Differences in neural activity when processing emotional arousal and valence in autism spectrum disorders

Abstract

Individuals with autism spectrum disorders (ASD) often have difficulty recognizing and interpreting facial expressions of emotion, which may impair their ability to navigate and communicate successfully in their social, interpersonal environments. Characterizing specific differences between individuals with ASD and their typically developing (TD) counterparts in the neural activity subserving their experience of emotional faces may provide distinct targets for ASD interventions. Thus we used functional magnetic resonance imaging (fMRI) and a parametric experimental design to identify brain regions in which neural activity correlated with ratings of arousal and valence for a broad range of emotional faces. Participants (51 ASD, 84 TD) were group‐matched by age, sex, IQ, race, and socioeconomic status. Using task‐related change in blood‐oxygen‐level‐dependent (BOLD) fMRI signal as a measure, and covarying for age, sex, FSIQ, and ADOS scores, we detected significant differences across diagnostic groups in the neural activity subserving the dimension of arousal but not valence. BOLD‐signal in TD participants correlated inversely with ratings of arousal in regions associated primarily with attentional functions, whereas BOLD‐signal in ASD participants correlated positively with arousal ratings in regions commonly associated with impulse control and default‐mode activity. Only minor differences were detected between groups in the BOLD signal correlates of valence ratings. Our findings provide unique insight into the emotional experiences of individuals with ASD. Although behavioral responses to face‐stimuli were comparable across diagnostic groups, the corresponding neural activity for our ASD and TD groups differed dramatically. The near absence of group differences for valence correlates and the presence of strong group differences for arousal correlates suggest that individuals with ASD are not atypical in all aspects of emotion‐processing. Studying these similarities and differences may help us to understand the origins of divergent interpersonal emotional experience in persons with ASD. Hum Brain Mapp 37:443–461, 2016. © 2015 Wiley Periodicals, Inc.

Abbreviation

ACC

anterior cingulate cortex

AMYG

amygdala

BOLD

blood oxygen level dependent

Broca

Broca's area

Cb

cerebellum

CN

caudate nucleus

Cu

cuneus

DLPFC

dorsolateral prefrontal cortex

FG

fusiform gyrus

HIPP

hippocampus

ILPFC

inferolateral prefrontal cortex

INS

insula

IPC

inferior parietal cortex

IPS

intraparietal sulcus

M1

primary motor cortex

MCC

middle cingulate cortex

MFG

middle frontal gyrus

MOG

middle occipital gyrus

MTG

middle temporal gyrus

OFC

orbitofrontal cortex

PC

parietal cortex

PCC

posterior cingulate cortex

PCu

precuneus

PreMC

premotor cortex

PrG

precentral gyrus

PUT

putamen

S1

primary somatosensory cortex

S2

secondary somatosensory cortex

SFG

superior frontal gyrus

SMA

supplementary motor area

SPL

superior parietal lobule

STS

superior temporal sulcus

THAL

thalamus

V1

primary visual cortex

INTRODUCTION

Autism spectrum disorders (ASD) are a set of complex neurodevelopmental disabilities that cause lifelong impairments in social ability, communication, and behavioral flexibility [American Psychiatric Association, 2000]. Individuals with ASD often have difficulty recognizing and interpreting facial expressions of emotions, which may impair their ability to understand the intentionality and minds of others, a capacity needed for successful social communication [Golan et al., 2006; Grelotti et al., 2002]. Despite having a general consensus that persons with ASD are atypical in their processing of human faces and emotional expressions [Harms et al., 2010; Sasson, 2006], researchers do not agree on the underlying brain and behavioral mechanisms through which individuals with ASD decode emotional faces. Some prior research suggests that individuals with ASD rely more on cognitive‐perceptual systems involving explicit cognitive or verbally mediated processes to interpret facial expressions of emotions, in contrast to neurotypical individuals who process emotions more automatically [Harms et al., 2010; Pelphrey et al., 2007].

Although ASD is generally considered to involve deficits in emotion recognition, prior studies have provided only inconsistent evidence for those deficits. For example, several studies have reported that adults and children with ASD have more difficulty recognizing, responding to, and expressing emotions than typically developing (TD) individuals [Ashwin et al., 2006; Tantam et al., 1989; Uljarevic and Hamilton, 2013] and more than persons with other neurodevelopmental disorders [Celani et al., 1999; Riby et al., 2008]. However, other studies have reported typical levels of facial emotion recognition in persons with ASD [Castelli, 2005; Harms et al., 2010; Ozonoff et al., 1990; Tseng et al., 2014].

Disparities in findings for the recognition and understanding of emotions in individuals with ASD may, to some extent, be due to fundamental differences in the underlying model of emotion implicitly assumed when designing those studies. That underlying model has most often been the theory of basic emotions [Ekman, 1992; Panksepp, 1992], which posits that each member of a core set of discrete, or “basic,” emotions (e.g., anger, sadness, or happiness) are subserved by its own distinct and independent neural system [Ekman, 1992; Panksepp, 1992]. Earlier reviews have documented the limitations and inconsistencies of this theory, including the absence of one‐to‐one mappings of individual emotions with specific facial expressions, motor behaviors, and autonomic responses, as well as the absence of evidence for a core set of emotions from which other emotions derive [Hamann, 2012; Posner et al., 2005; Russell, 1980; Vytal and Hamann, 2010]. Moreover, interpreting findings from neuroimaging studies based on the theory of basic emotions is complicated by the subtraction method employed in most functional imaging designs, in which brain activity is measured by comparing two tasks or stimuli that are assumed to differ only in the cognitive process of interest. Most functional imaging studies based on the theory of basic emotions have contrasted neural responses to individual emotions with neural responses to stimuli intended to be emotion‐neutral. Unfortunately, the use of “neutral” faces as control stimuli is an inherent confound in emotion research because of the difficulties involved in creating truly “neutral” stimuli [Killgore and Yurgelun‐Todd, 2004; Klein et al., 2015; Posner et al., 2005; Thomas et al., 2001].

Additionally, most imaging studies of the basic emotions theory have focused only on a small number of emotions, generally those of high arousal and negative valence (e.g., fear and anger), low arousal and negative valence (e.g., sadness), or moderate arousal and positive valence (e.g., happy). Consequently, researchers have had trouble disentangling the differing contributions of arousal and valence to the neural correlates of emotions. For example, “happy” stimuli are often the only positive valence emotions included in study designs. Comparing them to stimuli that are putatively neutral or even negatively valenced confounds the positive arousal component with the positive valence of the happy stimulus. In effect, even when comparing happy with putatively neutral faces, both of these types of stimuli have not only a valence, but also an arousal component that is never considered as contributing to the reported fMRI activation [Fusar‐Poli et al., 2009; Harms et al., 2010; Murphy et al., 2003].

An alternative theoretical framework to the theory of basic emotions is the “Circumplex Model of Affect,” which holds that the subjective experience of all emotions arises from the linear combination of two independent neurophysiological systems, valence and arousal. Valence refers to hedonic tone, or the degree to which an emotion is pleasant or unpleasant, whereas arousal represents the degree to which an emotion is associated with high or low energy. Under this model, a “happy” response to a stimulus arises from relatively intense activation of the neural system associated with positive valence and moderate activation of the neural system associated with positive arousal. Other emotional states thus arise from the same two underlying neurophysiological systems but differ in degree of activation of each. Because all emotions can be represented as a linear combination of the dimensions of arousal and valence, emotions shade imperceptibly from one into another along the contour of the two‐dimensional circumplex [Posner et al., 2005]. The subjective experience of the neurophysiological signals for valence and arousal is determined by interpretations of the signals in relation to the experiential context of the stimuli and memories of prior experiences of similar sensations [Posner et al., 2005; Russell, 2003]. Thus the labeling of our subjective experience as one emotion rather than another nearby emotion is the consequence, in part, of cognitive interpretation of the neurophysiological experiences of arousal and valence within the situational context [Russell, 2005].

