Visual Preference for Biological Motion in Children and Adults with Autism Spectrum Disorder: An Eye-Tracking Study

Participants aged ≥ 6 years with a confirmed diagnosis of ASD based on clinical examination, caregiver interview and use of the Autism Diagnostic Observation Schedule, 2nd edition (ADOS-2) (Lord et al. 2012) were enrolled. Key exclusion criteria were a measured composite score on the Kaufmann Brief Intelligence Test-2 (KBIT-2) (Kaufman and Kaufman 2004) of < 60, and history of or current significant medical illness. Each site also enrolled a control sample of TD participants, aged ≥ 6 years, with a score in the normal range on the Social Communication Questionnaire (Rutter et al. 2003) who did not meet criteria for any major mental health disorder (American Psychiatric Association 2013) assessed using the Mini-International Neuropsychiatric Interview (MINI) (Hergueta et al. 1998) for those 18 years old and above, and MINI-KID caregiver interview (Sheehan et al. 2010) for participants under 18 years. There were no exclusion criteria based specifically on vision. If participants required corrective lenses to view the screen and were able to obtain calibration on the measure at the start of each test set, then their eye-tracking data were included in the study.

In total, 136 individuals with ASD and 41 TD controls completed the study. After exclusions due to technical or calibration failures, the study population included 121 (89.0%) individuals with ASD and 40 (97.6%) TD controls. Table 1 lists characteristics of each analyzed group of participants. Further details on the participant characteristics can be found in Ness et al. (2019). Note that no socioeconomic data on the participants were collected.

The biological motion task consisted of videos in which two PLD animations were shown side-by-side. A PLD representing biological motion was shown on one side and a non-biological PLD was shown on the other. These stimuli were previously used by and reported by Umbricht et al. (2017). The PLDs were made up of dynamic black-point displays on a grey background (Supplementary Fig. 1). Each video lasted approximately four seconds. The biological/non-biological side was counterbalanced across presentations. Animations of biological motion were derived from a human actor’s performance and retrieved from the Carnegie Mellon University (CMU) Motion Capture Database (CMU Graphics Lab 2011). In contrast, animations of non-biological motion were computer-generated animations of moving dots, either phase-scrambled versions of the biological motion or continuous rotation (i.e. spinning around the vertical axis) of the first frame of the biological motion animation (see Supplementary Material for the analysis of attention to the two types of non-biological motion stimuli). All stimuli contained the same number of dots.

Biological and non-biological motion stimuli were matched in terms of motion complexity and speed of movement (see also Supplementary Material). There was an equal amount of novel stimuli in both the biological and non-biological motion condition. Both phase-scrambled and rotational control conditions were created using an original biological motion stimulus as a template. Phase-scrambled motion was created by time shifting each biological-motion point’s motion by a random time offset. These time offsets were randomly selected to sample uniformly between − 417 and 417 ms based on the number of points shown. The periodicity of this offset (833 ms) was ascertained by examining the dominant period of biological motion movement across clips (approximately 100 frames at 120 frames per second) via autocorrelation. This process guaranteed the total relative local motion of individual point-lights was comparable between phase-scrambled and unscrambled biological motion stimuli. The rotational point-light display was created by computing the dominant periodicity of hip movement in the source biological motion stimuli via autocorrelation and then rotating the first frame of the biological motion stimulus about the z-axis (following the spine in upright walking) at a constant speed given by hip periodicity. Stimuli were designed so that the first frame shown would be comparable across all conditions (unscrambled biological motion, scrambled biological motion, rotation).

This biological motion task was part of a large, observational, multi-center study conducted from 06 July 2015 to 14 October 2016 at nine study sites in the US (ClinicalTrials.gov, NCT02668991) and consisted of passive viewing tasks (Bangerter et al. 2020a, 2020b; Jagannatha et al. 2019; Manfredonia et al. 2018; Manyakov et al. 2018; Ness et al. 2019; Sargsyan et al. 2019). In this study, both groups of participants completed the same set of biosensor tasks. The total viewing time was approximately 40 min, including videos and other static stimuli presentations, divided into three sets, between which participants were allowed to take a break. Sixty different videos of biological motion were interspersed between these other stimuli and presented in 6 blocks of 10 videos (2 blocks within each set). An inter-stimulus slide that contained a cartoon image in the center of a grey screen was used for a duration of 1.5–2 s to re-orientate participants towards the center of the screen before each video, though no control was performed to verify the latter.

