Abstract
We investigated the intra-individual variability of face scanning in autistic children to represent a new avenue for understanding abnormal face scanning in autism spectrum condition. Across four studies, we used eye-tracking techniques to systematically examine the variability of face scanning patterns in autistic children when performing different tasks and scanning different types of faces. Autistic and non-autistic children were asked to complete a face judgment task (Study 1, age range: 4.9–7.2 years), a face recognition task (Study 2, age range: 4.7–7.6 years), a facial expression recognition task (Study 3, age range: 4.3–7.4 years), and a dynamic facial expression free viewing task (Study 4, age range: 2.5–5.6 years). In addition, we conducted Study 5 using houses as stimuli to test the specificity of the results to faces (age range: 4.9–7.2 years). We found that scan pattern similarity between different face presentations was lower in autistic children than non-autistic children, which was robust to variations in experimental methods. Furthermore, the decreased scan pattern similarity in autism spectrum condition was evident in both viewing faces and houses. These results suggest that the scanning patterns of autistic children are noisier and variable. It might represent a new avenue for the understanding of core symptoms in autism spectrum condition.
Lay abstract
Atypical face scanning is suggested to be related to social interactions and communicative deficits in autistic children. We systematically examined whether autistic and non-autistic children used consistent scanning patterns when performing different tasks and scanning different types of faces. We found that autistic children scanned faces more variably than non-autistic children: While non-autistic children used more consistent scanning patterns, autistic children’s scanning patterns changed frequently when watching different faces. Autistic children’s variable face scanning patterns might delay and impair face processing, resulting in a social interaction deficit. What’s more, variable scanning patterns may create an unstable and unpredictable perception of the environment for autistic children. Developing in such an unstable environment might motivate autistic children to retract from the environment, avoid social interaction, and focus instead on the performance of repetitive behavior. Therefore, studying face scanning variability might represent a new avenue for understanding core symptoms in autistic people.
Autism spectrum condition (ASC) is a neurodevelopmental disorder characterized by impairments in social interaction and communication and by the presence of restricted interests and repetitive behaviors (American Psychiatric Association [APA], 2013). ASC is recognized as highly heterogeneous, that is, possessing high inter-individual variability (Lord et al., 2018). However, growing evidence shows that ASC is also characterized by greater behavioral and neural intra-individual variability than non-autistic people. This is supported by evidence that autistic individuals show greater trial-by-trial reaction time variability for a variety of tasks (Geurts et al., 2008; Karalunas et al., 2014), and more variable neural responses to the same stimulus across visual, somatosensory, and auditory domains (Dinstein et al., 2012; Haigh et al., 2015). Specifically, excessively variable neural responses reflect a less stable and predictable perception of the surrounding world, especially the perception of social situations, which involve a higher level of unpredictability (humans are more likely to exhibit variable behaviors than objects) (Dinstein et al., 2015; Sinha et al., 2014). To retreat from this unpredictable-seeming world, autistic people might avoid social interactions and instead focus on performing repetitive behaviors that generate more predictable neural responses, according to the prediction theory (Sinha et al., 2014). While admittedly speculative, the prediction theory motivates further studies of behavioral and neural intra-individual variability in ASC. Here, we investigated the intra-individual variability of face scanning in autistic children.
Atypical face processing is very common in autistic people; they have difficulty recognizing facial identities and expressions (Uljarevic & Hamilton, 2013; Weigelt et al., 2012). Some researchers even suggest that atypical face processing may contribute to the emergence of socially interactive and communicative impairments in ASC (Klin et al., 2015; Schultz, 2005; Shah et al., 2013). Difficulties in face processing may occur when one fails to look at the faces or most relevant facial features, such as the eyes (Chawarska & Shic, 2009; Circelli et al., 2013; Mckelvie, 1976; Sekiguchi, 2011). Autistic adults (e.g. Pelphrey et al., 2002; Yi et al., 2014), children (e.g. Chawarska & Shic, 2009; Speer et al., 2007; Wang et al., 2020), and even infants (e.g. Chawarska et al., 2013; Jones & Klin, 2013; Shic et al., 2020) showed reduced face looking time compared to non-autistic ones. These results have been confirmed by a recent meta-analysis of eye-tracking studies (Frazier et al., 2017). Thus, failing to attend to the critical facial features might contribute to face processing and social interaction impairments.