Several studies have provided evidence for the existence of distinct neural systems that subserve the experience of emotional valence and arousal [Colibazzi et al., 2010; Gerber et al., 2008; Posner et al., 2009]. However, to our knowledge, no other studies have examined whether neural activity in circuits that subserve processing of the two dimensions of facial emotions differ between individuals with ASD and their TD counterparts. A prior publication from our laboratory reported that, in the same sample, the ASD group performed nearly as well as, and in a similar pattern to, the TD group when participants were asked to rate emotional faces for arousal and valence [Tseng et al., 2014]. However, without corresponding data on brain activity, determining whether the TD and ASD groups recruited the same neural systems to appraise emotional stimuli is impossible. Typical‐level behavioral performance on emotion‐processing tasks does not exclude the possibility of atypical neurocognitive processing of emotional information. Rather, abnormalities in emotion‐processing might be obscured in some individuals because they have developed compensatory strategies that yield “typical” levels of behavioral performance. Indeed, higher‐functioning individuals with ASD might capitalize on their cognitive resources to identify facial expressions. For example, studies employing emotion‐matching paradigms [Piggot et al., 2004; Rump et al., 2009] are more likely than studies using emotion‐labeling paradigms [Katsyri et al., 2008; Piggot et al., 2004; Rutherford and Towns, 2008] to reveal differences in behavioral performance between TD and higher‐functioning ASD groups. For some individuals with ASD, the use of emotion labels in a task may facilitate recognition of facial expressions of emotions, especially when they are trained to identify emotions as part of an intervention program [Tanaka et al., 2010].

Although functional imaging studies of emotion‐processing in ASD have yielded inconsistent findings, several have reported hypofunctioning in ASD in regions associated with socio‐emotional processing (e.g., INS, AMYG) [Di Martino et al., 2009], in extrastriate cortices [Deeley et al., 2007], ventral PFC [Ashwin et al., 2007; Hadjikhani et al., 2006], medial‐frontal and orbito‐frontal cortices [Bachevalier and Loveland, 2006; Loveland et al., 2008; Ogai et al., 2003], ACC and FG [Hall et al., 2003], striatum, and IFG [Dapretto et al., 2006] compared to TD controls. Conversely, studies have found increased activity for ASD compared to TD groups in STS, ACC [Ashwin et al., 2007; Hall et al., 2003; Pelphrey et al., 2007], and parieto‐occipital regions [Dapretto et al., 2006; Hubl et al., 2003; Wang et al., 2004] when viewing facial emotions. These increases in neural activity may derive from increased visual and motor attention [Dapretto et al., 2006], more effortful processing of specific facial features within the given social contexts [Ashwin et al., 2007], and increased attentional load [Wang et al., 2004], supporting the possibility that emotional processing is more effortful and less automatic in individuals with ASD than in their TD counterparts.

To address whether neural activity in circuits that subserve processing of arousal and valence differ between individuals with ASD and TD individuals, we applied a parametric experimental design to identify brain regions in which neural activity correlated with arousal and valence ratings for a broad range of facial emotions. The use of a parametric design allows us to compare emotional stimuli across multiple levels or through incremental changes along the affective dimensions of arousal and valence. For example, parametric manipulation of emotional stimuli that change incrementally in the degree of arousal or valence that they generate can be mapped against concomitant variations in neural activity. Accordingly, activity in neural structures or pathways that correlate with the degree of emotional arousal or valence induced by the emotional probes can be assessed in individual participants [Posner et al., 2005]. In the present study, we used BOLD‐signal intensity to index neural activity as participants viewed photographs depicting emotional faces. We identified brain regions in which BOLD‐signal systematically covaried with ratings of arousal or valence in ASD and TD groups, and we determined the areas in which these correlations differed statistically across emotional dimensions and diagnostic groups, indicating the differential associations of these regions with processing arousal or valence within each group. We sought to identify similarities and differences in neural activity when participants with ASD and TD participants view and rate these experiences of facial emotions [Gerber et al., 2008; Russell et al., 1989]. Given the socio‐emotional deficits associated with ASD, we hypothesize that the ASD group will show abnormal patterns of brain activation when compared to controls, particularly in brain regions associated with processing of emotional stimuli in persons with ASD, including AMY, INS, CN, PFC, OFC, and ACC.

MATERIALS AND METHODS

Study procedures were approved by New York State Psychiatric Institute's Institutional Review Board. All participants provided informed written consent or assent and received payment for participating (See Supporting Information for detailed consent procedures).

Participants

We recruited 51 individuals with ASD (6F, ages: 7–60 years, Mean: 27.5 ± 13.1 years) and 84 TD individuals (22F, ages: 7–60 yrs, Mean: 24.0 ± 11.4 years) from the New York City area. A wide age‐range was included in our sample in order to understand better the developmental trajectory of emotional processing in this under‐studied group. For example, if the child participants with ASD performed similarly to our adult participants with ASD, then we might infer that any emotional deficits found are likely a static, trait‐like disturbance. We also hoped to use cross‐sectional data from this investigation to generate hypotheses for future longitudinal research.

Participants were group‐matched by age, sex, IQ (Wechsler Abbreviated Scale of Intelligence) [Wechsler, 1999], handedness (Edinburgh Handedness Inventory) [Oldfield, 1971], race, and socioeconomic status (Hollingshead Index of Social Status) [Hollingshead, 1975]. Mean full scale IQ (FSIQ) was 109.2 ± 19.4 for the ASD group and 115.9 ± 12.4 for the TD group. Mean verbal IQ (VIQ) and mean performance IQ (PIQ) for both groups did not differ significantly so we opted to conduct further analyses of IQ using only FSIQ as a covariate (Table 1).

Table 1.

Participant characteristics

	ASD	TD
Participants (N)	51	84
ASD subtype
PDD‐NOS	9	—
Asperger's syndrome	24	—
Autistic disorder	18	—
Mean age (yrs)	27.5 ± 13.1	24.0 ± 11.4
Children (<18 yrs) (N (%))	12 (24%)	31 (37%)
Males (N (%))	45 (88%)	62 (74%)
Caucasian (N (%))	40 (78%)	60 (71%)
Mean SESa	50	53
Mean FSIQb	109.2 ± 19.4	115.9 ± 12.4
Mean VIQ	110.9 ± 20.9	115.7 ± 13.2
Mean PIQ Mean ADOS (social affect + restrictive, repetitive behaviors)c	105.0 ± 17.6 10.9 ± 3.1	112.9 ± 11.8 —
Mean ADOS—calibrated severity scores (modules 2 and 3)d	7.5 ± 1.8	—

^aSES scores for 7 TD and 14 ASD participants were unavailable.

^bFSIQ scores for 1 TD participant and 1 ASD participant were unavailable.

^cADOS scores for 4 ASD participants were unavailable.

^dADOS CSS scores were calculated for participants tested using Modules 2 and 3 (N = 10).

Additional individuals participated (N = 4 ASD; N = 1 TD) but were not included in the final sample due to excessive head motion in the scanner.