Scale data were collected concurrently with eye-tracking data. Parents or caregivers of individuals with ASD were required to spend at least 3 days per week with participants. They completed the following scales:

Each of the above scales and ADOS-2 consist of several subscales that reflect different ASD symptoms. All subscales used in this study are presented in Table 2. Standard scores for the above scales and the calibrated severity score of the ADOS-2 were used in the reported analyses.

Participants sat in a comfortable chair approximately 60 cm from a 23-inch computer screen (1920 × 1080 pixels). The height of the chair and screen were adjusted to ensure that participants’ eyes were level with the center of the screen. ET data were collected using the Tobii X2 eye tracker, with a sampling rate of 30 Hz, mounted below the screen. iMotions Biometric Research Platform (https://​imotions.​com/​) was used for stimuli presentation, data synchronization, and automatic calibration. Participants were allowed to freely observe presented stimuli. To ensure a high accuracy of the eye movement recordings (e.g., Blignaut and Wium 2014), before each experimental set a five-point calibration procedure consisting of animated cartoon characters paired with an auditory cue was performed. The calibration procedure was aimed to reach the mean distance between the participant’s gaze direction and the target points of less than 0.5° of visual angle.

For each frame, circles with the radius of approximately 3.7° of visual angle centered at each “black dot” corresponding to (non-)biological motion were drawn. Combination of such circles that corresponded to dots representing biological motion determined a biological motion region of interest (ROI). Similarly, a non-biological motion ROI was determined. Supplementary Fig. 1 shows two dynamic ROIs on a single example frame. Fixations were identified using the Binocular-Individual Threshold algorithm (van der Lans et al. 2011). The minimum required number of tracked samples and the maximum acceptable number of consecutive untracked samples in the eye tracking signal in a single fixation were both set to three. The following eye movement metrics were computed:

Analysis of covariance (ANCOVA) was conducted to compare eye movement metrics between the ASD and TD groups. Specifically, a single eye movement metric was used as a dependent variable, whereas participant group served as an independent variable. Participant’s age and gender entered the analysis as additional factors. Note that the KBIT-2 test was only administered in the ASD group and, thus, participant’s IQ score could not be included in the analysis as an additional factor. Supplementary Tables 6 and 7 show the results of tests for the homogeneity of variance assumption made in the ANCOVA models, and the results obtained with the same ANCOVA models as correcting for the violations of the assumption, respectively. Since ANCOVA models including two-factor interactions between participant group and the two other covariates did not significantly differ from those without any interaction (analysis of variance: all P-values > 0.18), statistical inference was based on the models without interactions. Note also that none of the scale data (see Behavior Rating Scales) entered any of the ANCOVA models. To test for effect of outliers on the obtained results, the same analyses were repeated after removing potential outliers from the analyzed dataset (Supplementary Table 1). To test for effect of site on the eye movement metrics, for each metric we conducted ANCOVA similar to that described above but additionally included a site identifier as an independent categorical variable. Similarly, to test for an interaction between participant’s age and group, the original ANCOVA models included an interaction term between these two factors. To test for effect of IQ on the differences in eye movement metrics between the two groups of participants, the entire ASD sample was divided into three separate groups based on the level of IQ. The tested levels represented the ASD participants with a low (range of the KBIT-2 IQ composite score: 60–84; n = 34 participants), average/normal (85–115; n = 63), and high (116–136; n = 24) IQ. The same ANCOVA’s as above were conducted combining the data of all TD participants and individuals with ASD with a specific level of IQ (Supplementary Table 2). The effect size of participant group on each eye movement metric was estimated using Cohen’s d and partial eta-squared (η_p²).