Here, we argue that studying looking time during face processing is not enough. Instead, the variability of the scanning pattern should also be considered as we further our understanding of abnormal face processing and social impairments in ASC. According to the signal detection theory (Green & Swets, 1966), detection capability is dependent on both the signal and the noise (i.e. signal-to-noise ratio). Although looking more at the relevant features could improve signal intensity, scanning pattern variability is more likely to be related to a person’s internal noise, which is suggested to be related to the natural variability of behavior. In fact, altering spatial attention could change neural noise (Arazi et al., 2019). For example, Arazi and colleagues (2019) demonstrated that when the non-autistic participants were guided by a cue to the position of an upcoming stimulus, perhaps inducing a more consistent looking pattern, their neural variability significantly decreased, and the magnitude of the decrease was associated with improved perception of the attended stimuli. Furthermore, some studies found that participants’ scanpaths become more similar as faces become increasingly familiar (in other words, the noise was reduced for these faces), reflecting greater efficiency in extracting facial information from familiar faces (Xiao et al., 2014). Thus, studying scan pattern variability might provide a new avenue to deepen our understanding of abnormal face processing and even social interaction and communicative impairments in ASC.
The current study used eye movements to systematically examine the variability of face scanning patterns in autistic children aged 2–7 years when performing different tasks and scanning different types of faces. As autistic children are usually first diagnosed in this age range, investigations based on children at this age may be less influenced by idiosyncratic compensatory strategies learned from diverse intervention experiences than those performed on adults or older children. We calculated the scan pattern similarity between different presentations of the same face or/and different faces for each child. We then compared the differences in the calculated scan pattern similarity between autistic children and non-autistic children in five studies. Study 1 asked children to judge whether they liked the presented faces. As identity and expression are the two main characteristics of a face, Study 2 asked children to memorize and recognize faces, and Study 3 asked children to identify happy and angry facial expressions. Studies 1–3 used static faces with a limited number of facial expressions, which may not fully represent the everyday social experiences of children. Study 4 sought to examine the scan pattern similarity when viewing neutral faces gradually changing into six different expressions in a dynamic fashion, which was designed to increase the real-world applicability of findings. We hypothesized that autistic children’s face scanning patterns were noisier and variable, thus exhibiting lower scan pattern similarity than non-autistic children across different studies. Finally, we conducted Study 5 using houses as stimuli to test whether the lower scan pattern similarity in ASC was restricted to the face stimuli. Table 1 summarizes the stimuli and tasks used in the five studies.
A summary of the stimuli and tasks used in this article.
Method
Participants
Participants were all Han Chinese children. The same children completed Study 1 and Study 5, and different children completed Studies 2–4. Autistic children were all previously diagnosed by professional pediatricians in licensed hospitals according to the criteria for ASC in DSM-5 (APA, 2013). ASC diagnoses were further confirmed using the Autism Diagnostic Observation Schedule (ADOS; Lord et al., 2000) in Studies 1, 4, and 5. Non-autistic children were recruited from typical kindergartens or primary schools, and their teachers and parents reported no concern about any potential developmental or psychiatric disorder.
Children’s IQ was measured by the Wechsler Intelligence Scale in Study 1 and Studies 3–5, Raven’s Test (CRT-C2) (Raven et al., 1998) and Peabody Picture Vocabulary Test-Revised (PPVT-R) (Vance & Stone, 1990) in Study 2. Children with an IQ lower than 65 (measured by the Wechsler Intelligence Scale) were excluded from the analysis (one autistic child in Studies 1 and 5; five autistic children and one non-autistic child in Study 4). These children were scored above the threshold cutoff for cognitive impairment, and the ASC and non-ASC groups were matched after excluding these children. One autistic child in Studies 1 and 5, one autistic child and two non-autistic children in Study 3, and three autistic children in Study 4 were further excluded from the analysis since they had too many invalid trials (see “Eye Movement Data Analysis” section in the supplementary material for the definition of invalid trials and exclusion criteria). Valid trial number and total trial number for each study are presented in Table 2.
Characteristics of the included participants.