TD participants were excluded if they met DSM‐IV‐TR criteria for a current Axis‐I‐disorder, or had lifetime history of developmental delay or other indicators of ASD, psychosis, substance abuse disorder, head trauma, seizure disorder, or other neurological illness. None of the TD participants were taking prescription or over‐the‐counter medications; however, the use of dietary supplements was not assessed.

Participants with ASD were evaluated by an expert clinician and met Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (DSM‐IV‐TR) [American Psychiatric Association, 2000] criteria for autistic disorder, Asperger syndrome, or pervasive developmental disorder‐not otherwise specified (PDD‐NOS). Diagnoses were confirmed with the Autism Diagnostic Interview‐Revised [Lord et al., 1994] and the Autism diagnostic observation schedule (ADOS) [Lord et al., 1989]. A detailed list of current medications was recorded for every participant (available on request).

At the beginning of each study session, an experienced member of the study team explained verbally the nature of the research protocol, including potential risks and benefits, to all potential participants and (in the case of a minor) to their parents. For participants 8 to <18 years old, assent was discussed and obtained. For adult participants (>18 years), the study member obtaining consent explained the protocol and associated risks to the prospective participant before asking participants to sign the consent form. For all adult participants with ASD, an independent assessment of capacity to consent was conducted by a clinical monitor who was independent of the study team. When the clinical monitor determined that the participant lacked the capacity to consent (i.e., the participant did not demonstrate an understanding of the procedures, alternatives, and potential risks and benefits of the study, and that participation was voluntary), an authorized legal representative was designated to provide informed consent.

Emotion Paradigm

We used a parametric experimental design to identify brain regions in which neural activity correlated with participant ratings of arousal or valence (Fig. 1).

Figure 1.

Emotion paradigm.

Neural activity was indexed using BOLD‐signal intensity as participants viewed photographs of emotional faces. After viewing each face, participants rated arousal and valence simultaneously by selecting an individual box on a 9 × 9 two‐dimensional grid. Location on the x‐axis indicated the participant's rating of valence (left = negative valence, right = positive valence), and location on the y‐axis indicated the rating of arousal (top = high arousal, bottom = low arousal). We recorded the selected box as two integer scores, each ranging from −4 to +4, representing valence and arousal.

Each trial consisted of 3 sequential epochs: (1) Visual presentation of a photograph of a human face for 18 s. The photographs were copied, with permission, from the 20 distinct stimuli used by Russell and Bullock [1985] for their studies of the affective circumplex. Thirteen of these 20 images were taken from Ekman and Friesen's [1976] “Pictures of Facial Affect” and depicted expressions of a number of emotions (two pictures each of emotional faces commonly classified as expressing happiness, surprise, fear, anger, disgust, or sadness, and one commonly classified as neutral). Russell and Bullock supplemented this set to represent better portions of the circumplex that the Ekman series under‐sampled (emotions associated with low arousal but positive or neutral valence). These included two photographs each of actors and actresses expressing boredom, contentment, and sleepiness, as well as one expressing excitement. (2) Visual presentation of a two‐dimensional grid on which participants indicated their ratings of arousal and valence for each stimulus by moving an arrow controlled by an MRI‐compatible computer mouse. This screen remained visible until the participant clicked the mouse button, up to a maximum of 20 s. (3) Visual presentation of a fixation point (+) at the center of the participant's visual field. The fixation point was displayed immediately following the rating of valence and arousal. The durations of rating and gaze fixation were each variable, but when summed together always equaled 20 s. One imaging run consisted of 20 trials presented in a pseudorandom order (but uniform from participant to participant).

Visual stimuli were presented to each participant via MRI‐compatible LCD goggles (Resonance Technology, Northridge, CA) using E‐Prime software, version 1.138 running on a Dell IBM‐compatible computer. Measures of stimulus durations and reaction times were accurate to 20 ms. Stimuli were presented at the center of the participant's visual field, subtending 19° of the vertical and 15° of the horizontal visual field.

Prior to the study session, all participants were given a practice session with the task so they could familiarize themselves with task instructions, the types of stimuli they would be seeing (practice stimuli were not shown during the study session), the grid on which they would be rating arousal and valence, and the computer mouse they would be clicking to indicate their ratings. Researchers were available to review the practice responses in detail, to explain the instructions further, or to answer any questions about the task during this practice round to ensure full comprehension. Each participant was told, “You will be shown a face that expresses a certain feeling. You will be asked to assess the feeling on the chart shown below….On the chart, the vertical dimension represents degree of arousal. Arousal has to do with how awake, alert, or energetic a person is…. The right half of the chart represents pleasant feelings—the farther to the right, the more pleasant. The left half represents unpleasant feelings—the farther to the left, the more unpleasant…. During the experiment, you will first be shown a face. This will appear on the screen for 18s. Then you will be shown the grid. When the grid appears, you will click on the area you think best describes the face…Try to think about the feeling expressed by the face during the 18s shown. It will not be on the screen when you are shown the grid.” At the time of instruction and during the experiment itself, the words “High Pleasure” appeared to the right of the grid, and “High Energy” above the grid. The shortened practice version consisted of three faces—one each expressing sadness, happiness, and anger. To minimize the possibility of habituation, none of the practice faces were identical to actual experimental stimuli. During the scan, researchers monitored on‐line behavioral responses in real‐time so that we could ensure attention to the task.

Behavioral Data Analyses

Behavioral data gathered from the present study were analyzed and reported in detail in a prior publication from our laboratory [Tseng et al., 2014]. The participants were identical across the two studies, with the exception that several participants were eliminated from fMRI analysis because of excessive head motion in the scanner. In addition, we collected data for an additional 3 TD adults, 1 TD child, and 4 adults with ASD after the publication of our prior study. We did not find any significant differences between our findings with or without the additional eight participants, so we included them in our fMRI sample. As described in our previous report, we divided participants into four groups by diagnosis and age to compare behavioral performances across groups: Adult ASD (N = 39, 5F, ages: 18–61 years, Mean: 31.9 ± 11.8 years), Adult TD (N = 53, 8F, ages: 18–60 years, Mean: 30.1 ± 10.2 years), Child ASD (N = 12, 1F, ages: 7–17 years, mean: 13.2 ± 3.1 years), and Child TD (N = 31,14F, ages: 7–17 years, Mean: 13.7 ± 2.7 years). Mean FSIQ scores in these groups were: Adult ASD (110.2 ± 18.4), Adult TD (116.42 ± 1.9), Child ASD (105.6 ± 22.2), and Child TD (115.0 ± 13.4). We also divided participants by diagnosis alone to compare the entire ASD and TD groups.

Multivariate ANCOVAs were conducted with arousal and valence ratings as dependent variables, group as the independent variable, and age and gender as covariates to assess emotion‐specific differences between groups. We used hierarchical multiple regressions for ASD and TD groups (controlling for age and sex) with arousal and valence ratings as dependent variables and FSIQ scores as the independent variable to assess whether IQ was significantly correlated with how participants rated each emotion‐type. Similar analyses were conducted with total ADOS scores (Social Affect (SA) + Restrictive, Repetitive Behaviors (RRB), Mean = 10.9 ± 3.1 [Gotham et al., 2007]. CSS conversion algorithms are not available for participants over the age of 16 or who were assessed with module 4 of the ADOS.

To assess whether severity of diagnosis significantly correlated with how participants rated each emotion‐type, we used hierarchical multiple regressions for analyses in the ASD group (controlling for age and sex) in which arousal or valence ratings were entered separately as the dependent variable and total ADOS score was the independent variable. These regressions were applied separately to each facial stimulus. We also conducted these analyses with only the social affect scores from the ADOS as the independent variable, because we expected the social affect measure alone might correlate more strongly with how participants with ASD rated these affective stimuli.

