Autism Digital Phenotyping in Preschool- and School-Age Children.

Autism Digital Phenotyping in Preschool- and School-Age Children - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Autism Res . Author manuscript; available in PMC: 2026 Jun 1. Published in final edited form as: Autism Res. 2025 Apr 2;18(6):1217–1233. doi: 10.1002/aur.70032 Autism Digital Phenotyping in Preschool- and School-Age Children Vikram Aikat 1, # , Kimberly LH Carpenter 2, 3, # , Pradeep Raj Krishnappa Babu 4 , J Matias Di Martino 4, 5 , Steven Espinosa 6 , Scott Compton 2, 3 , Naomi Davis 2, 3 , Lauren Franz 2, 3, 7 , Marina Spanos 2 , Guillermo Sapiro 8, * , Geraldine Dawson 2, 3, * Author information Article notes Copyright and License information 1 Department of Computer Science, Duke University, Durham, NC, USA; 2 Department of Psychiatry and Behavioral Sciences, Duke University, Durham, NC, USA; 3 Duke Center for Autism and Brain Development, Duke University, Durham, NC, USA; 4 Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA; 5 Department of Informatics and Computer Science, Universidad Catolica del Uruguay, Montevideo, Uruguay; 6 Office of Information Technology, Duke University, Durham, NC, USA; 7 Duke Global Health Institute, Duke University, Durham, NC, USA; 8 Department of Electrical and Computer Engineering, Princeton University, Princeton, NJ, USA. # Co-First authors, equal contribution. * Co-Senior authors. ✉ Corresponding Author: Vikram Aikat: vikram.aikat@duke.edu Issue date 2025 Jun. PMC Copyright notice PMCID: PMC12167152  NIHMSID: NIHMS2069551  PMID: 40176346 The publisher's version of this article is available at Autism Res Abstract There is a critical need for scalable and objective tools for autism screening and outcome monitoring, which can be used alongside traditional caregiver and clinical measures. To address this need, we developed SenseToKnow, a tablet- or smartphone-based digital phenotyping application (app), which uses computer vision and touch data to measure several autism-related behavioral features, such as social attention, facial and head movements, and visual-motor skills. Our previous work demonstrated that the SenseToKnow app can accurately detect and quantify behavioral signs of autism in 18–40-month-old toddlers. In the present study, we administered the SenseToKnow app on an iPad to 149 preschool- and school-age children (45 neurotypical and 104 autistic) between 3 and 8 years of age. Results revealed significant group differences between autistic and neurotypical children in terms of several behavioral features, which remained after controlling for sex and age. Repeat administration with a subgroup demonstrated stability in the individual digital phenotypes. Examining correlations between the Vineland Adaptive Behavior Scales and individual digital phenotypes, we found that autistic children with higher levels of communication, daily living, socialization, motor, and adaptive skills exhibited higher levels of social attention and coordinated gaze with speech, less frequent head movements, higher complexity of facial movements, higher overall attention, lower blink rates, and higher visual motor skills, demonstrating convergent validity between app features and clinical measures. App features were also significantly correlated with ratings on the Social Responsiveness Scale. These results suggest that the SenseToKnow app can be used as an accessible, scalable, and objective digital tool to measure autism-related behaviors in preschool- and school-age children. Keywords: Autism, Preschool- and School-Age, Digital Phenotyping, Computer Vision Lay Summary Previous research has shown that an application (app), SenseToKnow, administered on a smartphone or tablet, can be used to accurately detect and measure several early autism behavioral features in toddlers. The present study demonstrated that the SenseToKnow app detected significant group differences in behavioral characteristics between autistic and neurotypical children at preschool- and school-age. The app uses digital technology derived from computer vision and touch data to measure autism-related behaviors in a scalable, objective way. This work provides a foundation for future work on autism screening and outcome monitoring through digital phenotyping in preschool- and school-age children. Introduction Screening, diagnosis, and behavioral assessments of autism continue to rely on subjective clinical observation and caregiver report measures ( Dawson & Sapiro, 2019 ; McPartland, 2016 ). There is a critical need for scalable, quantitative, and objective tools for autism screening, assessment, and outcome monitoring to complement clinician and caregiver report measures ( McCracken et al., 2021 ; Uddin et al., 2017 ; Walsh et al., 2011 ). To meet this need, existing studies have employed various approaches to automatically quantify and analyze various autistic behaviors using modalities such as audio, video, and touch, assessed via computers or tablets ( Mukherjee et al., 2024 ). Methods for quantifying autism-related behaviors have included computer vision ( de Belen et al., 2020 ), language and voice acoustic features ( Fusaroli et al., 2017 ; MacFarlane et al., 2022 ), and eye-tracking (Wei et al., 2023), among others. Objective, quantitative assessments can be used alongside traditional caregiver report and clinical rating methods; this was demonstrated in a study of 18–72-month-old children that combined measures from various questionnaires (depending on the child’s age) along with video data to predict autism ( Kanne et al., 2018 ). While many of these studies of digital assessment approaches show promise, their limitations include focusing on a limited subset of behaviors, requiring overly lengthy assessments, and utilizing specialized equipment, such as depth cameras for motion tracking (Tunçgenç et al., 2019) or eye trackers for gaze monitoring ( Shic et al., 2022 ). For example, in their review of eye-tracking approaches for autism screening, Wei et al. found that while these methods show promise, particularly in preschool-age children, the significant variation in eye-tracking technology poses challenges for scalability (Wei et al., 2023). Given these challenges, it is essential to develop digital phenotyping technologies for autism that are both accessible and scalable ( Lidstone et al., 2021 ), while also capturing multiple behavioral features to comprehensively assess various autism-related behaviors. In this effort, we have previously demonstrated that digital phenotyping via computer vision conducted without any specialized hardware and using only the frontal camera and touchscreen of a tablet or smartphone can objectively detect and quantify multiple behavioral signs of autism in toddlers ( Perochon et al., 2023a ; Krishnappa Babu et al., 2024b ). This was achieved via an app (SenseToKnow) that displays brief, strategically designed movies on a tablet or smartphone while the device’s frontal camera records the child’s behavioral responses, which are then objectively quantified via automatic computer vision analysis (CVA). The app also includes a brief bubble popping game designed to assess visual motor skills detected via the device’s touchscreen-based inertia sensor. The SenseToKnow app has been shown to elicit a wide range of autism-related behaviors, including gaze ( Chang et al., 2021 ), head movements ( Dawson et al., 2018 ), facial expression and complexity ( Carpenter et al., 2021 , Krishnappa Babu et al., 2023b ; Krishnappa Babu et al., 2023c ), response to name ( Campbell et al., 2019 , Hashemi et al., 2021 , Perochon et al., 2021 ), blink rate ( Krishnappa Babu et al., 2023a ), and visual-motor skills ( Perochon et al., 2023b ) in toddlers with and without autism. The SenseToKnow app can identify autism in 18–36-month-old toddlers with a high diagnostic accuracy when delivered in primary care ( Perochon et al., 2023a ) or remotely by caregivers ( Krishnappa Babu et al., 2024a , Krishnappa Babu et al., 2024b ). In toddlers, digital phenotypes derived from the app have been found to correlate with standardized clinical measures ( Coffman et al., 2023 ). These prior studies provided evidence for the utility of using the SenseToKnow app to detect autism-related behaviors in toddlers. The present work expanded the age range into early childhood as a first step toward assessing whether the app could potentially be used for autism screening and assessment in older children, which is important given that the average age of autism diagnosis remains close to 5 years, and even older in low- and middle-income countries ( van ‘t Hof et al., 2021 ). There is also a need for objective, quantitative outcome measures for use in clinical trials for autism, including trials conducted with preschool- and school-age children. To address these gaps, the current study had three goals. First, we assessed whether the SenseToKnow app can detect and quantify autism-related behaviors in groups of 3- to 8-year-old children with and without a diagnosis of autism. Second, we