Note. Full-scale IQ was measured using the abbreviated Chinese Fourth Edition version of the Wechsler Intelligence Scale for Preschool and Primary Children (Wechsler, 2014b) for children below the age of 6 years and the abbreviated Chinese Fourth Edition version of the Wechsler Intelligence Scale for Children (Wechsler, 2014a) for children above the age of 6 years. In Study 2, nonverbal and verbal mental intelligence were measured by Raven’s Test (CRT-C2) (Raven et al., 1998) and Peabody Picture Vocabulary Test–Revised (PPVT-R) (Vance & Stone, 1990), respectively. Gender information for Study 2 was unavailable because the experimenter forgot to record it. SA and RRB Severity were calculated according to Gotham et al. (2009) and Hus et al. (2012). ADOS: Autism Diagnostic Observation Schedule; SA Severity: ADOS Social Affect Severity; RRB Severity: ADOS Restricted, Repetitive Behavior Severity.
p < 0.05, **p < 0.01, ***p < 0.001.
The final sample information is shown in Table 2. Children’s age range was 4.9–7.2 years for Studies 1 and 5, 4.7–7.6 years for Study 2, 4.3–7.4 years for Study 3, and 2.5–5.6 years for Study 4. Specific data on socioeconomic status were not recorded. The ASC and non-ASC groups were matched by chronological age and IQ (Table 2).
The present protocol was conducted according to the ethical standards laid down in the 1964 Declaration of Helsinki and approved by the sponsoring university’s Ethical Committee. We obtained oral consent from all of the children and written consent from all of their parents before conducting these experiments.
Materials, procedures, and data analysis
Detailed methods, statistical models, and other associated references are available in the supplementary material online.
In Study 1, faces in grayscale with a neutral expression were presented one by one to children for 2.5 s, and children were instructed to answer whether they liked the presented face. The task included two blocks, and the faces in Block 1 were repeated in Block 2. The faces within a block were different from each other.
In Study 2, face recognition was examined (all faces had a neutral expression). Children were first presented by a single face for 3 s and were asked to remember it (memory phase). Then they were presented with two faces, one of which matched the first face, and the other was a new face with the same gender (recognition phase). Children were asked to decide which face was the same as the one they had just seen. The stimuli remained until children gave their verbal responses. The experimenter immediately recorded the verbal answer by pressing one of the two keyboard buttons.
In Study 3, we examined children’s facial expression recognition by presenting them with a single face for 2 s and asking them to judge whether the face was angry or happy. The faces were artificially created. When children gave their verbal responses, an experimenter immediately recorded the verbal answer by pressing one of the two keyboard buttons.
In Study 4, we used dynamic faces (videos) instead of static faces. These faces were movies of artificially created faces. Children were presented with a single face gradually changing from a neutral expression to anger, disgust, fear, happiness, sadness, or surprise for about 3 s. There was no task in this study.
The procedure for Study 5 was the same as that of Study 1, except that it used houses instead of faces as the stimuli. The same participants completed the two studies. In brief, in Study 5, the house images in grayscale were presented one by one to children for 2.5 s, and children were instructed to answer whether they liked the presented house. There were two blocks: The house images in Block 1 were repeated in Block 2. Within each block, the houses were never repeated.
Eye movement data were recorded by a Tobii TX-120 eye-tracker (Studies 1, 2, and 5) or a Tobii Pro X3-120 eye-tracker (Studies 3 and 4). The sampling rate was 120 Hz.
We used two approaches to calculate the scan pattern similarity for two stimuli (the supplementary material online provides detailed calculation methods). The first one measures the similarities in the gaze patterns between two images based solely on the spatial (but not temporal) allocation of gaze, referred to as the space-based similarity score. If a participant looks at the same regions for two images, the space-based similarity score will be high (Figure 1).
Three versions of simulated gaze patterns. Spearman correlations are a straightforward method to capture similarity level between heatmaps and match closely with human intuition: the Spearman correlation coefficient was 0.82 for A and B, and 0.64 for A and C. A and B also show the temporal traces of fixations. Even though the overall spatial pattern may be similar between A and B, the temporal order of the gaze may differ dramatically.
It is possible that a participant might look at the same regions for two images, but the temporal order of fixations is very different (Figure 1). To consider both the spatial and temporal allocations of gaze, we used the second approach based on the ScanMatch method (Cristino et al., 2010) to calculate the scanpath-based similarity score (Figure 2). If a participant uses similar scanpaths (looks at similar regions with similar temporal orders) to scan two images, the scanpath-based similarity score will be high.