Finally, we conducted multivariate ANCOVAs with arousal or valence ratings entered separately as the dependent variable, ASD subtype (PDD‐NOS, Asperger's Syndrome, Autistic Disorder) entered as the independent variable, and age and gender entered as covariates to assess whether participant responses varied according to specific ASD subtype.

Task Performance

So that we could be as confident as possible that all participants were performing the task as instructed and to ensure the face validity of their responses, we first visually compared each individual's arousal and valence ratings qualitatively against the canonical circumplex to ensure that the responses seemed reasonable. Then, assuming that the responses of the healthy adults represent the end product of development, we used the arousal and valence scores from typically‐developing adults reported by Russell and Bullock [1985] as reference ratings for “correct” performance by assessing quantitatively the correlations of each individual participant's data with the reference ratings. Our rationale was that an individual responding at random to the stimuli or who was not understanding or following instructions would be unlikely to produce a similar response pattern to the reference ratings. Then, as a subset analysis, we removed participants (N = 13: 4 Child ASD, 4 Adult ASD, 5 Child TD) whose correlations between arousal or valence ratings with the reference values were significant at P > 0.2 (corresponding to a Pearson's r > 0.42). Similar to findings from our original analysis with the entire sample (N = 135), we detected with this smaller sample (N = 122) a main effect of diagnosis (P < 0.05). Thus, although we were unable to measure task comprehension directly during the scan, the use of prescan practice trials and the similarity of results in our subset analysis with those of the original analysis show that the vast majority of our participants were able to understand and perform the task as instructed.

Image Acquisition

Imaging was performed on a GE Signa 3T whole body scanner (Milwaukee, WI) using a GE single channel quadrature head‐coil. A 3D spoiled gradient recall (SPGR) image was acquired for coregistration with axial functional images and for coregistration with a standard reference image (Montreal Neurological Institute (MNI)). Functional images were acquired using a single shot gradient echo planar (EPI) pulse sequence in groups of 43 axial slices per volume and 273 volumes per run (preceded by six “dummy” volumes to ensure scanner stability). Parameters for the EPI images were: repetition time = 2,800 ms, echo time = 25 ms, flip angle = 90°, acquisition matrix = 64 × 64, field of view = 24 cm × 24 cm, slice thickness = 3 mm, skip = 0.5 mm, receiver bandwidth = 62.5 kHz, in‐plane resolution = 3.75 mm × 3.75 mm. Each run lasted 13 min 1 s, for a total EPI scan time of 39 min 3s.

Image Preprocessing

Prior to statistical analyses, we used SPM8 (http://www.fil.ion.ucl.ac.uk/spm/, run under MATLAB2009b) to preprocess the fMRI data. Slice timing was corrected using the middle slice (22 of 43) as the timing reference. Slice timing corrected functional images were then realigned to the middle image of the middle run for motion correction for three translational directions and rotations. Images with motion greater than one voxel were excluded from all subsequent analyses. Motion corrected images of each participant were coregistered to the corresponding T1‐weighted high‐resolution anatomical image, which in turn was spatially normalized to the standard MNI template with voxel dimensions of 3 mm³. These participant‐specific normalization parameters were then used to warp the functional images into the same MNI template. A spatial smoother with a Gaussian kernel of 8‐mm Full Width at Half Maximum was applied to the functional images, which were then temporally filtered using a Discrete Cosine Transform high‐pass filter with a cutoff frequency of 1/128 Hz to remove low frequency noise such as scanner drift.

We then assessed data quality by plotting motion parameters and mean intensity values for raw, normalized, and smoothed images for each run in each participant. Visual inspection allowed us to identify scans for which average intensity values across voxels were significantly outside the mean and which occurred at the same moment as a large head movement. We also used histogram plots for each contrast image in each participant to help identify outliers for mean intensity that might have been missed by the batch preprocessing procedure. We used the ArtRepair algorithm (http://cibsr.stanford.edu/tools/human-brain-project/artrepair-software.html) to detect and repair those image volumes that were contaminated by spiking motion artifacts and outliers [Mazaika et al., 2009]. Volumes with motion larger than 1mm were repaired. Participants for whom motion affected more than 15% of their data (>41/273 volumes per run) were excluded from further analyses; based on this criterion, we eliminated from our final analysis 1 TD and 4 ASD participants (from the original 140 participants).

Statistical Analyses

We analyzed fMRI data at the individual (first) level using a general linear model (GLM) to detect BOLD‐signal correlates of arousal or valence within each individual participant and at the group (second) level using Bayesian posterior inference [Neumann and Lohmann, 2003] at a posterior probability threshold of 98.75%, to detect random effects of arousal or valence correlates within and between diagnostic groups. We covaried for age and sex of the participants. We also conducted additional analyses covarying for FSIQ in all participants and for ADOS scores in the ASD group. We assessed the main effects of arousal and valence ratings on BOLD‐signal for each diagnostic group (TD, ASD). We also assessed BOLD‐signal correlates with quadratics of arousal and valence ratings, allowing us to assess at each voxel whether the correlation of ratings with BOLD‐signal had a significant curvilinear component. We included simultaneously in our model the main effects and quadratic values of arousal and valence ratings (including them separately yielded identical findings). Finally, we assessed whether the within‐group valence and arousal correlates differed significantly across ASD and TD participants by assessing the interactions of the correlations with diagnostic group. We included simultaneously in our model the main effects and their interactions with diagnostic group to ensure that the models were hierarchically well formulated. We plotted the scatters for the linear and quadratic associations of BOLD‐signal with ratings of arousal and valence to assess the distribution of data around the regressions and to determine the group contributions to significant interactions.

First‐level analysis

We used GLM in SPM8 for the analyses of data at the individual level. We modeled preprocessed BOLD time series data at each voxel, using 8 independent functions (Fn) or regressors that consisted of:

• Fn(1): the canonical hemodynamic response function (HRF) convolved with a box car function (BCF) derived from the onsets and durations of the presentation of facial stimuli

• Fn(2): Fn(1) modulated by the linear arousal rating for each stimulus

• Fn(3): Fn(1) modulated by the linear valence rating for each stimulus

• Fn(4): Fn(1) modulated by the quadratic arousal rating for each stimulus

• Fn(5): Fn(1) modulated by the quadratic valence rating for each stimulus

• Fn(6): the canonical HRF convolved with a BCF indexing the manual responses of each participant to the task stimuli

• Fn(7): the canonical HRF convolved with a BCF indexing the presentation of a fixation cross

• Fn(8): a constant

Our model, which included the main effects and quadratic values of arousal and valence ratings on a −4 to 4 scale for each participant, was estimated using the Restricted Maximum Likelihood (ReML) algorithm. Task‐related T‐contrast images were generated using SPM8 contrast manager. We ran our models for valence and arousal separately (i.e., with functions 1, 2, 4, 6, 7, and 8 for arousal and with functions 1, 3, 5, 7, and 8 for valence) and both with and without the quadratic terms (functions 4 and 5) to ensure that the model was not over‐specified; the findings for the linear arousal and valence terms in these reduced models were unchanged from findings for the model that included all eight functions. We thus elected to present findings for the full model so that we could account for every event that occurs during the task, allowing us to control for signal variability in each trial. Also, by including both linear and quadratic components we were able to assess whether the response is truly linear across the range of ratings or whether it is curvilinear [Acton and Friston, 1998; Buchel et al., 1996, 1998; Frackowiak, 2004]. We also ran our model using both the SPM default that orthogonalizes parametric variables, as well as without orthogonalization, because we were concerned that if our regressors were inter‐correlated and we did orthogonalize our modulators, then the explained variance in BOLD‐signal would not be assigned to any of the regressors and our power would be reduced for statistical testing. We also wanted to ensure that our findings were robust with respect to orthogonalization. Our findings were nearly identical with or without orthogonalization, so we elected to present our findings using the orthogonalized analyses. Finally, we also ran the GLMs with motion parameters as regressors and found that they had no significant effect on our findings, so we elected to present our findings without motion regressors in the model.

