen in mood disorders such as depression may arise, then, from the inability to integrate multiple components of affective stimuli or update affective perception effectively due to hypofunctional connectivity or hypo-activation of the dACC.^16,52 Direct electrical stimulation to this region could potentially restore this integrative and adaptive function of the dACC. In fact, our data showed that stimulation to ACC during the affective bias task was able to shift the interpretation of emotional facial expressions in a positive (i.e. happier) direction. Thus, dACC may serve as an ideal (though not the only) location from which to monitor brain activity to identify states (i.e. moments) of dysfunction, and apply stimulation in a closed-loop fashion.

Affective subnetworks

Disentangling the region-specific functions of the ASN is an essential step towards both characterizing emotional processing in the brain, and effectively treating emotion-related pathologies. Data from scalp EEG^53,54 and functional MRI^55,56 tend to support the valence hypothesis, which proposes that emotion is processed in a valence-specific manner such that the left hemisphere processes positive valence while the right hemisphere processes negative valence. However, the current data suggest that some ASN nodes are sensitive to both valence and intensity (dACC and vmPFC) while others are primarily sensitive to intensity but not valence (AMY and vmPFC), perhaps without preference by hemisphere.

An interesting point of mechanistic integration of the current findings is with the anatomical distribution of von Economo neurons (VENs)—a population of neurons with a unique bipolar morphology and a spindle-shaped soma, expressed almost exclusively in the human anterior cingulate cortex (concentrated in subgenual portions of the ACC, but still found in dACC) and aINS.⁵⁷ These cells are relatively large and are thought to facilitate rapid communication across distant brain regions⁵⁷ in addition to their putative role in integrating disparate information streams to mediate socio-affective functions. Abnormalities in VENs tend to translate to regional dysfunction in the aINS and ACC and are related to deficits in social communication, such as in autism and frontotemporal dementia.⁵⁸ These findings may, with further research, help characterize a unique means of socio-affective processing in the human brain, reliant on the ASN.

Limitations of the current work and future directions

Several open questions remain regarding how to assess emotion precisely but broadly from a neuroscientific perspective. One limitation of the current study concerns the use of emotional faces. The extent to which our findings generalize remains an open question. For instance, it is unknown whether our findings generalize to other domains of emotion both in terms of category (fear, anger, disgust, etc.), expression (facial, vocal), or perception (vocal pitch, image brightness, etc.). Further study with larger cohorts of participants are needed to establish the extent to which emotional evaluation and affective biasing in behaviour can be dissociated from other cognitive constructs such as attention, executive function, or arousal. Future research should determine whether the affective bias task can be used to classify or index mood, and whether it can predict treatment outcome and track changes in mood longitudinally. Clinical questionnaires including the Hamilton Depression Scale, Montgomery-Asberg Depression Rating Scale, and the Beck Depression Inventory, are only sensitive to changes in depression over relatively long periods of time. Hence, affective bias behavioural tasks can serve as a useful adjunct to these scales, especially when the need arises to test the efficacy of brain stimulation on a short time scale. Furthermore, some individuals are incapable of completing a depression inventory, for instance locked-in patients who are incapable of providing a vocal or written response, or for patients who suffer from alexithymia and are incapable of describing their own emotional state. The current study took a data-driven approach to quantifying electrophysiological activity related to emotional evaluation, which led to a focus on low frequency oscillations. There may be associated relationships in the higher frequency bands as well, which should be thoroughly characterized in future work with greater sampling of cortical and laminar grey matter structures. In addition, the current study was limited by its analysis of only electrophysiological data collected during stimulation-off periods. This approach was necessitated by stimulation duration limitations as well as by the frank limitations of stimulation artefact rejection techniques for stereo-EEG settings to date. Future studies are underway to characterize the impact of stimulation upon DSF (centre frequency), as well as summary variable changes, such as the aperiodic component of the PSD curve. Finally, the data strongly suggest dACC as a candidate stimulation target area. However, given the data, vmPFC may serve as an additional target area. Future research should seek to better understand the impact of stimulation in vmPFC on emotion perception and other affective processes.

In summary, the findings from LME modelling and direct intracranial electrical stimulation in this study suggest a causal role for dACC during the processing of emotional content. We have demonstrated the capacity for neural activity in the dACC to predict emotional behaviour at the level of individual trials (negative bias in the rating of emotional stimuli). We have also demonstrated significant changes in affective bias during stimulation of the dACC. These are the two elements necessary for the development of closed-loop neuromodulation for mood disorders. With the demonstration of modulation of sensitive and quantifiable mood-related behaviour, and with a neural target signature to track and predict modulation, this study may represent the first step toward a new future of adaptive closed-loop interventions for mood disorders targeting the dACC.

Data availability

MATLAB and R code that was used to run data analysis and generate figures are stored in a publicly available repository (https://github.com/SwatPANLab/AffectiveBias_ephys_stim). Analysed datasets are stored in the NIMH data archive. Preprocessing scripts will be provided upon request to the first author.

Funding

The project received funding support from the National Institutes of Health (Bijanki: R01-MH127006, K01-MH116364, R21-A1-NS104953; Willie: R01-MH120194, P41-EB01878), the Caroline Wiess Law Fund for Research in Molecular Medicine, the ARCO Foundation, and the McNair Foundation.

Competing interests

Unrelated to the current work, S.A.S. serves as a consultant for Boston Scientific, Zimmer Biomet, Neuropace, Abbott, and Koh Young. Unrelated to the current work, K.R.B. and J.T.W. are inventors on an issued United States patent (16121599) related to cingulum bundle stimulation. The patent applicant is individual Jon T. Willie; Inventors are Kelly R. Bijanki, Jon T. Willie, Nigel P. Pedersen, and Cory S. Inman. The patent is non-revenue-generating.

Supplementary material

Supplementary material is available at Brain online.