To calculate the scanpath-based similarity score, a grid (1 × 1° visual angle) was overlaid onto the image stimuli. We also examined whether our results were robust to grid sizes by increasing the size of the grid to 1.5° and 2° visual angles (presented in the supplementary material). The resulting rectangle regions were labeled with two letters. Each fixation sequence was recoded to a letter sequence representing fixation location and time order according to its spatial locations on the rectangle regions. In this example, the letter sequence was gO iB jM lL. Note that this letter sequence contains both spatial and temporal information. The letter sequences were then compared with each other, and a similarity score was calculated based on the Needleman–Wunsch algorithm (Needleman & Wunsch, 1970).
We examined the group differences in the scan pattern similarity score using a linear mixed model (LMM) in Studies 1–5. In Studies 3 and 4, the LMM also included expression and its interaction with the Group variable. Studies 2 and 3 also explored whether increased face scanning pattern variability was related to worse performance in identifying the face or recognizing the expression using a linear model. Specifically, we used Group, similarity score, and interaction between Group and similarity score to predict the face recognition accuracy (ACC) in Study 2 and discriminating ability d’ of expression recognition in Study 3. Response time (RT) was not analyzed because the experimenters, not the children, pressed buttons. Given that the space- and scanpath-based similarity scores were highly correlated (Pearson r = 0.73 in Study 2 and 0.79 in Study 3), we averaged the two types of similarity scores to create one similarity score for the simplicity of analysis. We also tested the correlation between individual similarity scores to faces and houses, given that the same participants were tested in Studies 1 and 5. This was to probe whether the profiles were specific to social stimuli or related to looking behavior in general. Furthermore, we tested the association between similarity scores and autistic symptoms in the ASC group in Studies 1 and 5. We used standardized ADOS scores, including social affect (SA), restricted and repetitive behavior (RRB), and total severity scores (Gotham et al., 2009; Hus et al., 2012), rather than ADOS raw scores, to measure autistic symptoms because standardized scores were less influenced by participant demographics than raw scores. Standardized ADOS scores were not available in Study 4 because some ADOS items were not measured (not measuring these items does not influence the diagnosis of ASC). Finally, because we conducted multiple studies using comparable measures to validate the hypothesis that autistic children scanned faces more variably than non-autistic children (Studies 1–4), we amalgamated these studies using a meta-analysis based on the random effects model. See the supplementary material online for these detailed statistic methods.
Community involvement
There is no community involved in this study.
Results
Study 1
The scan pattern similarity score was calculated for both the repeated faces (the same faces from Blocks 1 and 2) and different faces (between all stimuli regardless of blocks). Please see the details in the supplementary material. LMM revealed a significant main effect of Group, F(1, 48.18) = 12.66, p < 0.001, for space-based similarity of repeated faces; F(1, 49.98) = 12.03, p = 0.001, for space-based similarity of different faces; F(1, 49.14) = 8.63, p = 0.005, for scanpath-based similarity of repeated faces; and F(1, 49.97) = 7.80, p = 0.007, for scanpath-based similarity of different faces. Autistic children had a lower scan pattern similarity score than non-autistic children (Figure 3(a) and (b)), suggesting increased face scanning pattern variability in ASC.
Boxplots showing reduced scan pattern similarity for faces in autism spectrum condition (ASC) when performing (a and b) likeness judgment, (c) identity recognition, (d) expression recognition, and (e) passive viewing tasks. (f and g) Autistic children also showed reduced scan pattern similarity when viewing houses than non-autistic children. Each black triangle or black circle represents the mean value, each thick black vertical line representing the error bar, and each gray point representing one child’s data.
Study 2
The scan pattern similarity score was calculated for faces from different trials (between all stimuli) in the memory phase. We did not record eye movement data during the recognition phase because two faces were presented on the screen. LMM only revealed the marginally significant Group effects for both the space-based similarity score, F(1, 43.21) = 2.98, p = 0.091, and the scanpath-based similarity score, F(1, 42.95) = 4.02, p = 0.051. Autistic children tended to have lower scan pattern similarity scores than non-autistic children (Figure 3(c)).
To explore the relationship between the face scanning pattern variability and the identity recognition performance, we used Group (ASC or non-ASC), the similarity score (averaged across the space- and scanpath-based similarity scores), and the interaction between Group and similarity score to predict face recognition accuracy using a linear model. We only found the Group main effect, β = 0.20, t = 4.14, p < 0.001, suggesting that the accuracy was lower in autistic children than non-autistic children.