Second‐level analysis

We used Bayesian inference to detect random effects by assessing the posterior probability of detecting within or between group difference, β, given the activation map that we attained in a particular contrast. We used a posterior probability of greater than 98.75% as the threshold for statistical significance in each of the contrast maps and, in addition, required a spatial extent of at least eight contiguous voxels to further strengthen the biological validity of our findings. Unlike a more conventional second‐level analysis that uses classical parametric inference to detect a group effect in a statistical parametric map by disproving the null hypothesis (β = 0) at each voxel of the image, a group effect using the Bayesian method infers the posterior probability of detecting the observed group effects (β ≠ 0), given the data in a posterior probability map [Neumann and Lohmann, 2003]. Whereas the voxelwise tests in a statistical parametric map require correction for the number of statistical comparisons performed, the Bayesian method, because it infers posterior probability, by definition, does not generate false positives and does not require adjustment of its P values based on stringent P value thresholding (a feature of these analyses that has been confirmed in numerous simulations and empirical studies) [Friston and Penny, 2003; Friston et al., 2002].

Post‐Hoc Analyses

Several additional analyses ensured that possible confounding effects did not unduly influence our findings. We conducted post‐hoc analyses while covarying for FSIQ in all participants and ADOS scores in analyses involving only participants with ASD. Results did not differ significantly when we covaried for age, sex, FSIQ, or ADOS scores in participants with ASD. Additionally, we assessed the age‐by‐diagnosis interaction but found none; restricting our ASD sample to participants who were older than 18‐years did not change our findings from those for our overall sample (ASD: 24% of group (12/51) and TD: 37% of group (31/84)). Additionally, we analyzed our dataset with only male participants (45 ASD, 62 TD) and found the patterns of activation to be similar to those for our main model (Supporting Information Fig. S4). We also assessed age correlations within each group and detected none that were significant for valence or arousal. Finally, restricting analyses to participants who were medication‐naïve (ASD: 68% (34/51) yielded the same results as for our overall sample.

RESULTS

Behavioral Data

On the whole, our behavioral findings suggest that while participants in the ASD group rated arousal and valence for a wide range of emotions similar to individuals in the TD group, emotion ratings for the ASD groups along both valence and arousal dimensions were somewhat constricted in their ranges relative to those of the TD groups. These findings did not change when we covaried for overall intelligence. Also, for the ASD group, correlations of ADOS scores with arousal and valence yielded only one marginally significant finding: ADOS scores correlated with valence ratings for surprise faces (β = 0.32, t ₄₂ = 2.1, P = 0.05). Results did not vary by ASD subtype.

Emotion‐specific Analyses

Given that emotional processing in typically developing adults is presumably the desired outcome of emotional processing in typical and atypical development, we used our average Adult TD data as a point of reference for visually comparing data from the other three groups, even though our primary analyses of the behavioral data treated age as a continuous variable. We report subtle but significant differences between groups for specific emotions, although the overall assignment of valence and arousal scores across the all emotion‐types were similar for the ASD and TD groups (Fig. 2).

Figure 2.

Emotion‐specific group comparisons of behavioral findings.

Ratings for both valence and arousal dimensions of emotions in the child ASD group were somewhat constricted in their ranges relative to those of the Adult TD group: the child ASD group reported significantly lower arousal ratings for high arousal emotions such as Excited (F _3,63 = 3.53, P = 0.0008) and surprised (F _3,63 = 3.38, P = 0.0013), and higher arousal ratings for Sleepy, a low arousal emotion, (F _3,63 = 2.02, P = 0.048). They also reported significantly less negative valence ratings for negatively valenced emotions, including Disgusted (F _3,63 = 2.01, P = 0.049) and Sad (F _3,63 = 2.83, P = 0.006), and a trend for less positive valence ratings than the adult TD group for the positively valenced excited (F _3,63 = 1.96, P = 0.055) and happy (F _3,63 = 1.90, P = 0.061) (Fig. 2A).

Emotions for the adult ASD group relative to the adult TD group also showed a trend for constriction in their ranges; they reported significantly less negative valence ratings for the negatively valenced sad faces (F _3,90 = 2.33, P = 0.022) (Fig. 2B).

No significant age differences were detected within the ASD groups. However, adult TD participants did report higher arousal ratings than Child TD participants for the negatively valenced angry (F _3,82 = 2.64, P = 0.01), disgusted (F _3,82 = 2.46, P = 0.016), sad (F _3,82 = 2.82, P = 0.006), and scared faces (F _3,82 = 2.14, P = 0.036) and less positive valence ratings for excited faces (F _3,47 = 2.818, P = 0.045) (Fig. 2C).

fMRI Data

Main effects

Linear and quadratic correlates of arousal

These analyses revealed significant inverse linear and quadratic correlations of BOLD‐signal with arousal ratings for our TD participants in ILPFC and DLPFC, dorsal ACC, inferoposterior PC, dorsal PC, CN, and PUT. For ASD participants, we detected significant positive linear associations of BOLD‐signal with arousal ratings in the posterior temporal/inferior PC, mesial wall (pregenual and dorsal portions of SFG and ACC), premotor and supplementary motor regions, Cu and PCu, subcortical regions (all basal ganglia nuclei, THAL), and dorsal Cb (Fig. 3, Tables 2 and 4).

Figure 3.

Arousal correlates (A) Regions of significant correlations of BOLD‐signal with ratings of arousal for TD participants. (Positive correlations are coded in red to yellow, and inverse correlations are coded in green to blue.). (B) Scatterplots of correlations for BOLD‐signal with ratings of arousal for TD participants (green) or participants with ASD (purple). (C) Regions of significant correlations of BOLD‐signal with ratings of arousal for participants with ASD. (D) The regions where the correlation of arousal ratings with BOLD‐signal for participants with ASD differs significantly from the correlation of arousal ratings with BOLD‐signal for TD participants. (ASD > TD coded in red to yellow, TD >ASD coded in blue to green). (E) Scatterplots of correlations for BOLD‐signal with ratings of arousal for TD participants (green) and participants with ASD (purple) in regions where the correlation of arousal ratings with BOLD for participants with ASD differs significantly from the correlation of arousal ratings with BOLD for TD participants.

Table 2.

Regions of significant linear correlations of BOLD signal with ratings of AROUSAL for participants with ASD

Anatomical regionsa	Location		MNI coordinates			Z
	Side	BA	x	y	z		Correlation
Linear arousal
Premotor cortex	R	6, 1	39	−28	67	8.21	Positive
Dorsolateral prefrontal cortex	L	8	−6	47	25	8.13	Positive
Cerebellum	L		−6	−61	−20	8.04	Positive
Dorsolateral prefrontal cortex	L	9	−21	20	58	7.97	Positive
Dorsolateral prefrontal cortex	R		21	41	49	7.66	Positive
Broca's area	L		−51	20	13	4.23	Positive

^aRegions were significant at P < 0.001 or posterior probability > 99.0%. (All of the other tables in this document show significance at P < 0.0125 or posterior probability > 98.75%).