examined the test-retest stability of the behaviors detected via the app. Third, we examined the convergent validity of app features for the autistic sample with standardized clinical measures by analyzing the correlation between SenseToKnow app digital features with both the Vineland Adaptive Behavior Scales ( Sparrow et al., 2016 ), a commonly used, standardized, caregiver-report measure of communication, daily living, socialization, motor, and adaptive skills, and the Social Responsiveness Scale ( Constantino, 2012 ), a measure of the severity of autism-related social impairment. Methods Participants Participants were 149 children (45 neurotypical and 104 autistic) between 3 and 8 years of age at the time of app administration. Of these 149 children, 135 (41 neurotypical and 94 autistic) children had valid app data determined via an automated validity algorithm based on the proportion of frames the child is facing the screen, their distance from the camera, and their head pose relative to the camera ( Krishnappa Babu et al., 2024a ). There was no difference in the number of children with valid data between the two groups. The mean age of the study sample was 71.4 months (standard deviation 15.9 months, range 40.2–102.6 months), including 98 boys and 37 girls. The demographic characteristics of the two groups of children, including age, sex, race, ethnicity, caregiver education level, and developmental quotient (DQ) measured via standardized measures of intellectual ability are shown in Table 1 . Table 1. Study sample demographic characteristics Demographic Variable Total Sample (N=135; 100%) Neurotypical (N=41; 30.37%) Autism (N= 94; 69.63%) Age in Months Mean (SD) 71.44 (15.91) 67.77 (15.73) 73.05 (15.72) Developmental Quotient * Mean (SD) 94.95 (24.51) 116.05 (12.59) 85.74 (22.69) Sex Boys (%) 98 (72.59%) 29 (70.73%) 69 (73.40%) Girls (%) 37 (27.41%) 12 (29.27%) 25 (26.60%) Race White/Caucasian (%) 103 (76.30%) 35 (85.37%) 68 (72.34%) Black or African American (%) 7 (5.19%) 1 (2.44%) 6 (6.38%) Asian (%) 8 (5.93%) 1 (2.44%) 7 (7.45%) More Than One Race (%) 13 (9.63%) 3 (7.32%) 10 (10.64%) Other (%) 4 (2.96%) 1 (2.44%) 3 (3.19%) Ethnicity Hispanic\Latino (%) 14 (10.37%) 3 (7.32%) 11 (11.70%) Not Hispanic/Latino (%) 121 (89.63%) 38 (92.68%) 83 (88.30%) Caregivers’ Education Level High School Diploma (%) 4 (2.96%) 1 (2.44%) 3 (3.19%) Some College Education (%) 15 (11.11%) 1 (2.44%) 14 (14.89%) 4-Year College Degree or More (%) 116 (85.93%) 39 (95.12%) 77 (81.91%) Open in a new tab * p<.05 for autism versus neurotypical comparison ANOVA and Chi-square tests were used for numerical and categorical comparisons, respectively. Children were recruited from the community through the research center’s volunteer research registry, flyers and brochures, emails, and social media posts. The study design and methods were approved by Duke University School of Medicine Institutional Review Board (IRB-Pro00085156). The parent or legal guardian signed informed consent forms approved by the Duke Institutional Review Board. Children over the age of 6 years provided verbal assent if they were capable of doing so. For all groups, inclusion criteria included being between 3 and 8 years of age at the initial eligibility visit and, for children on stimulant medication, being willing to undergo a 24-hour washout. Exclusion criteria included any known genetic or neurological syndrome or condition with established links to autism (e.g., Fragile X). Children with genetic findings in which the link to autism is less well known/established (e.g., 16p11.2 CNVs) were not excluded from the study. Children with a history of epilepsy or seizure disorders were also excluded, except for those with a history of simple febrile seizures or if the child had been seizure-free for the 12 months prior to their study visits. Children with motor or sensory impairment that would interfere with the valid completion of study measures, including significant hearing or vision impairment not correctable by a hearing aid or glasses/contact lenses, were not eligible for the study. Other exclusionary criteria included the history of neonatal brain damage (e.g., with a diagnosed hypoxic or ischemic event) and any known environmental circumstance that likely accounted for the diagnosis of autism (e.g., severe nutritional or psychological deprivation). For the autistic children, additional inclusion criteria included a DSM-5 diagnosis of autism spectrum disorder. Prior to the Covid-19 pandemic, the diagnosis was based on the Autism Diagnostic Interview-Revised (ADI-R; Lord et al., 1994 ) and the Autism Diagnostic Observation Schedule-2 (ADOS-2; Lord et al., 2012 ). Following the resumption of research during the Covid-19 pandemic, the ADOS-2 (N=85) was replaced with the Brief Observation of Symptoms of Autism (BOSA (N=15); Dow et al., 2022 ) which was administered by clinicians who were research reliable on the ADOS-2. Inclusion criteria in the neurotypical group included having cognitive abilities in the average range (IQ>70), Vineland Adaptive Behavior Scale-3 (VABS-3; Sparrow et al., 2016 ) scores in the average range, and no clinical elevations on any of the following: the Child Behavior Checklist (except for on the anxiety subscale; Achenbach & Ruffle, 2000 ), the Attention-Deficit/Hyperactivity Disorder (ADHD)-Rating Scale (ADHD-RS; DuPaul et al., 1998a), or the Social Responsiveness Scale 2nd Edition (SRS-2; Constantino, 2012 ). Children were excluded from the neurotypical group if they had a sibling or first-degree relative diagnosed with either autism spectrum disorder or attention-deficit/hyperactivity disorder. Developmental Quotient (DQ) was assessed using the Differential Ability Scales, 2nd Edition (DAS-II; Elliott, 1990 ) prior to the COVID-19 pandemic in March 2020. After March 2020, in keeping with social distancing measures, we modified the protocol to allow for validated, tablet-based cognitive measures. Children between 36 and 89 months of age completed the Wechsler Preschool and Primary Scale of Intelligence-Fourth Ed. (WPPSI-IV; Wechsler, 2012 ), and children 90 months and older completed the Wechsler Intelligence Scale for Children-Fifth Ed. (WISC-V; Wechsler, 2015 ). Although the cognitive measures were delivered via a tablet, they were still administered in the clinic under the supervision of one of the research clinicians. To ensure that cognitive ability was standardized across these assessments, we calculated a developmental quotient (DQ) for each child by taking the average of all age equivalents from the associated measures divided by the chronological age of the child and multiplied by 100. SenseToKnow app administration and CVA and inertia features The app was administered on an iPad positioned with a stand on a table. Children watched a series of brief dynamic movies and participated in a bubble-popping game ( Figure 1 ). The movies were created in both English and Spanish to accommodate Spanish-speaking children. Actors in the movies were of diverse ethnic/racial backgrounds. Right and left placement of the social and nonsocial stimuli were counter-balanced across movies. The app administration took less than 10 minutes. Figure 1. Open in a new tab Movies shown in the SenseToKnow app, presented in order of appearance. Using the frontal camera of the iPad, the child’s body and facial movements were recorded at 1920×1080 spatial resolution and at 30 frames per second. Facial bounding boxes for the child were automatically detected using InsightFace ( Deng et al., 2022 ) and manually verified using human verification. For each frame containing the child’s face, CVA features were extracted. All head movement, facial complexity, and facing forward features were separately computed for social and nonsocial stimuli. A summary of CVA features can be found in Table 2 . Table 2. SenseToKnow app dependent variables, associated stimuli, and previous references Dependent Variables Associated Stimuli a Description Primary Reference(s) Response to Name 1, 3 Proportion Proportion of name calls in which the child’s head turns Campbell, et al. (2019) Delay Average delay between the name call and the child’s head turn Perochon, et al. (2021) Gaze Measures Gaze Silhouette Score 2, 4 Measure of the overlap and distinctive nature of gaze clusters. High silhouette scores correspond to well defined and distinctive clusters, whereas scores closer to zero signify weak clustering results. Chang, et al. (2021) Gaze at Social Stimuli 2, 4 Percent of the video the child is looking at the social stimuli Gaze Speech Correlation 9 Correlation between the child’s gaze patterns and the actor’s speech patterns Head Movement Average Movement, Social and Nonsocial 2–9 Average rate of head movement to social versus non-social videos Dawson, et al. (2018) Acceleration, Social and Nonsocial 2–9 Measure of the change in head movement rate (i.e. acceleration) Krishnappa Babu, et al. (2023b) Head Movement and Facial Complexity Complexity, Social and Nonsocial for Head, Eyebrows and Mouth 2–9 Level of predictability and