Study 3
Scan pattern similarity score was calculated for faces from different trials (between all stimuli) separately for angry and happy faces to examine whether expressions influenced face scanning pattern variability. LMM revealed that the main effect of Group was significant for both the space-based similarity score, F(1, 84.03) = 7.31, p = 0.008, and the scanpath-based similarity score, F(1, 83.48) = 5.14, p = 0.026. Autistic children had a lower scan pattern similarity score than non-autistic children (Figure 3(d)). The main effect of Expression was significant for the space-based similarity score, F(1, 83.15) = 4.14, p = 0.045, but not significant for the scanpath-based similarity score, F(1, 84.79) = 0.14, p = 0.709. There was no evidence that expression modulated the difference in scanning pattern variability between autistic and non-autistic children since the interaction effect between Group and Expression was not significant for both the space-based similarity score, F(1, 83.15) = 0.43, p = 0.516, and the scanpath-based similarity score, F(1, 84.79) = 0.02, p = 0.876.
To explore the relationship between face scanning pattern variability and expression recognition performance, we used the Group variable (ASC or non-ASC), the similarity score (averaged across the space- and scanpath-based similarity scores), and the interaction between Group and similarity score to predict the discriminating ability d’ of expression recognition using a linear model. We only found the significant main effect of similarity score, β = 2.58, t = 2.42, p = 0.018, suggesting that increased face scanning pattern variability is associated with poorer facial expression recognition performance.
Study 4
The scan pattern similarity score was calculated for dynamic faces from different trials/videos (between all stimuli), separately for different expressions. In line with the results from Study 3, LMM revealed that the main effect of Group was significant for both the space-based similarity score, F(1, 37.53) = 5.28, p = 0.027, and the scanpath-based similarity score, F(1, 37.45) = 8.94, p = 0.005. Autistic children had a lower scan pattern similarity score than non-autistic children (Figure 3(e)). The main effect of Expression was not significant for both the space-based similarity score, F(5, 37.40) = 2.41, p = 0.055, and the scanpath-based similarity score, F(5, 45.83) = 2.14, p = 0.077. There was no evidence that expression modulated the difference in scanning pattern variability between autistic and non-autistic children since the interaction between Group and Expression was not significant for both the space-based similarity score, F(5, 37.40) = 1.20, p = 0.329, and the scanpath-based similarity score, F(5, 45.83) = 1.01, p = 0.420.
Study 5
Unlike the face images, in which different faces had the same size and shape, and in which the facial features (eyes and mouth) were almost at the same position across different faces, different houses had varied sizes, shapes, and so on. Therefore, the scan pattern similarity score was only calculated for the repeated house images. LMM revealed that the main effect of Group was significant, F(1, 48.84) = 10.68, p = 0.002, for the space-based similarity score; F(1, 46.09) = 6.69, p = 0.013, for the scanpath-based similarity scores. Autistic children had a lower scan pattern similarity score than non-autistic children, whether based on the space or scanpath (Figure 3(f)), suggesting that the increased scanning pattern variability in ASC is not limited to faces.
Studies 1 and 5 combined
We combined the data from Studies 1 and 5, included Stimuli Category (face vs house) to an LMM, and found that the main effect of Group was significant, F(1, 48.55) = 15.66, p < 0.001, for the space-based similarity score; F(1, 47.50) = 10.84, p = 0.002, for the scanpath-based similarity score (Figure 3(g)). The main effect of Stimuli Category was significant, F(1, 49.11) = 27.72, p < 0.001, for the space-based similarity score; F(1, 50.61) = 4.87, p = 0.032, for the scanpath-based similarity score. Children scanned faces more consistently than houses. The interaction effect was not significant, F(1, 49.11) = 2.01, p = 0.163, for the space-based similarity score; F(1, 50.61) = 1.71, p = 0.197, for the scanpath-based similarity score.
We further tested the correlation between individual similarity scores to faces and houses. For the space-based similarity score, the correlation between viewing faces and houses was significant for the autistic children, r = 0.67, p < 0.001, but not significant for the non-autistic children, r = 0.18, p = 0.440. For the scanpath-based similarity score, the correlation was also significant for the autistic children, r = 0.64, p < 0.001, but not significant for the non-autistic children, r = 0.22, p = 0.350. The results are shown in Figure 4.