Table 4.

Regions of significant linear correlations of BOLD signal with ratings of AROUSAL for TD participants

Anatomical regions	Location		MNI coordinates			Z
	Side	BA	x	y	z		Correlation
Linear arousal
Inferolateral prefrontal cortex	R		27	62	19	7.29	Negative
Broca's area	R		51	41	16	6.96	Negative
Posterior parietal cortex	R		30	−76	52	6.8	Negative
Dorsolateral prefrontal cortex	R		39	56	4	6.74	Negative
Primary motor cortex	R		57	11	34	6.74	Negative
Intraparietal sulcus	L		−44	−46	48	6.73	Negative
Parietal‐occipital‐temporal cortex	L		−42	−55	61	6.73	Negative
Precuneus	R		12	−70	64	6.34	Negative
Cuneus	R		21	−85	46	6.31	Negative
Broca's area	R		48	20	34	5.69	Negative
Insula	L		−39	−13	10	5.04	Negative
Insula	R		36	26	−2	4.53	Negative
Caudate nucleus	R		15	8	16	4.3	Negative
Visual association cortex	L		−3	−88	−5	4.21	Negative
Visual association cortex	L		−60	−10	−17	4.09	Negative
Hippocampus	R		18	−34	10	4.09	Negative
Putamen	L		−21	2	7	4.03	Negative

Conjunction maps of the linear and quadratic effects of arousal in each group show regions where both linear and quadratic effects were detected (e.g., Cb, Broca's, CN, DLPFC, PCu), and scatterplots show the combined effect of the linear and curvilinear components of the correlation (Fig. 4; Supporting Information Fig. S‐2, Tables S‐1A, S‐1C).

Figure 4.

Linear and quadratic correlations of arousal with BOLD‐signal in ASD (A) Regions of significant linear correlations of BOLD‐signal with ratings of arousal for participants with ASD. (Positive correlations are coded in red to yellow, and inverse correlations are coded in green to blue.). (B) Scatterplots of linear correlations for BOLD‐signal with ratings of arousal for participants with ASD (orange). (C) Regions of significant linear correlations of BOLD‐signal with ratings of arousal combined with regions of significant quadratic correlations of arousal ratings for participants with ASD. Positive linear correlations with arousal are shown alone (orange), quadratic correlations with arousal alone (fuschia), and linear+quadratic correlations with arousal (red). (D) Scatterplots of correlations between BOLD‐signal change and ratings of linear+quadratic correlations with arousal for participants with ASD (red) in regions where BOLD correlated positively for both linear and quadratic correlations with arousal ratings.

Linear and quadratic correlates of valence

BOLD‐signal correlated with ratings of linear valence and quadratic valence similarly for both diagnostic groups. In both TD and ASD participants, valence ratings correlated positively with BOLD‐signal in ACC, FG, and, and inversely with BOLD‐signal in posterior PC, S1, M1 and SMA. We did find regions where the correlation of valence ratings with BOLD‐signal differed significantly across the ASD and TD groups but they were very small in spatial extent and of questionable biological significance (Fig. 5, Tables 3 and 5; Supporting Information Tables S‐1B, S‐1D, S‐3).

Figure 5.

Valence correlates (A) Regions of significant correlations of BOLD‐signal with ratings of valence for TD participants. (Positive correlations are coded in red to yellow, and inverse correlations are coded in green to blue.). (B) Scatterplots of correlations for BOLD‐signal with ratings of valence for TD participants (green) or participants with ASD (purple). (C) Regions of significant correlations of BOLD‐signal with ratings of valence for participants with ASD. (D) The regions where the correlation of valence ratings with BOLD for participants with ASD differs significantly from the correlation of valence ratings with BOLD for TD participants. (ASD > TD coded in red to yellow, TD >ASD coded in blue to green). (E) Scatterplots of correlations for BOLD‐signal with ratings of valence for TD participants (green) and participants with ASD (purple) in regions where the correlation of valence ratings with BOLD for participants with ASD differs significantly from the correlation of valence ratings with BOLD for TD participants.

Table 3.

Regions of significant linear correlations of BOLD signal with ratings of VALENCE for participants with ASD

	Location		MNI coordinates
Anatomical regions	Side	BA	x	Y	z	Z	Correlation
Linear Valence
Cerebellum	L		−3	−52	−38	4.94	Positive
Cerebellum	R		9	−55	−38	4.69	Positive
Primary visual cortex	R	17/18	30	−58	4	4.58	Positive
Lingual gyrus	R	18	12	−88	−5	4.28	Positive
Dorsolateral prefrontal cortex	L		−18	62	10	4.01	Positive
Dorsolateral prefrontal cortex	R		24	62	10	3.97	Positive
Hippocampus	R		24	−16	−14	3.79	Positive
Dorsolateral prefrontal cortex	R		15	53	1	3.79	Positive
Fusiform gyrus	R		39	−22	−23	3.75	Positive
Primary visual cortex	L	18	−15	−55	10	3.74	Positive
Anterior cingulate cortex	L		−9	35	−5	3.55	Positive
Hippocampus	R		27	−40	1	3.54	Positive
Inferotemporal visual area	L		−39	−31	−14	3.29	Positive
Caudate nucleus	L		−3	20	−2	3.2	Positive
Dorsolateral prefrontal cortex	R		30	44	46	3.2	Positive
Orbitofrontal cortex	R		27	32	−11	3.04	Positive
Broca's area	L	44	−54	20	13	6.93	Negative
Primary somatosensory cortex	L	3b	−60	−4	19	6	Negative
Primary motor cortex	L		−54	5	28	5.93	Negative
Secondary somatosensory cortex	L		−60	−7	13	5.7	Negative
Insula	L		−33	23	10	5.33	Negative
Posterior parietal cortex	R		24	−70	49	5.28	Negative
Parietal‐temporal cortex	L		−48	−40	43	4.85	Negative
Primary somatosensory cortex	R	1	63	−13	31	4.85	Negative
Dorsolateral prefrontal cortex	R		27	5	55	4.58	Negative
Dorsolateral prefrontal cortex	L		−36	53	13	4.56	Negative
Visual association cortex	R		36	−82	34	4.52	Negative
Supplementary motor area	R	6	12	11	46	4.48	Negative
Parietal‐temporal cortex	R		45	−40	40	4.45	Negative
Dorsolateral prefrontal cortex	R	44	39	5	43	4.39	Negative
Broca's area	R	44	45	14	34	4.23	Negative
Dorsolateral prefrontal cortex	L		0	29	43	4.18	Negative
Inferotemporal visual area	R		54	−67	−8	4.14	Negative
Inferotemporal visual area	R		63	−55	−2	3.77	Negative
Primary visual cortex	L	18	−9	−88	−8	3.2	Negative
Broca's area	R	45	54	20	1	3.09	Negative

Table 5.