stability of head movements. This was done for the full head, as well as for the mouth and the eyebrows separately. Krishnappa Babu, et al. (2023c) Head Position Chang, et al. (2021) ; Perochon, et al. (2023b) Facing Forward, Social and Nonsocial 2–9 Forward facing head position, which served as a proxy for attention to stimuli was calculated using the facial landmarks detected using InsightFace and included times during the stimulus presentation in which the child’s eyes were open, their gaze was directed towards or near the screen, and their face remained relatively stable Blink Krishnappa Babu, et al. (2023a) Blink Rate, Social and Nonsocial 2–9 Average blink rate during social versus non-social movies Open in a new tab a Associated Movie Stimuli: 1 Floating Bubbles, 2 Spinning Top, 3 Mechanical Puppy, 4 Blowing Bubbles, 5 Rhymes and Toys, 6 Make Me Laugh, 7 Dog in Grass, 8 Playing with Blocks, 9 Fun at the Park Response to name During the app administration, at three pre-specified times, an examiner positioned behind the child called the child’s name. From this, we computed an average delay in turning the head in response to the name call and the proportion of name calls to which the child turned their head. See Perochon et al. (2021) for more details. Gaze Using the child’s head pose and facial landmarks, along with a deep neural network, we computed the proportion of time the child gazed at social stimuli in strategically-designed movies that displayed social stimuli on one side of the screen and nonsocial stimuli (toys) on the opposite side of the screen (gaze at social stimuli), as well as a gaze silhouette score, a metric of how well their gaze patterns clustered during this movie. Additionally, we calculated the correlation between the child’s gaze and the occurrence of a person speaking in a movie in which two people were having a conversation (gaze speech correlation). See Chang et al. (2021) for more details. Head movements Across all movies, using head and face landmarks, we computed the rate of head movements and the derivative of this metric to measure the head movement acceleration. We then estimated the complexity of the head movement rate at multiple timescales using multiscale entropy. See Krishnappa et al. (2023b) for more details. Facial complexity Across all movies, we used the same complexity analysis performed on head movement rate, analyzing the movements of the eyebrows and mouth landmarks to compute eyebrows complexity and mouth complexity. See Krishnappa et al. (2023c) for more details. Facing the screen and blink rate, two measures of attentional engagement Using head and face landmarks, we computed the roll, pitch, and yaw for each child and calculated overall attention, i.e., the proportion of time they spent facing the screen, calculated separately for movies with high versus low social content. In addition, using OpenFace ( Baltrusaitis et al., 2018 ), we computed facial action units and extracted blinking events, computing the rate of blinking during movies with high social versus low (nonsocial) content. Studies have found that lower blink rates have been associated with higher attentional engagement ( Ranti et al., 2020 ). See Krishnappa et al. (2023a) for more details. In order to better isolate the effect of each stimulus on the participant’s behavioral responses, all features such as gaze, head movement, facial complexity, and blink rate were computed only during frames in which the child is facing towards the screen. Bubble popping game Following the movies, the children played a short bubble popping game in which floating bubbles containing marine animals floated across the screen. The child popped the bubbles while the iPad touchscreen recorded the locations touched as well as the force applied during each touch. Dependent inertia features from the bubble popping game were the same as defined in previous work ( Perochon et al., 2023b ) and included touch popping ratio, touch average applied force and length, and touchscreen exploratory percentage. The touch popping ratio, which is a measure of the child’s accuracy and precision, was calculated by dividing the total number of popped bubbles over the total number of touches. The touch average applied force uses the kinetics (inertia sensors) information from the device to find the average force of the child’s touches on the screen during the game, while the touch average length feature measured the average length for each touch motion. Finally, the touchscreen exploratory percentage feature was defined as the proportion of the device screen that was touched during the entirety of the game. Test-retest stability across two administrations To assess test-retest stability of the digital phenotypes derived by the app, 74 participants (11 neurotypical and 63 autistic) who were part of the 135 participants in the larger study agreed to participate in an additional administration of the app, which was conducted an average of 10 days after the first administration. Seven of those called back had an invalid second administration, resulting in 67 participants (10 neurotypical and 57 autistic) with two valid administrations, which were used in the test-retest analysis. Only half of the sample completed a second app administration due to changes implemented in the study protocol when we returned to in-person research during the COVID-19 pandemic, which minimized the time participants and staff were in direct contact with one another. These changes led to some participants only completing a subset of the original study tasks, including only completing a single app administration. The participants who participated in the second administration did not differ in demographic characteristics, including sex, age, race, ethnicity and caregiver’s education level, compared to the larger sample of participants in the first administration, reducing any sample potential bias in our test-retest analysis. Statistical analyses For each app feature, we applied the Mann-Whitney-U Test to evaluate the statistical significance between the distributions of the feature for the autistic group versus the neurotypical group. Additionally, to understand the relationship between each feature and group, we used separate analyses of covariance (ANCOVA) with age, sex, and DQ as covariates. We utilized the Mann-Whitney-U Test, paired t-test and Bland-Altman analysis to assess the test-retest stability of repeated administrations in our study. The proportion of agreement between repeated administrations was found using Bland-Altman analysis. From the standard deviation of differences between each feature in repeated administrations, we calculated an upper and lower limit of agreement, where points within the limit display consistent measurement of that feature. Finally, for each app feature, we computed the Spearman’s correlation with each Vineland-3 domain standard score, including communication, daily living skills, socialization, and fine motor skills, as well as the summary metric of adaptive behavior composite (ABC) and with the Social Responsiveness Scale (SRS) scaled scores in the subdomains of Social Awareness, Restricted and Repetitive Behaviors, and the overall SRS score. We chose the Social Awareness subdomain because it is designed to measure the child’s sensitivity to social cues (e.g. social gaze, attention to speech and facial expressions) which are displayed in the SenseToKnow movies, and the Restricted Interests/Repetitive Behaviors subdomain because of its potential relationship with the frequency of the child’s head movements, which we have previously hypothesized to reflect a sensory regulatory function. Results Group differences in SenseToKnow features Results of the Mann-Whitney-U Test comparing autistic versus neurotypical children for each app feature are shown in Table 3 , along with mean values of each feature for both diagnostic groups. We found significant group differences between the autistic and neurotypical group for the majority of digital phenotypes, with the exception of the touch variables from the bubble popping game. Table 3. Results of Mann-Whitney U Test and mean feature value for each SenseToKnow feature comparing autistic (N=94) and neurotypical (N=41) groups Digital Phenotype SenseToKnow Feature AUT Mean NT Mean p-value Effect Size Response to Name Response to name delay (seconds) 1.13 1.17 0.684 0.06 Proportion of name call responses (%) 55.9 37.0 0.0127 0.262 Gaze Gaze at social stimuli (%) 46.2 52.3 <.0001 0.382 Gaze speech correlation .333 .461 <.005 0.342 Gaze silhouette score .481 .568 <.00005 0.475 Head Movements Rate of head movements during social movies 1.78 .851 <.005 0.351 Rate of head movements during nonsocial movies 2.27 .870 <.00005 0.471 Rate of head movements during social movies derivative .0114 .00452 <.001 0.367 Rate of head movements during nonsocial movies derivative .0175 .00575 <.000005 0.515 Head movements complexity during social movies 16.1 23.0 <.0001 0.423 Head movements complexity during nonsocial movies 7.4 11.9 <.000005 0.497 