Correlation between individual similarity scores to faces (Study 1) and houses (Study 5).
Finally, we tested the relationship between similarity scores and autistic symptoms. We conducted several Pearson correlations between the space-based and scanpath-based similarity scores (face and house separately) and standardized ADOS scores (SA, RRB, and total severity scores separately). No significant correlations were found (rs were from −0.25 to 0.08; ps were from 0.18 to 0.96).
Meta-analysis from Studies 1–4
Detailed results are shown in Figure 5. In brief, the overall effect was significant for both the space-based similarity score, t = −7.42, p = 0.005, and the scanpath-based similarity score, t = −5.56, p = 0.012, suggesting that autistic children scanned the faces more variably than non-autistic children.
Meta-analysis results for the (a) space-based and (b) scanpath-based similarity score from Studies 1–4.
Additional analysis
We compared group differences in number of valid trials, image viewing time, and oculomotor function (mean fixation duration and saccade amplitude).
In brief, we found that autistic children had fewer valid trials than non-autistic children in all studies. Number of valid trials was also related to similarity scores in all studies. Therefore, we reran the LMMs to reveal Group differences in scanning pattern similarity scores with the number of valid trials included as a covariate. We still found the group differences in at least one type of similarity score across all studies except for Study 2.
As for the image viewing time, autistic children viewed images for shorter periods of time than non-autistic children in Studies 1 and 5 combined and Study 2. Finally, for the oculomotor function, we found autistic children had a shorter mean fixation duration than non-autistic children in Studies 1 and 5. There were no other significant differences.
We also examined whether our results of scanpath-based similarity score were robust to grid sizes by increasing the grid size to 1.5° and 2° visual angles. The results remained the same.
Detailed results are presented in the supplementary material online.
Discussion
Across four studies and using the meta-analysis, we found that autistic children scanned faces more variably than non-autistic children. We conclude from these findings that autistic children scanning faces less reliably is a reproducible finding that appears relatively robust to variations in experimental methods. Furthermore, the increased scanning pattern variability in ASC was evident in both viewing faces and houses, and the correlation between individual similarity scores to faces and houses was also significant in autistic children. Thus, this visual scanning atypicality in ASC is not restricted to faces or social stimuli but may reflect a more generalized atypicality.
The group differences were not likely to be explained by the group differences in oculomotor function, as the two groups had similar mean fixation duration and saccade amplitude. Beyond investigating looking times to the face or the social scene, in recent research, experimenters have also used diverse eye movement indexes to understand the circumscribed viewing patterns in autistic people. Mean fixation duration and saccade amplitude are two such indexes—smaller saccade amplitude was described as a tendency to explore areas closer to the current fixation and longer fixation duration as “sticky attention” or a tendency to view persistently in one location. Studies have found that autistic people or people with more autistic traits exhibit circumscribed viewing patterns when looking at a social scene or during face-to-face interactions (Heaton & Freeth, 2016; Vabalas & Freeth, 2015). Increased circumscribed viewing patterns were also related to a more “rigid” personality (i.e. having little interest in change or difficulty adjusting to change) (Heaton & Freeth, 2016). However, opposite results also exist. For example, Vettori et al. (2020) found that the autistic group used a more exploratory and less stable scanning strategy when viewing isolated static faces than non-autistic group. It is possible that the circumscribed viewing patterns may depend on stimuli content. Despite the inconsistent results, these studies and our own underline the importance of using diverse data analysis methods beyond calculating looking times to pinpoint face scanning patterns, which could shed new light on the core symptoms in ASC.