The Wilcoxon signed rank test was applied to compare preference for biological motion and “% time the first fixation was on biological motion” against the chance level (50%) in each of the two groups of participants separately. Similarly, to compare average latency of the first fixation between biological and non-biological stimuli within a single group of participants, the Wilcoxon matched pairs signed rank test was used.

Relationships between the eye movement metrics and different ASD symptoms were assessed using Spearman partial correlations, with participant’s age, gender and KBIT-2 IQ composite score serving as covariates (Supplementary Table 5). Spearman correlations were also used to assess relationships between the eye movement metrics and KBIT-2 IQ composite score (Supplementary Table 3) as well as between preference for biological motion and participant’s age (Supplementary Fig. 2). All reported correlations (r_S) were computed using the data of all ASD participants. The choice of Spearman correlations was attributed to a generally lower susceptibility of this type of correlations to potential outliers present in the data, as compared to Pearson correlations. Note, however, that qualitatively similar results were also observed when using Pearson correlations.

All statistical analyses were performed in MATLAB and R, with package “ppcor” in R used to compute partial correlations (Kim 2015). All reported P-values were two-sided, except those generated by ANCOVAs and based on F-statistic. The P-values corresponding to the differences in eye movement metrics between the two groups of participants were adjusted for multiple comparisons (five tests) using the Benjamini–Hochberg procedure (false discovery rate = 5%) and are reported as such throughout the main text. Non-adjusted, original P-values of all ANCOVA’s are listed in Supplementary Tables 1 and 2. No other correction for multiple comparisons was performed, since the study was exploratory rather than hypothesis-driven. Note that the number of statistical comparisons and, thus, the exact cut-off for significant P-values in each analysis was debatable. For example, the total number of tests entering the only reported adjustment for multiple comparisons (see above) was set to five. Alternatively, it could have been set to 15 when additionally counting the tests run to assess the effect of participant’s gender and age on the observed results. Adjustment for multiple comparisons for the relationships between the eye movement metrics and ASD symptoms (Supplementary Table 5) could have been done in multiple ways: for each symptom and behavior rating scale separately but across all five eye movement metrics, for each eye movement metric separately but across all symptoms and behavior rating scales, and combining all tests regardless eye movement metric, symptom and behavior rating scale. For the reasons outlined above, the P-values were reported “as is” with values < 0.05 considered significant.

Differences in eye movement metrics between ASD (n = 121 participants) and TD (n = 40) groups are shown in Fig. 1. Analysis of covariance revealed a significant effect of participant group on all metrics (all P-values corrected for multiple comparisons < 0.03), except “% time the first fixation was on biological motion” (Supplementary Table 1). On average, the ASD group showed a significantly lower level of visual attention to presented stimuli than the TD group (% total valid time; ASD vs. TD, mean: 80.0% vs. 89.2%; Cohen’s d = 0.68, η_p² = 0.08, P < 10^–3). In comparison to the TD group, the ASD group spent significantly less time looking at stimuli that depicted biological as opposed to non-biological motion (% preference for biological motion; ASD vs. TD, mean: 54.5% vs. 61.9%; Cohen’s d = 0.87, η_p² = 0.14, P < 10^–5). Overall preference for biological motion was significantly greater than the chance level (50%, equal looking time for both types of stimuli) in both groups of participants (both P-values < 10^–8). Preference for biological motion was the only eye movement metric that was significantly related to age, with older individuals (ASD and TD) showing a decrease in preference with age (Supplementary Table 1) (Supplementary Fig. 2). Percentage of time the first fixation was on biological motion was similar between the two groups of participants (ASD vs. TD, mean: 54.2% vs. 52.7%; Cohen’s d = 0.25, η_p² = 0.01, P = 0.26). In both groups of participants, percentage of time the first fixation was on biological motion significantly exceeded the chance level (in both groups P-value < 0.01). Individuals with ASD showed greater average latencies of the first fixation on both biological (ASD vs. TD, mean: 649.1 ms vs. 521.3 ms; Cohen’s d = 0.56, η_p² = 0.05, P < 0.01) and non-biological motion (639.7 ms vs. 530.9 ms; Cohen’s d = 0.46, η_p² = 0.04, P < 0.03), as compared to TD controls. There was no significant difference in average latency of the first fixation between biological and non-biological stimuli when data from each group were analyzed separately (both P-values > 0.94). Qualitatively similar results were observed when accounting for an interaction between participant’s age and group (the interaction term in ANCOVA’s: all five P-values > 0.24), a potential effect of site on the eye movement metrics (the effect of site in ANCOVA’s: all five P-values > 0.10), and the violations of the homogeneity of variance assumption in the analyses (Supplementary Tables 6 and 7). Furthermore, the effect of participant group on all five metrics persisted after removing potential outliers present in the data (Supplementary Table 1).