Facial Complexity Eyebrow complexity during social movies 27.5 29.6 0.22 0.133 Eyebrow complexity during nonsocial movies 22.9 28.1 <.0005 0.378 Mouth complexity during social movies 26.3 29.4 0.0175 0.258 Mouth complexity during nonsocial movies 22.6 27.8 <.0005 0.407 Facing Forward Time spent facing toward screen during social movies (%) 83.7 93.7 <.0001 0.423 Time spent facing toward screen during nonsocial movies (%) 81.7 92.6 <.000005 0.507 Blink Blink rate during social movies .00776 .00436 <.005 0.329 Blink rate during nonsocial movies .00799 .00439 <.005 0.355 Touch Touch popping ratio (%) 92.1 95.2 0.0945 0.195 Touchscreen exploratory percentage (%) 70.3 70.2 0.947 0.008 Touch average applied force .239 .107 0.103 0.197 Touch average length (mm) 3.73 2.10 0.0912 0.203 Open in a new tab Contrary to our earlier findings with younger children, compared to the neurotypical children, the autistic children had a higher proportion of name calls to which they turned their head (p<.05). However, numerous other features were consistent with our findings with younger children; specifically, autistic children spent less time gazing at social stimuli (p<.0001), displayed a more diffuse gaze pattern (silhouette score; p<.00005), and exhibited lower correlations between their gaze and the actors’ speech (p<.005). Furthermore, compared to neurotypical children, autistic children displayed higher rates of head movements during social (p<.005) and nonsocial (p<.00005) movies and lower complexity (greater predictability) of their head movements during social (p<.001) and nonsocial movies (p<.000005). Autistic children also showed significantly lower complexity of their eyebrow and mouth movements during all nonsocial movies (p< .0005). We examined two measures of overall attention: facing forward to the screen and blink rate (lower blink rate = higher engagement). Autistic children exhibited significantly less time facing the screen for both social (p<.0001) and nonsocial (p<.000005) movies and higher blink rates during social (p<.005) and nonsocial movies (p<.005). Examining the inertia features derived from the bubble-popping game revealed that these features did not discriminate between the autistic and neurotypical groups. Group differences with age, sex, and DQ as covariates Age. We used ANCOVA to examine differences in each SenseToKnow feature between diagnosis groups with age as a covariate. The results are shown in Table 4 . Similar to analyses conducted without age as a covariate, for every feature except touch (derived from the bubble-popping game) and response to name delay, significant group differences remained. For touch related features, we found significant group differences in the touch average applied force with age as a covariate, a feature that did not show significant group differences when age was not considered in the analysis. Linear regression models illustrating the relationship between age and each behavioral feature are shown in the Supplementary Materials ( Figure S1 ), as well as correlation (Spearman’s) of each feature with age and DQ for the neurotypical group ( Table S1 ) and the autistic group ( Table S2 ). Table 4. Results of ANCOVA analysis with age as a covariate for each SenseToKnow feature comparing autistic (N=94) and neurotypical (N=41) groups. Digital Phenotype SenseToKnow Feature p-value Response to Name Response to name delay 0.780 Gaze Gaze at social stimuli <.0005 Gaze speech correlation <.0005 Gaze silhouette score <.00001 Head Movements Rate of head movements during social movies <.001 Rate of head movements during nonsocial movies <.0005 Rate of head movements during social movies derivative <.0005 Rate of head movements during nonsocial movies derivative <.0005 Head movements complexity during social movies <.00001 Head movements complexity during nonsocial movies <.000001 Facial Complexity Eyebrow complexity during social movies <.05 Eyebrow complexity during nonsocial movies <.00005 Mouth complexity during social movies <.001 Mouth complexity during nonsocial movies <.00005 Facing Forward Time spent facing toward screen during social movies <.0001 Time spent facing toward screen during nonsocial movies <.00001 Blink Blink rate during social movies <.005 Blink rate during nonsocial movies <.005 Touch Touch popping ratio .0625 Touchscreen exploratory percentage .622 Touch average applied force <.05 Touch average length .065 Open in a new tab Sex. Using sex as a covariate in an ANCOVA analysis, group differences remained for the majority of features, as shown in Table 5 . Except for touch average applied force, every feature that was found to be significant when comparing means using the Mann Whitney U-test was also significant when performing ANCOVA between diagnosis groups with sex as a covariate. Table 5. Results of ANCOVA analysis with sex as a covariate for each SenseToKnow feature between autistic (N=94) and neurotypical (N=41) groups. Digital Phenotype SenseToKnow Feature p-value Response to Name Response to name delay .963 Gaze Gaze at social stimuli <.0005 Gaze speech correlation <.001 Gaze silhouette score <.00001 Head Movements Rate of head movements during social movies <.005 Rate of head movements during nonsocial movies <.005 Rate of head movements during social movies derivative <.005 Rate of head movements during nonsocial movies derivative <.001 Head movements complexity during social movies <.00005 Head movements complexity during nonsocial movies <.000005 Facial Complexity Eyebrow complexity during social movies .0679 Eyebrow complexity during nonsocial movies <.0001 Mouth complexity during social movies <.01 Mouth complexity during nonsocial movies <.00005 Facing Forward Time spent facing toward screen during social movies <.0005 Time spent facing toward screen during nonsocial movies <.00005 Blink Blink rate during social movies <.005 Blink rate during nonsocial movies <.005 Touch Touch popping ratio .157 Touchscreen exploratory percentage .841 Touch average applied force <.05 Touch average length .226 Open in a new tab DQ. We found that when using DQ as a covariate, a number of features continued to distinguish between the autistic and neurotypical groups; autistic children spent a lower amount of time spent gazing at social stimuli (p<.05), exhibited more diffuse gaze patterns (gaze silhouette score; p<.005), exhibited a higher proportion of name call responses (p<.005), and had lower head movement complexity during nonsocial movies (p<.05). This is notable given that there is a significant group difference in DQ with a large effect size (p<10 −11 , r=0.75). Test-retest stability For each feature, we analyzed whether there is a significant difference between the two app administrations using the Mann-Whitney U test, paired t-test, Pearson correlation coefficient, and Bland-Altman analysis. The results are shown in Table 6 . The Mann-Whitney U test results revealed that none of the SenseToKnow features showed significantly different distributions between the first and second administrations. This is true of the overall group and when examining the autistic or neurotypical groups separately. Paired t-test results revealed that for all features, except the complexity of head movements during social stimuli (p<.05), differences between repeated administrations were not significant (p>.05). Using a Bland-Altman analysis, all CVA and inertia features have 94% or more of participants within the limits of agreement between repeated administrations. We further analyzed the stability of repeated administrations using the Pearson correlation coefficient. The majority of features, except for nonsocial head movement rate and derivative, response to name delay, and all touch features, show stability between first and second administrations. While this method is sensitive to outliers at the individual level, reduced stability for a feature may indicate that while feature distributions are not significantly different, individuals may be showing variability in a single feature between repeated administrations. Table 6. Results of Mann-Whitney U Test, paired t-test and Bland-Altman analysis for each SenseToKnow feature between first and second administration (N=67). Digital Phenotype SenseToKnow Feature Mann-Whitney U Test Paired t-Test Pearson’s r Bland-Altman Effect Size p-value t-statistic p-value r p-value Agreement Response to Name Response to name delay 0.021 0.851 −.290 .773 .015 .912 .94 Proportion of name call responses 0.048 0.612 −.853 .407 .515 <.0001 .99 Gaze Gaze at social stimuli 0.04 0.692 −1.10 .274 .285 .019 .95 Gaze speech correlation 0.097 0.333 −1.57 .121 .622 <.0001 .95 Gaze silhouette score 0.047 0.64 −0.630 .528 .398 <.001 .96 Head Movements Rate of head movements during social movies 0.084 0.403 0.050 .960 .564 <.0001 .97 Rate of head movements during nonsocial movies 0.03 0.769 −0.101 .920 .130 0.296 .96 Rate of head movements during social movies derivative 0.093 0.355 −0.357 .722 .561 <.0001 .94 Rate of head movements during nonsocial movies derivative 0.033 0.742 0.232 .817 .058 .641 .99 Head movements complexity during social movies 0.108 0.282 −2.37 <.05 .647 <.0001 .96 Head movements complexity during nonsocial movies 0.013 0.901 −0.314 .755 .604 <.0001 .97 Facial Complexity Eyebrow complexity during social movies 0.065 0.516 −1.59 .117 .494 <.0001 .94 Eyebrow complexity during nonsocial movies 0.115 0.253 −0.875 .385 .372 <.005 .96 Mouth complexity during social movies 0.067 0.507 −1.48 .143 .555 <.0001 .96 Mouth complexity during nonsocial movies 0.071 0.479 −0.498 .620 .377 <.005 .96 Facing Forward Time spent facing toward screen during social movies 0.093 0.355 −1.63 .108 .571 <.0001 .94 Time spent facing toward screen during nonsocial movies 0.059 0.56 0.649 .518 .457 <.0005 .97 Blink Blink rate during social movies 0.038 0.709 −0.233 .816 .374 <.005 .97 Blink rate during nonsocial movies 0.135 0.177 −1.36 .179 .397 <.001 .99 Touch Touch popping ratio 0.043 0.698 −1.73 .090 .275 .061 .96 Touchscreen exploratory percentage 0.051 0.653 1.42 .161 .171 .250 .98 Touch average applied force 0.035 0.768 0.064 .949 .027 .862 .94 Touch average length 0.048 0.69 1.45 .153 .063 .676 1.00 Open in a new tab Correlations between digital phenotypes and clinical measures Vineland Adaptive Behavior Scales. Correlations for autistic participants between the SenseToKnow app features and Vineland-3 clinical measures are shown in Table 7 . Across adaptive functioning domains, higher scores on subscales of the Vineland-3 were positively correlated with numerous app features, including gaze metrics, facial and head movement complexity, facing forward metrics and touch popping ratio. Additionally, app features such as blink rate, head movement rate and acceleration, and touch average length also show significant negative correlation with Vineland-3 measures. We found no significant correlations between name call related features and Vineland-3 clinical measures. Table 7. Correlation of each SenseToKnow feature with Vineland-3 clinical measures within the autistic group (N=94). Digital Phenotype SenseToKnow Feature Communication Daily Living Skills Socialization Motor Skills Adaptive Behavior Composite Response to Name Response to name delay −.065 −.04 −.098 −.047 −.073 Proportion of name call responses −.114 −.086 −.010 −.088 −.092 Gaze Gaze at social stimuli .251 *** .197 * .294 **** .245 *** .254 *** Gaze speech correlation .459 **** .426 **** .463 **** .399 **** .448 **** Gaze silhouette score .315 **** .255 *** .246 *** .253 *** .304 **** Head Movements Rate of head movements during social movies −.412 **** −.308 **** −.355 **** −.192 * −.363 **** Rate of head movements during nonsocial movies −.377 **** −.318 **** −.339 **** −.335 **** −.351 **** Rate of head movements during social movies derivative −.417 **** −.306 **** −.354 **** −.219 * −.366 **** Rate of head movements during nonsocial movies derivative −.377 **** −.318 **** −.344 **** −.337 **** −.354 **** Head movements complexity during social movies .417 **** .355 **** .357 **** .323 **** .388 **** Head movements complexity during nonsocial movies .357 **** .274 *** .283 **** .326 **** .331 **** Facial Complexity Eyebrow complexity during social movies .415 **** .348 **** .365 **** .249 *** .360 **** Eyebrow complexity during nonsocial movies .441 **** .380 **** .386 **** .362 **** .401 **** Mouth complexity during social movies .469 **** .384 **** .413 **** .293 **** .416 **** Mouth complexity during nonsocial movies .465 **** .401 **** .410 **** .368 **** .422 **** Facing Forward Time spent facing toward screen during social movies .514 **** .429 **** .479 **** .383 **** .471 **** Time spent facing toward screen during nonsocial movies .413 **** .344 **** .370 **** .324 **** .382 **** Blink Blink rate during social movies −.339 **** −.320 **** −.316 **** −.291 **** −.326 **** Blink rate during nonsocial movies −.306 **** −.288 **** −.251 *** −.300 **** −.287 **** Touch Touch popping ratio .322 **** .221 * .219 * .318 **** .257 ** Touchscreen exploratory percentage −.071 .002 −.005 −.057 −.004 Touch average applied force −.109 −.133 −.137 −.154 −.142 Touch average length −.354 **** −.263 ** −.241 * −.270 ** −.278 *** Open in a new tab Notes: *: p<.05, **: p<.01, ***: p<.005, ****: p<.001. Values in this table represent Spearman’s rank correlation coefficient (r s ). Social Responsiveness Scale. Several app features were found to be correlated with measures derived from the SRS, shown in Table 8 . Higher scores (increased traits) in the Social Awareness subdomain were associated with a lower proportion of responses to name, reduced gaze at social stimuli and reduced gaze clustering (gaze silhouette score), reduced correlation between gaze and speech, reduced head movement, eye brow, and mouth complexity, and reduced time facing toward the screen, as well as higher rates of head movements and blink rates during social movies, and longer touch average length during the bubble-popping task. Largely similar patterns were found in terms of the correlations between app features and the SRS Restricted and Repetitive Behavior subdomain scores and for the total SRS scaled scores, as shown in Table 8 . Table 8. Correlation of each SenseToKnow feature with Social Responsiveness Scale (SRS) clinical measures within the autistic group (N=94). Digital Phenotype SenseToKnow Feature Social Awareness Total Scaled Score Restricted and Repetitive Behaviors Total Scaled Score Overall Total Scaled Score Response to Name Response to name delay .19 −.05 .14 Proportion of name call responses −.22 * .15 −.14 Gaze Gaze at social stimuli −.37 *** −.11 −.35 *** Gaze speech correlation −.52 **** −.38 *** −.58 **** Gaze silhouette score −.14 −.08 −.15 Head Movements Rate of head movements during social movies .31 ** .09 .29 ** Rate of head movements during nonsocial movies .19 .24 * .25 * Rate of head movements during social movies derivative .29 ** .11 .29 ** Rate of head movements during nonsocial movies derivative .12 .22 * .18 Head movements complexity during social movies −.32 ** −.19 −.33 *** Head movements complexity during nonsocial movies −.16 −.14 −.18 Facial Complexity Eyebrow complexity during social movies −.37 *** −.32 ** −.43 **** Eyebrow complexity during nonsocial movies −.22 * −.29 ** −.29 ** Mouth complexity during social movies −.35 *** −.30 ** −.40 **** Mouth complexity during nonsocial movies −.20 −.32 ** −.28 ** Facing Forward Time spent facing toward screen during social movies −.42 **** −.32 ** −.47 **** Time spent facing toward screen during nonsocial movies −.22 * −.24 * −.27 ** Blink Blink rate during social movies .24 * .15 .26 * Blink rate during nonsocial movies .18 .12 .20 Touch Touch popping ratio −.21 −.04 −.20 Touchscreen exploratory percentage −.11 −.02 −.10 Touch average applied force .10 .16 .14 Touch average length .25 * .20 .28 * Open in a new tab Notes: *: p<.05, **: p<.01, ***: p<.001, ****: p<.0001. Values in this table represent Spearman’s rank correlation coefficient (r s ). Discussion Research utilizing digital phenotyping to improve the understanding of various conditions, including autism, has demonstrated consistent growth in recent years ( dos Santos et al., 2024 ). This growth reflects advancements in technology that enable continuous, passive, and active data collection, offering deeper insights into behavioral and physiological patterns that may enhance diagnosis and intervention strategies across diverse conditions. In a similar fashion, we have previously shown that the SenseToKnow app accurately detects autism-related features in toddlers who are 16–40 months of age ( Perochon et al., 2023a ; Krishnappa Babu et al., 2024b ). The current study builds upon this earlier work showing that the majority of digital features derived from the SenseToKnow app also demonstrate significant group differences between older autistic and neurotypical children who are preschool- and school-age, with small to medium effect sizes. Further work is needed to determine whether the SenseToKnow app can detect behaviors that distinguish autism from other neurodevelopmental conditions, which was not the focus of the present study. In the present study, autistic and neurotypical children differed in terms of multiple gaze features, consistent with previous research using the SenseToKnow app with toddlers ( Chang et al., 2021 ). Specifically, the autistic group exhibited reduced attention to the social elements in the movies and lower gaze silhouette scores, the latter indicating less cohesive and less distinctive clustering for the gaze patterns in movies that included both social and nonsocial elements, consistent with previous work showing reduction of gaze fixation in school-age autistic children ( Wan et al., 2019 ). These findings are consistent with previous eye-tracking studies examining eye gaze in autistic children viewing social and nonsocial stimuli. Autistic children also exhibited reduced gaze speech correlations, indicating a lower ability to coordinate gaze with speech while observing a conversation between two adults. Such studies support the hypothesis that autistic