We were not the first to study scanning pattern variability in autistic individuals: three studies have explored this issue to our knowledge (Avni et al., 2019; Król & Król, 2019a, 2019b). All these studies found that autistic individuals scanned the same social scene in a more variable way than non-autistic individuals. In those studies, social stimuli were relatively complex, which required integrating social interaction information (e.g. mutual gaze) and nonsocial information related to social information (e.g. cups being looked at). These experimental settings made it impossible to disentangle the contributions of social and nonsocial information to the variable scanning patterns in ASC. We extended those findings by focusing on face processing, in particular, one important component of social interaction. Furthermore, we conducted multiple experiments in one study to confirm that variable face scanning patterns in ASC is a reproducible finding that appears to be relatively robust to various experimental settings and tasks. However, in contrast to our study, Król and Król (2019b) found that autistic individuals had a higher scanpath similarity in response to nonsocial natural scenes than non-autistic individuals. This discrepancy is probably due to the fact that the natural scene stimuli used in that study included objects related to circumscribed interests typical for ASC (i.e. vehicles), and those objects might create a stable perceptual experience for autistic individuals, reflected in their more stable eye movement pattern. Future studies might compare the differences between scanning objects related to and unrelated to circumscribed interests to illustrate this issue further. In addition, the current sample included children aged 2–7 years, whereas the Król and Król (2019b) study tested adolescents or young adults (mean age was around 17 years). This age difference between the samples could contribute to the inconsistency between their findings and ours. It is possible that autistic people’s scanning patterns for nonsocial objects are variable in childhood and normalize or become less variable with development. Considering that early development has cascading developmental consequences, it will be essential to understand if there are divergent developmental trajectories for scanning social and nonsocial objects from infancy to adults, and if so, how these contribute to autistic symptoms throughout development.
We speculate that the lower reliability of visual scanning patterns in autistic children may be related to their core symptoms. Variable scanning patterns in autistic children may create an unstable and unpredictable perception of the environment for them. This could explain why autistic individuals may be less able to learn the correct probabilities and statistics from external events, especially from everyday social situations, and provides novel evidence for the notion that impairment within the ASC phenotype is linked to their poor predictive ability (Lawson et al., 2013; Pellicano & Burr, 2012; Sinha et al., 2014; Van de Cruys et al., 2014). Developing in such an unstable environment might motivate autistic individuals to retract from the environment, avoid social interaction, and focus instead on the performance of repetitive behavior, which is more stable. If our speculation is true, scanning pattern variability will be etiologically meaningful, which poses limitations on only studying looking time in ASC. However, our subanalysis among the ASC group of Studies 1 and 5 indicated no evidence that the scanning variability and symptom severity were related. One possible explanation is that scanning variability only affects social and repetitive behaviors during a sensitive development window in infancy, which further determines later autistic core symptoms in childhood. Future longitudinal studies from both individuals with and without ASC should consider the trial-to-trial variability to gain further insight into the developmental trajectory of scanning pattern variability and to contribute to the outstanding fundamental question of whether scanning pattern variability is etiologically meaningful.
However, core symptoms of ASC resulting from reduced social motivation (Chevallier et al., 2012) might also increase face scanning variability in autistic children. The social motivation theory suggests that autistic children may find faces less meaningful and thus have less motivation to process them (Chevallier et al., 2012). A previous study found that viewers’ eye movements became more random when the environment was perceived as less meaningful (Jordan & Slater, 2009). Similarly, if faces are less meaningful to autistic children in accordance with the social motivation account, their face scanning patterns should also be more random and less strategic, leading to increased face scanning variability. However, the social motivation theory could not explain why autistic children in the current study also scanned houses more variably, as houses are not assumed to be less meaningful to autistic children than non-autistic children. To further illustrate this issue, future investigations comparing the differences between the scanning pattern variability of objects that are meaningful and not meaningful to autistic children are recommended.
As we mentioned in the introduction, scanning pattern variability is likely to influence face processing, resulting in a social interaction impairment. The face scanpath of autistic children is “noisier” and less predictable. Noisier perception of the face in ASC might delay their face processing and increase errors when extracting facial information. Our study partially supported this hypothesis. Based on the current data, increased face scanning pattern variability was associated with worse expression recognition performance but not identity recognition performance. We proposed three explanations for the inconsistent results. First, the number of participants in Study 3 (expression task) was almost twice that of Study 2 (identity task). It is possible that when we increase participant numbers in Study 2, we will get a similar result as revealed in Study 3. Second, a noisier perception of the face may increase the time to extract facial information. However, in both studies, experimenters helped children press key buttons, while children only made verbal responses. Thus, the RT might not be very reliable, and we did not report the RT results. Even so, we still analyzed the RT data in an exploratory way. The results revealed negative correlations between scan pattern similarity score (scanpath-based in Study 2 and space-based in Study 3) and RT in both studies. Thus, it appears that higher face scanning variability could increase face processing time. Considering the limitation of the current designs, the approach used should be regarded as exploratory. Thus we suggest that future studies should use more sophisticated designs, such as asking participants to press keys by themselves to further understand the relationship between face scanning variability and face processing ability. The last possibility is that face processing performance is related to face scanning pattern variability during recognition, but not memorizing, considering that scanning pattern was analyzed during the recognition phase in Study 3 but not Study 2. In Study 2, we did not record eye movement data during the recognition phase because two faces were presented on the screen. Future studies could test this possibility by presenting one face during the recognition phase.