Regions of significant linear correlations of BOLD signal with ratings of VALENCE for TD participants

	Location		MNI coordinates			Z
Anatomical regions	Side	BA	x	y	z		Correlation
Linear valence
Primary visual cortex	R	17	12	−97	4	7.87	Positive
Cuneus	R	18	21	−100	10	6.44	Positive
Orbitofrontal cortex	L		0	56	−8	4.65	Positive
Precuneus	L		0	−58	19	3.42	Positive
Visual association cortex	L	18	−15	−103	1	3.23	Positive
Dorsolateral prefrontal cortex	R		3	59	10	3.2	Positive
Inferotemporal visual area	R		45	−73	−5	3.15	Positive
Anterior cingulate cortex	L		−3	26	−11	3.05	Positive
Visual association cortex	L		−42	−73	−5	2.91	Positive
Fusiform gyrus	L		−30	−31	−26	2.89	Positive
Posterior parietal cortex	R		18	−76	58	4.51	Negative
Parietal‐temporal cortex	L		−45	−55	58	4.25	Negative
Supplementary motor area	L		0	20	58	4.16	Negative
Precuneus	R		9	−61	49	4.11	Negative
Dorsolateral prefrontal cortex	R		18	8	70	3.98	Negative
Parietal‐occipital‐temporal cortex	L		−30	−52	37	3.97	Negative
Rolandic operculum	R		39	−19	22	3.95	Negative
Parietal‐temporal cortex	L		−51	−49	55	3.85	Negative
Precuneus	L		−6	−79	43	3.84	Negative
Middle cingulate cortex	L		−9	11	40	3.72	Negative
Primary motor cortex	R	6	51	−10	52	3.68	Negative
Primary somatosensory cortex	R	6	30	−31	67	3.67	Negative
Dorsolateral prefrontal cortex	L		−6	17	43	3.6	Negative
Lingual gyrus	L		−15	−85	−8	3.57	Negative
Inferotemporal visual area	R		63	−55	7	3.56	Negative
Cuneus	L		−6	−85	34	3.54	Negative
Primary motor cortex	L		−39	5	37	3.51	Negative
Broca's area	L		−48	41	1	3.33	Negative
Dorsolateral prefrontal cortex	L		−36	44	28	3.23	Negative
Auditory association cortex	R		54	−40	13	3.23	Negative
Dorsolateral prefrontal cortex	R		36	35	22	3.03	Negative
Middle cingulate cortex	L		−6	−19	34	2.93	Negative
Anterior cingulate cortex	L		−9	26	28	2.72	Negative
Middle cingulate cortex	R		12	−22	34	2.6	Negative

Interactions

Diagnosis‐by‐arousal

Our analyses revealed significant differences in the correlation of arousal ratings with BOLD‐signal across ASD and TD participants (ASD > TD) in AMYG, HIPP, CN, ACC, and SFG (Fig. 3, Table 6).

Table 6.

Centers of activation in regions where the linear correlation of BOLD signal intensity with ratings of AROUSAL for participants with ASD differs significantly from the linear correlation of BOLD with ratings of AROUSAL for TD participants

Anatomical regionsa	Location		MNI coordinates			Z
	Side	BA	x	Y	z		Correlation
Linear arousal
Broca's area	L	45	−48	41	−5	5.68	ASD>TD
Dorsolateral prefrontal cortex	L	9	−24	38	46	5.12	ASD>TD
Dorsolateral prefrontal cortex	R	8	12	50	49	5.09	ASD>TD
Broca's area	R	45	57	26	16	4.95	ASD>TD
Caudate nucleus	R		18	14	13	4.68	ASD>TD
Hippocampus	R		21	−28	−5	4.56	ASD>TD
Amygdala	R		27	−1	−17	4.41	ASD>TD
Anterior cingulate cortex	R	32	12	44	16	4.19	ASD>TD

^aRegions were significant at P < 0.001 or posterior probability > 99.0%. (All of the other tables in this document show significance at P < 0.0125 or posterior probability > 98.75%).

Diagnosis‐by‐valence

We only detected minor differences across groups in the correlation of valence ratings with BOLD‐signal, suggesting that brain activity in individuals with ASD is typical for some components of emotion‐processing (Fig. 5).

DISCUSSION

Our goal was to assess whether activity in neural systems underlying the processing of arousal and valence of facial emotions in individuals with ASD differ from their TD counterparts. Although participants from both diagnostic groups performed with a comparable level of accuracy on the emotion‐rating task [Tseng et al., 2014], we detected dramatic differences across diagnostic groups in the neural activity subserving the dimensions of emotion, particularly arousal. BOLD‐signal correlated linearly with ratings of emotional arousal, but in opposite directions and in differing locations for the two groups. BOLD‐signal in TD participants correlated inversely with ratings of arousal in regions associated primarily with attentional functions, whereas BOLD‐signal in ASD participants correlated positively with arousal ratings in regions most commonly associated with impulse control and with default‐mode activity. In contrast, we found that BOLD‐signal correlated with ratings of valence similarly across the TD and ASD groups, positively in regions associated previously with processing of emotional faces, and inversely in sensorimotor regions. Only minor group differences were detected in the BOLD‐signal correlates with valence ratings.

Arousal‐related Activity in TD Participants

Brain regions that correlated inversely with ratings of arousal in our TD group included the ILPFC, DLPFC, dorsal ACC, and inferoposterior and dorsal PC, as well as CN and PUT, regions thought to support attentional functions [Petersen and Posner, 2012]. Attention is the means by which brains optimize the flexible use of limited cognitive resources to process prioritized stimuli (i.e., to select task‐relevant and ignore task‐irrelevant information) [Mansouri et al., 2009], plan for future contingencies, inhibit competing responses, and pursue long‐term goals [Pennington and Ozonoff, 1996]. The neural substrates of attention are generally thought to include ACC, inferolateral and prefrontal cortices, and basal ganglia [Fan et al., 2005; Raz and Buhle, 2006], regions that correlated inversely with arousal ratings in the TD group.

The inverse linear correlations of arousal ratings with BOLD‐signal in attentional networks suggests that TD participants may have had to allocate progressively more attentional resources to processing incrementally less‐arousing face‐stimuli. Numerous studies have reported that TD participants view low‐arousal faces as ambiguous—i.e., as stimuli that may be construed in more than one way, or whose content requires greater examination of contextual cues to decipher [Kryklywy et al., 2013; Rosen and Donley, 2006]. Although earlier studies of ambiguity in emotional faces focused on fear‐related, high‐arousal stimuli (such as a fearful face signaling potential threat in the environment) [Adolphs, 2010; Whalen, 1998; Whalen et al., 2001], more recent studies have highlighted the inherent ambiguity in many low‐arousal faces (e.g., faces with neutral, bored, or calm expressions) [Thomas et al., 2001; Tottenham et al., 2013]. These low‐arousal faces provide neither clear safety nor danger signals about the immediate vicinity and so individuals must allocate more attentional resources to gauge the potential danger of the faces [Adolphs, 2010]. Thus, we posit that TD participants enlisted more attentional resources when viewing low‐arousal faces because they were attempting to assess and classify low‐arousal, ambiguous stimuli [Petersen and Posner, 2012; Posner and Petersen, 1990; Reiman et al., 1997]. Consistent with this interpretation, prior studies have reported that attentional networks activate more strongly when processing ambiguous stimuli [Mushtaq et al., 2011; Raz et al., 2007; Volz et al., 2004].

Stimulus Specificity in Activating Attention Networks during Emotional Tasks

Our current findings and their interpretations, when combined with those in our prior studies using differing types of emotional stimuli, suggest that the allocation of attentional resources may depend critically upon the nature of the emotional stimulus [Colibazzi et al., 2010; Gerber et al., 2008; Landa et al., 2013; Posner et al., 2009]. In the present study, for example, we found that neural activity within attention networks of the TD group correlated inversely with ratings of arousal in emotional faces. Yet in our prior study using a mood‐induction task, neural activity in attention networks of a different TD group correlated inversely with ratings of emotional valence, not arousal [Colibazzi et al., 2010; Landa et al., 2013] (i.e., neural activity during mood‐induction increased in attention networks as the valence of mood‐inducing stimuli became more unpleasant and aversive). Moreover, these correlations during mood‐induction were significantly stronger for interpersonal than for non‐interpersonal stimuli [Landa et al., 2013]. The inherently interpersonal nature of faces may account for the similar neural responses of potentially threatening faces with aversive interpersonal, mood‐inducing situations.