individuals demonstrate reduced attention to social stimuli (Stuart et al., 2022; Kim et al., 2024 ). Across the different movies shown in the app, autistic children displayed greater head movement rates and acceleration, consistent with our previous work with toddlers ( Dawson et al., 2018 and Krishnappa Babu et al., 2023b ). However, we found that preschool- and school-age autistic participants displayed lower head movement complexity and facial movement complexity (i.e., greater predictability) compared to neurotypical children, while in past studies autistic participants in toddlerhood were found to have greater complexity than neurotypical children across eyebrows, mouth, and head movements ( Krishnappa Babu et al., 2023b , Krishnappa Babu et al., 2023c ). Research has shown that facial expression complexity tends to differ between autistic and neurotypical toddlers ( Webster et al., 2021 ). Specifically, autistic children often produce more ambiguous facial expressions compared to their neurotypical peers. For example, a study found that autistic children exhibit fewer spontaneous expressions, and their facial expressions are perceived as less frequent, fleeting, or harder to interpret in social contexts. These differences can become more pronounced over time as neurotypical children develop more varied and complex facial expressions, while autistic children may continue to show lower levels of expressiveness, especially in spontaneous social situations​ ( Grossard et al., 2020 ). Autistic children showed significantly higher blink rates (reflecting lower attentional engagement) during both social and nonsocial movies, similar to previous findings using the SenseToKnow app with toddlers ( Krishnappa Babu et al., 2023a ). In the current study, we also found that neurotypical children spent a greater proportion of time facing forward while watching the stimuli than autistic children during both social and nonsocial movies, whereas in previous work this was evident only during social movies ( Krishnappa Babu et al., 2023a ). Reduced attentional engagement for autistic participants is also consistent with other digital phenotyping work (Bazelmans et al., 2021). In contrast to findings from our previous studies with toddler-age autistic and neurotypical children ( Perochon, et al. 2021 ), unexpectedly, the preschool- and school-age autistic children were more likely to respond to their name than their neurotypical counterparts, an effect that was no longer significant once we controlled for the age of the participants. Findings of differences between autistic and neurotypical children in multiple digital phenotypes remained after controlling for age and sex. When controlling for developmental quotient, it was found that autistic children showed reduced gaze at social stimuli, more diffuse gaze patterns, lower correlations between their gaze patterns and speech, a higher proportion of responses to name, and lower head movement complexity during the nonsocial movies. In the present study, autistic and neurotypical children, even after correcting for age, did not differ in their performance on the majority of features computed from the bubble-popping game, including touch popping ratio, touchscreen exploratory percentage and touch average length, contrasting with findings in our studies of autistic toddlers ( Perochon et al., 2023b ). This suggests that, as motor skills develop, this component of the app is less sensitive in detecting differences in fine motor features between autistic and neurotypical participants. This may reflect a ceiling effect for both groups in older children, i.e., past a certain age, regardless of diagnosis group, the children are able to successfully and efficiently complete the bubble popping game. In this current study, autistic and neurotypical participants had a touch popping ratio of 92% and 95% respectively, which is notably higher than in toddlers ( Perochon et al., 2023b ) i.e., ~64% and ~75% respectively, indicating a far greater proportion of successful and purposeful touches on the screen in the older children. We found that the SenseToKnow features were stable across repeated administrations shown by consistent feature distributions between repeated administrations. The majority of features also displayed test-retest stability at the individual participant level. The consistency of these features across testing repetitions is an important characteristic for quantitative assessments that are being developed for screening, diagnosis, and clinical trials. When we examined correlations between the Vineland Adaptive Behavior Scales (VABS) and the digital phenotypes, we found that autistic children with higher communication, daily living, socialization, motor, and adaptive skills on the VABS exhibited significantly higher levels of social attention and coordinated gaze with speech, less frequent head movements, higher complexity of facial movements, higher overall attention, lower blink rates (greater attentional engagement), and higher visual motor skills. Within the VABS subdomains, we found that app features showed the strongest correlations with the communication subdomain scores (18 of the 19 features were correlated ranging from .25–.51) and the socialization subdomain scores (14 of the 19 features were correlated ranging from .22–.48). Correlations with daily living and motor skills scores were not as strong, while still significant. Stronger correlations for the subdomains of communication and socialization as compared to the daily living and motor skills scores provides some support for divergent validity. We also found that higher scores (increased autism traits) on the SRS Social Awareness subdomain were associated with a lower proportion of responses to name, reduced gaze at social stimuli and reduced gaze clustering (gaze silhouette score); lower correlation between gaze and speech; reduced head movement, eye brow, and mouth complexity, and less time facing toward the screen; as well as more frequent head movements and higher blink rates (indicated lower attentional engagement) during social movies, and longer touch average length during the bubble-popping task. Similar patterns were found for the SRS Restricted and Repetitive Behavior subdomain scores and for the total SRS scaled scores. These findings indicate the convergent validity of app features with clinical measures, suggesting that SenseToKnow features are related to relevant clinical measures in multiple behavioral domains in children across a wide age range. Although this work represents an important first step toward demonstrating that SenseToKnow, a digital phenotyping app, can detect autism-related behaviors in preschool- and school-age children, future work will be needed to determine if the SenseToKnow app is useful for outcome monitoring in clinical trials. Moreover, several limitations of this study should be noted. First, the two groups significantly differed in terms of their intellectual ability, a factor that influenced some of the digital phenotypes. Second, this study sample did not account for the presence of co-occurring conditions with autism. Future studies will need to disentangle the effects of intellectual ability and autism diagnosis on app feature presentation and further examine the influence of common co-occurring conditions, including intellectual disability, language delay, and attention-deficit/hyperactivity disorder, among others. Future work should aim to include DQ-matched samples across diagnoses to better understand group differences. Third, while the neurotypical and autistic groups are matched in sex, race, and caregiver education level, the study sample is primarily composed of White participants with highly educated caregivers. Finally, a longitudinal study is needed with larger and more diverse samples to better understand the generalizability of our findings and to carefully assess developmental changes in the features detected by the SenseToKnow app. Altogether, more diverse and representative neurotypical and autism samples, alongside children with other neurodevelopmental conditions, would provide stronger evidence for the use of the SenseToKnow app for measuring autism-related behaviors in this age range. These current group differences, along with significant correlations with various clinical features, suggest that the SenseToKnow app might be useful as a tool for screening and assessment of autism in older children. However, in order to better evaluate the use of the SenseToKnow app as a screener in older children, future prospective studies will need a much larger study cohort, with a more representative non-autism comparison group that includes both neurotypical and other non-autism developmental disabilities, such as children with language and developmental delay. This would allow for the use of deep learning models to better understand feature-based cutoffs between diagnosis groups and allow for better predictive separation on the individual level, rather than at the group level shown here. In