Restricting analyses to the groups of individuals with ASD with a specific level of IQ (low: n = 34 participants; average/normal: n = 63; high: n = 24) showed qualitatively similar results (Supplementary Table 2). Interestingly, as the IQ of the ASD group increased, “% total valid time” and average latency of the first fixation on either type of stimuli were more similar to those of the TD group. On the contrary, as the IQ of the ASD group increased, preference for biological motion deviated more from that observed in the TD group. Importantly, no systematic relationship between any of the five eye movement metrics and KBIT-2 IQ composite score was observed across the three tested IQ levels (Supplementary Table 3). Altogether, this rules out an explanation of the observed differences in the eye movement metrics between the two groups of participants by IQ.

To test for effect of severity of ASD symptoms on preference for biological motion for each behavior rating scale and symptom, participants with ASD were ranked according to the numeric score of that symptom and then split into three similarly-sized, non-overlapping groups (mild, moderate, severe) (Supplementary Table 4). For all five caregiver-reported behavior rating scales and ADOS-2, all symptoms, and three severity groupings, mean preference for biological motion exceeded the chance level (50%), reaching statistical significance in 80 out of 87 (92%) comparisons (all P-values < 0.05) (Table 2). All non-significant comparisons were in the mild symptom severity category.

To assess relationships between eye movement metrics and different ASD symptoms, as assessed by five caregiver-reported behavior rating scales and ADOS-2, each metric (n = 5) was examined for correlation with each symptom numeric score (n = 29). Only 5 of 145 computed correlation coefficients (3.4%) reached statistical significance (Supplementary Table 5; Supplementary Figs. 4–8). The ADOS-2 “restricted and repetitive behavior” significantly and positively correlated with % total valid time (r_S = 0.192, P < 0.04). Similarly, preference for biological motion correlated with the ABI “mental health” (r_S = 0.185, P < 0.05). Lastly, average latency of the first fixation on non-biological motion significantly and negatively correlated with the ABC “stereotypic behavior” (r_S = − 0.245, P < 0.01), the RBS-R “restricted behavior” (r_S = − 0.192, P < 0.04) and “stereotyped behavior” (r_S = -− 0.191, P < 0.04).

This study was part of a larger prospective, non-observational study designed to develop measures of change in ASD (Ness et al. 2019). One focus of the study was on obtaining quality data across a number of tasks and biosensors within the ASD population. Given the heterogeneity of this group, the aim was to maximize the collection of data from this sample. A group of TD individuals was added for comparison. As this was not the main focus of the study, this led to a smaller and less characterized TD group. In particular, the TD group did not have IQ measures, which limits inferences about the impact of IQ on performance. In addition, controls for multiple comparison were not performed in these analyses. The current results should be considered exploratory, and further validation in additional studies is required.

This study did not involve an intervention, and therefore the responsiveness to change of the features that discriminate between the ASD and TD groups could not be assessed, and this is a next step for an ongoing intervention study (NCT03664232). However, unlike Umbricht et al. (2017) who determined that first orientation to biological motion was less in an ASD group, and increased following vasopressin 1a receptor antagonist administration, we did not find evidence of between-group differences for this feature. Future studies are required to determine whether attention to biological motion can be increased in response to intervention, and whether this could be a precursor to other observable change in social interaction.