What’s the neural mechanism underlying scanning pattern variability in autistic individuals? Recently, an emerging body of evidence suggests autistic individuals exhibit more variable neural responses to the same stimuli (David et al., 2016; Dinstein et al., 2012; Edgar et al., 2015; Gandal et al., 2010; Haigh et al., 2015, 2016; Noordt et al., 2017; Sun et al., 2012). Thus, it is reasonable to assume that less reliable neural activity may generate variability in the scanning pattern. However, the situation is not so simple: Attention could also change neural variability. For example, when participants’ attention was guided to the position of upcoming stimuli, neural variability decreased (Arazi et al., 2019). Therefore, less reliable perceptual experiences, reflected in less stable scanpaths, may also lead to less reliable neural activity. Future studies aiming to understand the neural mechanism of scanning pattern variability in autistic individuals should consider the possibility of mutual influence between scanning pattern and neural activity variabilities.
Neural variability is not unique to autistic children—other disorders such as attention deficit hyperactivity disorder (ADHD) also exhibit excessive neural variability compared to non-ADHD individuals. Furthermore, numerous studies have reported increased RT variability across trials in individuals with ADHD (Adamo et al., 2018; Karalunas et al., 2014). Therefore, future research is required to establish the universality of increased scanning pattern variability in ASC and other developmental disorders (ADHD in particular). When evaluating this issue, the primary goal is to perform identical experiments and analyses with subgroups of individuals who fulfill clinical criteria for ASC, ADHD, and both disorders. Additional distinctions across these two disorders should use the longitudinal method to understand the developmental timing of scanning pattern variability (e.g. early infancy versus late childhood). Finally, it is entirely plausible that scanning pattern variability may appear across both autistic people and those with ADHD and affect their social interaction.
Several other potential future directions should be considered. First, our study excluded children with low IQ and therefore described only a subset of autistic children. Further research should be carried out to obtain a more fine-grained understanding of scanning variability across all autism profiles. Second, it is surprising that the group effect was not significant for the face identification task, given that face identification was impaired in our study and in previous studies (e.g. Weigelt et al., 2012). As mentioned before, we only recorded data during the memory phase, but future studies should also test the group effect during the recognition phase. Finally, the methods for calculating similarity scores were sensitive number of valid trials, which was reduced in autistic than non-autistic children. Although results were somewhat robust to differences in the number of trials, it is still important for future research to elaborate methods that are not sensitive to differences in the number of valid trials or test this method in equally engaging tasks for children with and without ASC.
In sum, our results revealed a consistently larger face scanning pattern variability in autistic children than non-autistic children across multiple types of stimuli, tasks, and experimental designs. Furthermore, visual scanning variability in ASC also extended to nonsocial objects (i.e. houses). We suggest that the notion of increased scanning pattern variability in ASC might represent a new avenue for understanding core symptoms in ASC. However, our study only represents the first step. More future studies are required to gain further insights into the developmental trajectory, neural mechanism, and clinical implication of scanning pattern variability in ASC and other developmental disorders.
Supplemental Material
sj-docx-1-aut-10.1177_13623613211064373 – Supplemental material for Investigating intra-individual variability of face scanning in autistic children
Supplemental material, sj-docx-1-aut-10.1177_13623613211064373 for Investigating intra-individual variability of face scanning in autistic children by Qiandong Wang, Haoyang Lu, Shuyuan Feng, Ci Song, Yixiao Hu and Li Yi in Autism
Footnotes
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Fundamental Research Funds for the Central Universities (Q.W., grant number 2021NTSS60), China Postdoctoral Science Foundation (Q.W., grant number 2021M690443), National Natural Science Foundation of China (L.Y., grant number 31871116), Beijing Natural Science Foundation (L.Y., grant number S170003), and Guangdong Key Project in “Development of new tools for diagnosis and treatment of Autism” (L.Y., grant number 2018B030335001).
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