summary, we found that several behavioral features derived from SenseToKnow, a digital phenotyping app, differed between 3- to 8-year-old autistic and neurotypical children. Particularly across gaze, head movement rate and acceleration, facing forward and blink related features, we find several consistent findings between this current study and previous work using the SenseToKnow app with toddlers. Repeat administration on a subset of children indicated that these features show stability across a short time period. Finally, we found that several of the individual digital features derived from SenseToKnow correlate with standardized clinical measures of higher communication, daily living, socialization, motor, and adaptive skills, as well as level of autism traits. These findings are encouraging for the potential of using digital phenotyping as an objective and quantitative way of assessing autism-related behaviors in preschool- and school-age children. Supplementary Material Supinfo NIHMS2069551-supplement-Supinfo.docx (636.2KB, docx) Acknowledgements This work is supported by NIH Autism Center of Excellence Award NICHD P50HD093074 (Dawson, PI), NIMH R01MH121329 (Dawson and Sapiro, MPI), NIMH R01MH120093 (Sapiro and Dawson, MPI), and the Simons Foundation, with additional support provided by the National Science Foundation, and Office of Naval Research. The funders had no direct participation in the preparation of this manuscript, decision to publish, study design or data analysis for this project. The authors wish to extend their sincere thanks to the numerous caregivers and children whose involvement was essential for this research. Additionally, the authors appreciate the valuable contributions from many research staff and clinical research specialists. Footnotes Conflict of Interests Statement Geraldine Dawson is on the scientific advisory boards of Tris Pharma, Inc. and the Nonverbal Learning Disability Project, is a consultant to YAMO Pharmaceuticals and Guidepoint Global, and receives book royalties from Guilford Press and Springer Nature Press. Dr. Carpenter, Dr. Dawson, and Dr. Sapiro have developed technology, data, and/or products that have been licensed, and they and Duke University have benefited financially. Guillermo Sapiro is affiliated with Apple; the work here reported is independent of such affiliation. Patient Consent Statement The study design and methods were approved by Duke University School of Medicine Institutional Review Board (IRB-Pro00085156). The parent or legal guardian signed informed consent forms approved by the Duke Institutional Review Board. Children over the age of 6 years provided verbal assent if they were capable of doing so. References Achenbach TM, & Ruffle TM (2000). The child behavior checklist and related forms for assessing behavioral/emotional problems and competencies. Pediatrics in Review, 21(8), 265–271. 10.1542/pir.21-8-265 [ DOI ] [ PubMed ] [ Google Scholar ] Baltrusaitis T, Zadeh A, Lim YC, & Morency LP (2018). OpenFace 2.0: Facial behavior analysis toolkit. Proceedings - 13th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2018, 59–66. 10.1109/FG.2018.00019 [ DOI ] [ Google Scholar ] Bazelmans T, Jones EJH, Ghods S, Corrigan S, Toth K, Charman T, & Webb SJ (2023). Identifying phenotypic and physiological subgroups of preschoolers with autism spectrum disorder. Psychological Medicine, 53(4), 1592–1602. 10.1017/S0033291721003172 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Campbell K, Carpenter KLH, Hashemi J, Espinosa S, Marsan S, Borg JS, Chang Z, Qiu Q, Vermeer S, Adler E, Tepper M, Egger HL, Baker JP, Sapiro G, & Dawson G (2019). Computer vision analysis captures atypical attention in toddlers with autism. Autism, 23(3), 619–628. 10.1177/1362361318766247 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Carpenter KLH, Hashemi J, Campbell K, Lippmann SJ, Baker JP, Egger HL, Espinosa S, Vermeer S, Sapiro G, & Dawson G (2021). Digital behavioral phenotyping detects atypical pattern of facial expression in toddlers with autism. Autism Research, 14(3), 488–499. 10.1002/aur.2391 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chang Z, Di Martino JM, Aiello R, Baker J, Carpenter K, Compton S, Davis N, Eichner B, Espinosa S, Flowers J, Franz L, Harris A, Howard J, Perochon S, Perrin EM, Krishnappa Babu PR, Spanos M, Sullivan C, Walter BK, … & Sapiro G (2021). Computational methods to measure patterns of gaze in toddlers with autism spectrum disorder. JAMA Pediatrics, 175(8), 827–836. 10.1001/jamapediatrics.2021.0530 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Coffman M, Di Martino JM, Aiello R, Carpenter KLH, Chang Z, Compton S, Eichner B, Espinosa S, Flowers J, Franz L, Perochon S, Krishnappa Babu PR, Sapiro G, & Dawson G (2023). Relationship between quantitative digital behavioral features and clinical profiles in young autistic children. Autism Research, 16(7), 1360–1374. 10.1002/aur.2955 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Constantino JN (2012). Social responsiveness scale. Encyclopedia of Autism Spectrum Disorders, 4457–4467. 10.1007/978-3-319-91280-6_296 [ DOI ] [ Google Scholar ] Dawson G, Campbell K, Hashemi J, Lippmann SJ, Smith V, Carpenter K, Egger H, Espinosa S, Vermeer S, Baker J, & Sapiro G (2018). Atypical postural control can be detected via computer vision analysis in toddlers with autism spectrum disorder. Scientific Reports, 8(1). 10.1038/s41598-018-35215-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Dawson G, & Sapiro G (2019). Potential for digital behavioral measurement tools to transform the detection and diagnosis of autism spectrum disorder. JAMA Pediatrics, 173(4), 305–306. 10.1001/jamapediatrics.2018.5269 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] de Belen RAJ, Bednarz T, Sowmya A, & Del Favero D (2020). Computer vision in autism spectrum disorder research: a systematic review of published studies from 2009 to 2019. Translational Psychiatry, 10(1). 10.1038/s41398-020-01015-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Deng J, Guo J, Yang J, Xue N, Kotsia I, & Zafeiriou S (2022). ArcFace: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10), 5962–5979. 10.1109/TPAMI.2021.3087709 [ DOI ] [ PubMed ] [ Google Scholar ] Dow D, Holbrook A, Toolan C, McDonald N, Sterrett K, Rosen N, Kim SH, & Lord C (2022). The brief observation of symptoms of autism (BOSA): Development of a new adapted assessment measure for remote telehealth administration through COVID-19 and beyond. Journal of Autism and Developmental Disorders, 52(12), 5383–5394. 10.1007/s10803-021-05395-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] dos Santos MP, Heckler WF, Bavaresco RS, & Barbosa JLV (2024). Machine learning applied to digital phenotyping: A systematic literature review and taxonomy. Computers in Human Behavior, 161. 10.1016/j.chb.2024.108422 [ DOI ] [ Google Scholar ] DuPaul GJ, Power TJ, Anastopoulos AD, & Reid R (1998). ADHD rating scale — IV : Checklists, norms, and clinical. Guildford Publishing. [ Google Scholar ] Elliot CD (1990). Differential ability scales: Handbook. The Psychological Corporation. [ Google Scholar ] Fusaroli R, Lambrechts A, Bang D, Bowler DM, & Gaigg SB (2017). Is voice a marker for autism spectrum disorder? A systematic review and meta-analysis. Autism Research, 10(3), 384–407. 10.1002/aur.1678 [ DOI ] [ PubMed ] [ Google Scholar ] Grossard C, Dapogny A, Cohen D, Bernheim S, Juillet E, Hamel F, Hun S, Bourgeois J, Pellerin H, Serret S, Bailly K, & Chaby L (2020). Children with autism spectrum disorder produce more ambiguous and less socially meaningful facial expressions: An experimental study using random forest classifiers. Molecular Autism, 11(1). 10.1186/s13229-020-0312-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gu Z, Dawson G, & Engelhard M (2023). Sex differences in the age of childhood autism diagnosis and the impact of co-occurring conditions. Autism Research, 16(12), 2391–2402. 10.1002/aur.3043 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hashemi J, Dawson G, Carpenter KLH, Campbell K, Qiu Q, Espinosa S, Marsan S, Baker JP, Egger HL, & Sapiro G (2021). Computer vision analysis for quantification of autism risk behaviors. IEEE Transactions on Affective Computing, 12(1), 215–226. 10.1109/TAFFC.2018.2868196 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kanne SM, Carpenter LA, & Warren Z (2018). Screening in toddlers and preschoolers at risk for autism spectrum disorder: Evaluating a novel mobile-health screening tool. Autism Research, 11(7), 1038–1049. 10.1002/aur.1959 [ DOI ] [ PubMed ] [ Google Scholar ] Kim NY, He J, Wu Q, Dai N, Kohlhoff K, Turner J, Paul LK, Kennedy DP, Adolphs R, & Navalpakkam V (2024). Smartphone-based gaze estimation for in-home autism research. Autism Research, 17(6), 1140–1148. 10.1002/aur.3140 [ DOI ] [ PubMed ] [ Google Scholar ] Krishnappa Babu PR, Aikat V, Di Martino JM, Chang Z, Perochon S, Espinosa S, Aiello R, L. H. Carpenter K, Compton S, Davis N, Eichner B, Flowers J, Franz L, Dawson G, & Sapiro G (2023a). Blink rate and facial orientation reveal distinctive patterns of attentional engagement in autistic toddlers: A digital phenotyping approach. Scientific Reports, 13(1). 