**1. Introduction**

Throughout the last decades, the number of people diagnosed with an Autism Spectrum Disorder (ASD) increased dramatically [1,2] and so did the need for high-quality diagnostic protocols and therapies. With the ongoing progress in computer sciences and hardware, a lot of creative ideas emerged on how to use sensor data to identify and observe autistic markers, support diagnostic procedures and enhance specific therapies to improve individuals' outcomes.

ASD is a behaviorally defined group of neurodevelopmental disorders that are specified by impaired reciprocal social communication and restricted, repetitive patterns of behavior or activities (DSM-5), [3]. The symptoms are usually apparent from early childhood and tend to persist throughout life [4]. Common social impairments include a lack of social attention as evident in abnormal eye gaze or eye contact [5] and social reciprocity such as in reduced sharing of emotions in facial [6] or vocal behavior [7]. Further, only a minority of the affected people report having mutual friendships [8]. Related to the restricted and repetitive behaviors, stereotyped motor movements and speech are the stand-out features in many people with ASD [9]. Other symptoms are insisting on sameness and

routines [10], special interests and hyper- or hyporeactivity to sensory input from various modalities [11]. The exact profile and severity of symptoms in people with ASD as well as their personal strengths and coping capabilities vary to a great degree, and so does their need for support.

Reasons for the increased prevalence over the last decades include a more formalized diagnostic approach and heightened awareness. The current 'gold standard' for a diagnosis of ASD consists of an assessment of current behavior, a biographical anamnesis, and a parental report, all collected and evaluated by a trained multi-professional team [12]. Although the screening and diagnostic methods for ASD improved throughout the last years, many affected people, especially women [13] and high functioning people, still receive a late diagnosis. Since early interventions have been shown to be most effective for improving adaptive behavior, as well as IQ and language skills [14], there is continued demand for methods promoting early assessment in order to avoid follow-up problems. In this context, progress in automatic and sensor-assisted identification of ASD-specific behavioral patterns could make an important contribution to an earlier and less biased diagnosis.

Even beyond assessment, advances in digital technology are highly relevant for autism, and in more than just one way. First, there is some evidence that many autistic people show behavioral tendencies to interact with technology and to potentially prefer such interactions to interactions with humans. It is thought that autistic traits are related to systemizing, the drive to analyze how systems work, as well as to predict, control and construct systems [15]. In this context, high information technology (IT) employment rates are often used as a proxy for higher rates of strong systemizers in a population. Intriguingly, recent research from the Netherlands reported that the prevalence of childhood autism diagnoses, but not of two control neurodevelopmental diagnoses (i.e., ADHD and dyspraxia), was substantially higher in Eindhoven, a classical (IT) region, when compared to two control regions (Utrecht and Haarlem) that had been selected for high demographic and socio-economic similarity in criteria other than the proportion of IT-related jobs [16]. Second, and qualifying any simplistic interpretation of this correlative (but not necessarily causal) relationship, there is evidence that technology can potentially provide powerful social support for children with autism. For instance, children with ASD often perform better with a social robot than a human partner (e.g., in terms of enhanced levels of social behavior towards robots), tend to perceive interactions with robots as positive [17], and subsequently show reduced levels of repetitive or stereotyped actions. For a recent review, see Pennisi et al. [18].

Scientific interest in the utilization of sensor technology to gain an understanding of people with ASD has increased considerably in the recent past. Some fields of research focus on different neurobiological assessments and try to identify autism-specific signals or 'biomarkers' to better understand the neurobiological underpinnings of the disorder. Good overviews covering methods including electroencephalography (EEG), magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI) are provided by Billeci et al. [19] and by Marco et al. [20]. Other research focused on an autonomic activity such as heart rate variability (HRV) or skin conductance responses (SCR). These can be studied with Wearable devices, typically in the context of emotional monitoring in ASD as seen in a review by Taj-Eldin, et al. [21]. Applications in VR environments [22–24] have also been reviewed as promising methods to train and practice social skills.

The aim of this review is to provide an overview of the current state of research using sensor-based technology in the context of ASD. We focus on sensor technology that is applicable without constraining natural movement, and on sensory input from the face, the voice, or body movements. Accordingly, this review does not consider evidence from wearable technology, VR, or psychophysiological and neurophysiological recordings. Note also that while we provide details of our procedure for identifying relevant original findings to enhance reproducibility, this paper represents a thematic review in which we pre-selected for contents as described below, and in which we occasionally discuss additional relevant findings that were not formally identified by this literature search. For instance, we may refer to some key findings regarding psychological theories of human social or emotional communication where relevant, even when the findings were not obtained with individuals with autism.
