*3.1. Facial Information*

#### 3.1.1. Facial Movements

For facial movements, researchers tended to focus on emotional facial information, following the idea that impaired social communication in ASD can be framed as deficits in emotional communication regarding both perception and expression [25], and the finding that people with ASD may show reduced or idiosyncratic emotional expressions [6,26]. Against this context, it may be useful to keep in mind that ASD-specific impairments in face perception are not restricted to emotional expressions, but also affect other aspects such as facial identity [27]. Similarly, expression in social communication may be affected in subtle ways that go beyond emotional expressions [28]. The original articles on sensor-based assessment in terms of identification of autism spectrum disorders (ASD)-related features, regarding facial movements and other forms, represented in this review are listed in Table 1.

Samad et al. [29] compared facial expressions from an ASD and a typically developing (TD) group, with 8 participants each. Computing facial curvatures from 3D point cloud data, they found equally intense but more asymmetrical facial expressions in the ASD group. Another study [30] used a single webcam mounted on a TV screen to record toddlers' spontaneous facial expressions when confronted with emotional cartoons. Comparing descriptive data from small groups with five children each, they reported that the lower face seemed to be more important in distinguishing between the ASD and TD groups. Leo et al. [31] presented a new processing pipeline on 2D video data that aimed at assessing facial expressions in ASD children specifically. They estimated the production skills for individual children based on verbal instructions to express these emotions, separately for individual face parts and emotions. The performance of 17 boys with ASD was variable, and some boys exhibited differential production scores for different categories of emotions. Note that this finding is in broad agreement with theories that emphasize category-specific mechanisms, and with componential approaches to emotion [32,33]. Samad et al. [34] used a storytelling avatar, also with the aim of eliciting spontaneous emotions to differentiate between ASD and TD groups with 10 participants each. Comparing 3D data on the level of facial action units (AUs) they found overall lower AU activations in the ASD group and lower correlations between the AUs. The deviant activation in AUs 6, 12, and 15 (cheek raiser, lip corner puller, and lip corner depressor, respectively), were found to be promising markers for ASD. As a more serious screening approach, the App 'Autism & Beyond' [35] recorded toddlers' facial responses to certain stimuli with the front camera and classified them into positive, neutral, or negative. Using a large database of 1756 children, the associated facial expressions, along with eye gaze and parental reports, with Autism Spectrum risk status. Comparisons between high- and low-risk groups

revealed that high risk for ASD was associated with higher frequencies of neutral facial expressions, and with lower frequencies of positive expressions.

#### 3.1.2. Eye Gaze

One major domain in which autistic people are frequently described to behave unusually relates to oculomotor behavior, including low levels of eye contact during communication, and low levels of directional signaling via eye gaze. In fact, a current target article challenges the common belief that autistic people lack social interest in others, and suggests that this interpretation could be erroneously elicited by unusual behavior such as low levels of eye contact [36]. While this may be a useful context to keep in mind when reading this section, there is agreement that the assessment of eye gaze can help to identify behavioral patterns that are relevant for autism [37].

Chawarska and Shic [38] compared toddlers with ASD and TD of different age groups (2 and 4 years old) with an eye-tracking system while they watched neutral faces. Although all groups spent equal time looking at the screen, a restricted scanning pattern in the ASD group was found, with relative neglect of the mouth area. Additionally, the older ASD group spent less time looking at inner facial features in general, even compared to the younger ASD group. According to the authors, this indicates a different scanning pattern of children with ASD that emerges throughout early childhood and suggests less looking at the mouth as the best early predictor for ASD. Liu et al. [39] used a machine-learning framework on eye-tracking data of 29 ASD vs. 29 TD children (mean age 7.90 years) looking at images of faces embedded in a learning task with repeated presentations. Their proposed framework was able to identify ASD children both from age- (mean age 7.86 years) and IQ-matched (mean age 5.74 years) control groups with high classification accuracy (up to 88.5%). Król and Król [40] used eye-tracking to study the effect of including temporal information into spatial eye-tracking of face stimuli, thus creating scan paths for 21 ASD and 23 TD individuals (mean age around 16 years). They found a difference in face-scanning not only in spatial properties but also in temporal aspects of eye gaze, even within short exposures (about 2 s) to a facial image. Classification of group membership based on a machine-learning algorithm on spatial and temporal data led to better accuracy (55.5%) than classification based on spatial data alone (53.9%), although overall accuracy was rather low. Note that discrepancies between classification accuracies in this study and the study by Liu et al. [39] need to be seen in the context of specific conditions of each study, and could potentially be attributed to the facial stimuli and trial numbers used, the experimental task instructions, methodological differences in data analysis, or differences in the samples tested.

In a more general study on the visual scanning of natural scene images, Wang et al. [41] compared an ASD and a TD group (*N* = 20 and 19, respectively) at different levels of perception. The ASD group was found to fixate more towards the center of an image (pixel level), less on objects in general (object level) and less on certain objects (e.g., faces or objects indicated by social gaze), but more on manipulatable objects (semantic level).


OriginalArticlesonSensor-BasedAssessmentinTermsofIdentificationofautismspectrumdisorders(ASD)-related





or did not specify exact age.

Overall, facial data suggest that facial movements, either in response to emotional stimuli or as an imitation of seen facial expressions, comprise relevant features that may be markers for autism. Recordings of spontaneous emotional responses could be of particular benefit when assessing small children or nonverbal people, although the importance of age-appropriate task instructions should be considered. Across several studies, eye gaze data reveal different scanning patterns for people with ASD, particularly when viewing faces. However, we believe that further systematic research into fixation patterns and scan paths for more complex natural scenes could enrich our current understanding with additional insights. Similarly, the observation of reduced eye contact/mutual gaze in ASD, reviewed in Jaiswal et al. [50], points at a technologically challenging but theoretically relevant field of future investigations using sensor technology.