10.1038/s41598-023-34293-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Krishnappa Babu PR, Di Martino JM, Chang Z, Perochon S, Aiello R, Carpenter KLH, Compton S, Davis N, Franz L, Espinosa S, Flowers J, Dawson G, & Sapiro G (2023b). Complexity analysis of head movements in autistic toddlers. Journal of Child Psychology and Psychiatry and Allied Disciplines, 64(1), 156–166. 10.1111/jcpp.13681 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Krishnappa Babu PR, Di Martino JM, Chang Z, Perochon S, Carpenter KLH, Compton S, Espinosa S, Dawson G, & Sapiro G (2023c). Exploring complexity of facial dynamics in autism spectrum disorder. IEEE Transactions on Affective Computing, 14(2), 919–930. 10.1109/TAFFC.2021.3113876 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Krishnappa Babu PR, Di Martino JM, Carpenter KLH, Compton S, Davis N, Eichner B, Espinosa S, Franz L, Perochon S, Dawson G, & Sapiro G (2024a). Large-scale validation of a scalable and portable behavioral digital screening tool for autism at home. Conference on Human Factors in Computing Systems - Proceedings. 10.1145/3613905.3650995 [ DOI ] [ Google Scholar ] Krishnappa Babu PR, Di Martino JM, Aiello R, Eichner B, Espinosa S, Green J, Howard J, Perochon S, Spanos M, Vermeer S, Dawson G, & Sapiro G (2024b). Validation of a mobile app for remote autism screening in toddlers. New England Journal of Medicine AI. 10.1056/AIcs2400510 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lidstone DE, Rochowiak R, Pacheco C, Tunçgenç B, Vidal R, & Mostofsky SH (2021). Automated and scalable computerized assessment of motor imitation (CAMI) in children with autism spectrum disorder using a single 2D camera: A pilot study. Research in Autism Spectrum Disorders, 87. 10.1016/j.rasd.2021.101840 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lord C, Rutter M, DiLavore P, Risi S, Gotham K, & Bishop S (2012). Autism diagnostic observation schedule. In Springer Reference. Western Psychological Services. 10.1007/springerreference_184220 [ DOI ] [ Google Scholar ] Lord C, Rutter M, & Le Couteur A (1994). Autism Diagnostic Interview-Revised: A revised version of a diagnostic interview for caregivers of individuals with possible pervasive developmental disorders. Journal of Autism and Developmental Disorders, 24(5), 659–685. 10.1007/BF02172145 [ DOI ] [ PubMed ] [ Google Scholar ] McCracken JT, Anagnostou E, Arango C, Dawson G, Farchione T, Mantua V, McPartland J, Murphy D, Pandina G, & Veenstra-VanderWeele J (2021). Drug development for autism spectrum disorder (ASD): Progress, challenges, and future directions. European Neuropsychopharmacology, 48, 3–31. 10.1016/j.euroneuro.2021.05.010 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] MacFarlane H, Salem AC, Chen L, Asgari M, & Fombonne E (2022). Combining voice and language features improves automated autism detection. Autism Research, 15(7), 1288–1300. 10.1002/aur.2733 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] McPartland JC (2016). Considerations in biomarker development for neurodevelopmental disorders. Current Opinion in Neurology, 29(2), 118–122. 10.1097/WCO.0000000000000300 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Mukherjee D, Bhavnani S, Lockwood Estrin G, Rao V, Dasgupta J, Irfan H, Chakrabarti B, Patel V, & Belmonte MK (2024). Digital tools for direct assessment of autism risk during early childhood: A systematic review. Autism, 28(1), 6–31. 10.1177/13623613221133176 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Perochon S, Di Martino JM, Aiello R, Baker J, Carpenter K, Chang Z, Compton S, Davis N, Eichner B, Espinosa S, Flowers J, Franz L, Gagliano M, Harris A, Howard J, Kollins SH, Perrin EM, Raj P, Spanos M, … Dawson G (2021). A scalable computational approach to assessing response to name in toddlers with autism. Journal of Child Psychology and Psychiatry and Allied Disciplines, 62(9), 1120–1131. 10.1111/jcpp.13381 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Perochon S, Di Martino JM, Carpenter KLH, Compton S, Davis N, Eichner B, Espinosa S, Franz L, Krishnappa Babu PR, Sapiro G, & Dawson G (2023a). Early detection of autism using digital behavioral phenotyping. Nature Medicine, 29(10), 2489–2497. 10.1038/s41591-023-02574-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Perochon S, Di Martino JM, Carpenter KLH, Compton S, Davis N, Espinosa S, Franz L, Rieder AD, Sullivan C, Sapiro G, & Dawson G (2023b). A tablet-based game for the assessment of visual motor skills in autistic children. Npj Digital Medicine, 6(1). 10.1038/s41746-023-00762-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ranti C, Jones W, Klin A & Shultz S (2020). Blink rate patterns provide a reliable measure of individual engagement with scene content. Sci. Rep. 10, 1–10. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Shic F, Naples AJ, Barney EC, Chang SA, Li B, McAllister T, Kim M, Dommer KJ, Hasselmo S, Atyabi A, Wang Q, Helleman G, Levin AR, Seow H, Bernier R, Charwaska K, Dawson G, Dziura J, Faja S, … & McPartland JC (2022). The autism biomarkers consortium for clinical trials: Evaluation of a battery of candidate eye-tracking biomarkers for use in autism clinical trials. Molecular Autism, 13(1). 10.1186/s13229-021-00482-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sparrow S, Cicchetti DV, & Balla DA (2016). Vineland adaptive behavior scales - 3 (3rd ed.). Pearson Inc. [ Google Scholar ] Stuart N, Whitehouse A, Palermo R, Bothe E, & Badcock N (2023). Eye gaze in autism spectrum disorder: A review of neural evidence for the eye avoidance hypothesis. Journal of Autism and Developmental Disorders, 53(5), 1884–1905. 10.1007/s10803-022-05443-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tunçgenç B, Pacheco C, Rochowiak R, Nicholas R, Rengarajan S, Zou E, Messenger B, Vidal R, & Mostofsky SH (2021). Computerized assessment of motor imitation as a scalable method for distinguishing children with autism. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 6(3), 321–328. 10.1016/j.bpsc.2020.09.001 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Uddin LQ, Dajani DR, Voorhies W, Bednarz H, & Kana RK (2017). Progress and roadblocks in the search for brain-based biomarkers of autism and attention-deficit/hyperactivity disorder. Translational Psychiatry, 7(8). 10.1038/TP.2017.164 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] van ‘t Hof M, Tisseur C, van Berckelear-Onnes I, van Nieuwenhuyzen A, Daniels AM, Deen M, Hoek HW, & Ester WA (2021). Age at autism spectrum disorder diagnosis: A systematic review and meta-analysis from 2012 to 2019. Autism, 25(4), 862–873. 10.1177/1362361320971107 [ DOI ] [ PubMed ] [ Google Scholar ] Walsh P, Elsabbagh M, Bolton P, & Singh I (2011). In search of biomarkers for autism: Scientific, social and ethical challenges. Nature Reviews Neuroscience, 12(10), 603–612. 10.1038/nrn3113 [ DOI ] [ PubMed ] [ Google Scholar ] Wan Kong X, Sun B, Yu S, Tu Y, Park J, Lang C, Koh M, Wei Z, Feng Z, Lin Y, Kong J,G, Wan G, Kong X, Sun B, Yu S, Tu Y, Park J, Lang C, Koh M, Wei Z, Feng Z, Lin Y, & Kong J (2019). Applying eye tracking to identify autism spectrum disorder in children. Journal of Autism and Developmental Disorders, 49(1), 209–215. https://pubmed.ncbi.nlm.nih.gov/30097760/ [ DOI ] [ PubMed ] [ Google Scholar ] Webster PJ, Wang S, & Li X (2021). Review: Posed vs. genuine facial emotion recognition and expression in autism and implications for intervention. Frontiers in Psychology, 12. 10.3389/fpsyg.2021.653112 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wechsler D (2012). WPPSI-IV: Wechsler preschool and primary scale of intelligence (4th ed.). Bloomington, MN: Pearson, Psychological Corporation. [ Google Scholar ] Wechsler D (2015). WISC-V Integrated: Wechsler intelligence scale for children - integrated. (Fifth edition : Integrated. ed.). San Antonio, Tex: NCS Pearson, Inc. [ Google Scholar ] Wei Q, Cao H, Shi Y, Xu X, & Li T (2022). Machine learning based on eye-tracking data to identify autism spectrum disorder: A systematic review and meta-analysis. SSRN Electronic Journal. 10.2139/ssrn.4100664 [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supinfo NIHMS2069551-supplement-Supinfo.docx (636.2KB, docx) ACTIONS View on publisher site PDF (801.5 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Visual Integration Mechanisms in Young Adults with High and Low Autistic Traits#1133193
• ‘It is my life don’t take that choice from me’: Autistic young adults’ experiences, preferences and needs regarding autistic peer support#428543
• A Stronger Association Between Screen Time and Externalizing Problems in Typically Developing Children than in Children with Autism Spectrum Disorder#314140
• Autism Beyond the Checklist: Phenocopies, Diagnostic Overshadowing, and the Clinical Workup#381243
• Promoting Positive Outcomes for Individuals With ASD: Linking Early Detection, Treatment, and Long-term Outcomes#1072642