Atypical Value-Driven Selective Attention in Young Children With Autism Spectrum Disorder

This case-control study examines value learning of social and nonsocial stimuli and its association with selective attention in preschoolers with and without autism spectrum disorder.

Edition (5/128 or 4% of participants), or the Differential Ability Scales-II Early Years(DAS-II) (33/128 or 26% of participants) depending on age and ability level. 1 Autism severity was quantified using the Autism Diagnostic Observation Schedule-2 (ADOS-2). Ten out of 128 children (9 in the TD group and 1 in the DD group) did not receive the ADOS-2 assessment due to time constraints. The ADOS-2 Toddler Module was administered to 43/118 or 36% of participants, while Modules 1, 2, and 3 were administered to 15/118 (13%), 42/118 (36%), and 18/118 (15%) of the sample, respectively. The ASD group consisted of children who received a clinical best estimate (CBE) diagnosis of ASD and had an ADOS-2 calibrated severity score (CSS) above the ASD diagnostic 4-point cut-off. The DD group included children with specific or global developmental delays (n = 29), as well as ADHD, anxiety, social difficulties, or behavioral problem symptoms (n = 6). Children in the DD group underwent evaluation of their socialcommunicative skills using the ADOS-2 and an ASD diagnosis was ruled out by expert clinicians.
Children in the TD group had a CBE classification of normative development, ADOS-2 CSS of 3 or below, and verbal and nonverbal IQ scores above 85.
Participant data were excluded from the analysis due to low quality of eye-tracking data (calibration error > 2 degrees), insufficient number of valid trials (< 2) during Baseline or Choice Test phases, or because the participants completed fewer than 1/3 (< 16) of valid trials in the Training phase of the experiment. Based on these criteria, 7/55 (12.71%) ASD participants were excluded, compared to 4/53 (11.43%) DD and 2/38 (5.26%) TD participants (P = .49). The children excluded from the analysis did not differ significantly from the retained sample in age  Table 1 for sample characteristics).

Stimuli
A single pair of faces and a single pair of fractals were presented as stimuli, and all participants were exposed to these same four stimuli. To generate these stimuli, 12 fractal videos were produced using a generic algorithm-based fractal generation software 2 and 12 videos of female happy facial expressions were drawn from the BU-4DFE 3D Dynamic Facial Database. 3 The opening frames of the videos were selected as the static stimuli. To ensure that the two classes of stimuli were equivalently perceptually salient, the perceptual salience of each stimulus in all 24 videos was rated on 10-point scales by 17 adults through crowdsourcing 4 via Qualtrics. For the face stimuli, the magnitude of each emotional expression (neutral and happy) was also rated. Each stimulus was rated twice, first as a static still-frame representing the first frame of the video and again as a dynamic display. Means of all ratings were calculated across all stimuli. Subsequently, one pair of faces and one pair of fractals with saliency rating within a half point of the overall mean were selected to be used in this study. Stimuli sets were also standardized regarding their luminance, contrast density, texture similarity, size, color, intensity, and positioning. 5 See Figure 1 for a representation of the static versions of the stimuli.

Procedure
Each Baseline phase consisted of four free-viewing trials per participant to help examine any pre-existing attentional biases among the images prior to Training. During each Baseline phase trial, the HV stimulus (i.e., the stimulus that was later reinforced during Training) was presented simultaneously with the LV stimulus (i.e., the stimulus that was not reinforced during The Training phase consisted of 48 trials during which the HV and LV stimuli were presented 24 times in a random order and stimuli were never presented in the same location for more than two consecutive trials. Each trial started the presentation of a central fixation point (multi-color pulsating circle) for 600 ms, followed by the presentation of either the HV or the LV stimulus in one of four randomly selected locations for 2000 ms. There was a 300 ms overlap between the presentation of the central fixation point and the stimulus to diminish the likelihood that participants would look away from the screen prior to the onset of the peripheral stimulus. If a child fixated a HV stimulus, it underwent a visual transformation: the fractal would revolve and the face would smile. However, when the child failed to fixate on the HV stimulus, it remained static. If a child fixated a LV stimulus, the stimulus remained static. The reward value of each stimulus within a pair was randomized across subjects, so that in approximately half of the subjects, face 1 (or fractal 1) was reinforced during Training (HV stimulus), while face 2 (or fractal 2) was not reinforced (LV stimulus).
The free-viewing Choice Test phase consisted of six trials with the same structure as the Baseline trials, whereby HV and LV stimuli were presented simultaneously for 4000 ms in two out of four possible location selected at random. No reinforcement was given for looking at HV stimuli. To minimize the effect of memory load, choice tests were administered immediately after each participant completed the training phase. Eye tracking data were collected throughout the experiment.

Analysis of valid trials counts
During Baseline and Choice Test, a trial was considered valid if the participant had a total mean dwell time on either one or both of the stimuli greater than 100 ms and the calibration accuracy was below 2 degrees of visual angle. During Training, a trial was considered valid it the child shifted gaze toward the peripheral stimulus. Means and standard deviations for the counts of valid trials across all groups and phases of the experiment are shown in eTable 1.
In the Baseline phase, linear mixed effects model analysis on the number of valid trials, with group, condition, group x condition, and age as fixed effects revealed no effects of group, 043, but no effect of condition, F(1,91) = 1.32, P = .253, or interaction between group and condition, F(2,91) = 1.50, P = .228. The effect of age was not significant (P = .376). Post-hoc between-group comparisons indicated that children in the ASD group completed more Training trials than children in the DD group (P = .039) and a comparable number of trials to the TD group (P = 1.00); DD and TD groups did not differ (P = .280). Finally, during Choice Test, LMM analysis indicated no effect of group, F(2,112) = 0.88, P = .419, or condition, F(1,91) = 3.92, P = .051, but a significant group x condition interaction, F(2,91) = 3.33, P = .040. The effect of age was not significant (P = .243). Post-hoc between-group comparisons across the Face condition revealed no significant differences between ASD and DD groups (P = .218) or TD (P = .336) groups, or between DD and TD groups (P = 1.00). Similarly, there were no group differences in the Fractal condition between ASD and DD (P = .168) or TD (P = 1.00) groups, or between DD and TD groups (P = .187). Within-group comparisons indicated that ASD and TD groups had a comparable number of valid trials during Choice Test in the Fractal and Face conditions (P = .589 and P = .293, respectively), but the DD group contributed fewer trials in the Fractal than the Face condition (P = .007).

Accuracy of eye tracking data
After invalid trials were excluded, calibration accuracy was computed for the three groups and two conditions. Linear mixed effects model analysis on calibration accuracy, with group, condition, age, and age x group interaction as fixed effects, indicated no significant effects of group (P = .34), condition (P = .61), age (P = .57), group x condition interaction (P=.77), or age x group interaction (P = .69). Thus, there were no significant differences between groups in calibration accuracy in either the Face or the Fractal conditions (eTable 2). proportions. All models include main effects and interactions for diagnosis and condition. For age effects, "age|dx" includes an interaction between age and dx, and for covariance structures, the "|dx" notation indicates that a different covariance matrix is estimated for each diagnostic group. Akaike Information Criterion (AIC) values for maximum likelihood (ML) and restricted maximum likelihood (REML) fits are shown. For AIC, the smaller the better, so AIC favors the models in the top row of each The best model for the HV preference proportions analysis according to AIC included fixed effects of diagnosis, condition, and their interaction, but not age, and the best covariance matrix structure was diagonal, that is, with preference proportions in the Face and Fractal conditions being uncorrelated with different variances.