The differential activation of attention circuits in response to differing stimuli (emotional faces or induced mood) suggests that we cannot easily distinguish or disentangle the allocation of attentional resources from the two major dimensions of emotion, valence and arousal. It also demonstrates that instead of identifying neural activity that subserves or produces the arousal and valence components of emotional experience, our fMRI paradigm may be detecting varying activity in attention circuits associated with varying individual experiences of emotion. Our findings show that attention‐related activation can dominate overall neural activity during emotional tasks, overwhelming the variance in neural activity that produces arousal and valence.

Arousal in Participants with ASD

Regions where BOLD‐signal correlated positively with arousal ratings in the ASD group included posterior temporal and inferior PC, the mesial wall (pregenual and dorsal portions of the superior frontal and anterior cingulate gyri), PreMC and supplementary motor cortices, Cu and PCu, all basal ganglia nuclei, THAL, and dorsal Cb. Much of the pattern of neural activity associated with arousal in participants with ASD was similar to the pattern of regional activation in tasks that require suppression of an automatic response [Peterson et al., 2002; Viviani, 2013] and in preparing for action. The function of arousal is to prepare an organism for action, and the restraint of that motor preparedness from execution during the fMRI experiment would require activation of an inhibitory response. Alternatively, participants with ASD may find more‐arousing faces to be aversive, thereby requiring activation of control networks to suppress their motoric urge to withdraw from the putatively noxious stimulus [Peterson, 2003; Tabu et al., 2012]. Consistent with both interpretations, children with ASD reportedly have a stronger skin conductance response, and therefore are more aroused, when viewing faces with a direct compared with averted gaze, and compared to TD children in either condition [Kylliainen et al., 2012]. Also consistent with both interpretations are the maps of curvilinear (quadratic) responses of neural activity related to arousal ratings, which showed disproportionately stronger BOLD responses to more arousing stimuli in regions that support inhibitory responses, including dorsal frontal, ACC, and basal ganglia regions (Fig. 4).

Several regions that correlated positively with arousal ratings in the ASD group, including medial prefrontal, anterior cingulate, posterior cingulate, and precuneus cortices, are also components of the default mode network (DMN). Activity in the DMN is greatest at rest and declines during cognitively demanding tasks [Raichle et al., 2001]. In TD individuals, the DMN activates during performance of various social, emotional, and introspective tasks [Gusnard et al., 2001; Raichle et al., 2001]. Prior studies suggest that persons with ASD may not deactivate the DMN sufficiently when performing cognitively demanding tasks [Kennedy and Courchesne, 2008]. In participants with ASD, however, we found proportionately greater neural activity in DMN regions, presumably representing a more strongly felt social and emotional experience with progressively more‐arousing faces.

Group Similarities and Differences in Arousal Correlates

We were surprised to find the strikingly different patterns of activation to arousal ratings across the TD and ASD groups. Participants in our TD group increased neural activity in their attention networks when rating less‐arousing faces, a pattern absent in the ASD group. These strong inverse correlations of BOLD signal with arousal ratings in attentional networks of the TD group is counter to the more intuitive expectation that highly arousing or emotionally engaging stimuli would be more attention‐grabbing than less arousing and less emotionally engaging stimuli.

In contrast, individuals in the ASD group demonstrated strong positive correlations with arousal ratings in regulatory systems (IPC, SFG, THAL, CN, Cu) and the DMN (ACC, PCC), a pattern similar to that shown in TD controls in a prior mood‐induction study from our laboratory [Colibazzi et al., 2010]. This similarity suggests that participants with ASD, more than the TD controls, may have been increasingly immersed in the emotions expressed by the more arousing face stimuli of the present study. In other words, activation of the regulatory systems and DMN may require the absorbing, vitalizing experience of interpersonal stimuli, whether they are more highly arousing faces for persons with ASD in the present study or more arousing interpersonal moods for TD individuals in our prior study.

Thus TD participants in the present study may have instead approached the rating of facial emotions (as opposed to the mood induction task of our prior study) as more of a cognitive than a mood‐inducing task, with the rating of progressively less arousing, more ambiguous, and potentially threatening faces requiring a progressively greater allocation of attentional resources. Individuals with ASD, in contrast, seem either not to have experienced low‐arousal faces as threatening, or they found those faces less salient and less attention‐grabbing than did the TD controls.

Valence

Only minor differences were detected between diagnostic groups in the BOLD signal correlates of valence ratings. In ASD and TD groups, BOLD‐signal correlated positively with valence ratings in ACC, FG, and OFC, regions implicated in facial emotion processing. Studies of adults suggest, for example, that FG supports the perceptual identification of faces [Haxby et al., 2000] and more specifically the coding of fearful (high‐arousal, high‐valence) faces [Pessoa et al., 2002; Vuilleumier et al., 2001]. The near absence of significant group differences suggests that individuals with ASD do not activate the brain atypically for all components of emotion‐processing.

Limitations

One prominent limitation is the absence of eye‐tracking data during participants’ viewing of emotional faces. Prior studies have shown that individuals with ASD do not spontaneously attend to, and they may even avoid, the eyes of face‐stimuli, even though the eyes are a rich source of information about emotional states [Klin et al., 2002]. Although less attention to the eyes of our face‐stimuli could have impaired the ability of participants with ASD to recognize and rate accurately both valence and arousal [Kliemann et al., 2010], if participants with ASD had not attended to task‐stimuli, their valence and arousal ratings would have differed from the ratings of TD participants [Tseng et al., 2014]. Nevertheless, we cannot exclude the possibility that subtle group differences in attention to specific facial features influenced our findings.

Additionally, as is often the case in research on ASD, we struggled when designing our experiment with the trade‐offs between task difficulty, selection of a task that can provide scientifically important information, and the generalizability of the study and its findings to the entire autism spectrum. We considered multiple issues simultaneously. In particular, we needed to include in our study individuals who would and did understand a task that addressed meaningfully our fundamental research questions. Nonverbal individuals, for example, would be unlikely to understand or perform our task adequately. Also, if we had included lower functioning persons with ASD (i.e., those with lower IQs), we would have had to include control participants with comparable levels of intelligence, which in turn would introduce a host of confounding variables and sample heterogeneity that would make interpretation of findings difficult. We were careful to also covary for full‐scale IQ, as well as for age and sex, and detected no significant differences in our findings for diagnosis. Additionally, the individuals with ASD in our sample ranged in ASD diagnosis from PDD‐NOS to Asperger's to Autism (mean ADOS Score = 10.9 ± 3.1), suggesting that we can extrapolate our findings to individuals with moderate to high‐functioning ASD.

CONCLUSIONS

Our findings provide unique insight into the emotional experiences of individuals with ASD. Although behavioral responses to face‐stimuli were similar across diagnostic groups, the corresponding neural activity of the behavioral responses for arousal differed prominently across groups. TD individuals and persons with ASD seem to find differing aspects of emotional stimuli to be salient and relevant. Studying these differences may help us to understand the origins of atypical interpersonal emotional experiences in individuals with ASD.

Supporting information

Supporting Information