A tablet-based game for the assessment of visual motor skills in autistic children

The study addresses the need for objective, scalable tools to assess early motor abilities that frequently differ in autistic children and can manifest as early as infancy. Traditional screening tools like the M-CHAT-R/F rely on caregiver report and can have reduced accuracy in real-world settings and be influenced by socioeconomic and demographic factors. Given that motor impairments (including visual-motor integration deficits) are commonly observed in autism and may relate to later language and developmental outcomes, the authors propose a tablet-based bubble-popping game to directly measure visual-motor skills. They hypothesize that autistic children, especially those with co-occurring ADHD, will show distinct motor performance patterns compared to neurotypical peers, and that game-derived digital motor phenotypes will correlate with standardized clinical assessments.

Prior work indicates motor impairments are prevalent in autism (estimated 50–85%) and often among the earliest observable differences, including in children without cognitive impairment. Deficits span gait, balance, coordination, movement accuracy, reaction time, manual dexterity, tone, hyperkinesis, and praxis, with particular challenges in tasks requiring visual-motor integration. Early motor skills have been linked to later autism diagnosis and expressive language outcomes. Evidence suggests autistic individuals may weigh proprioceptive feedback more heavily than visual input. Technology-driven assessments using mobile devices and sensors have successfully characterized motor behavior, including tablet-based approaches that identify autism-related motor signatures and quantify impairments. ADHD is also associated with motor difficulties (notably increased variability in speed and accuracy), and comorbid ADHD may exacerbate motor impairment in autism.

Design: Two cross-sectional studies using a tablet-based bubble-popping game to assess visual-motor performance, with feature extraction from touch and inertial data and statistical comparisons across diagnostic groups. Participants: Study 1 included 151 children aged 18–36 months recruited during well-child visits at four Duke pediatric primary care clinics; 23 were subsequently diagnosed with ASD. Inclusion: age 16–38 months, not ill, caregiver language English or Spanish. Exclusions: sensory/motor impairment precluding app, caregiver uninterested/unavailable, child distress, caregiver popping bubbles, insufficient clinical info. Neurotypical (NT) participants were randomly age-matched to the ASD group from a larger pool. Study 2 included 82 children aged 36–120 months recruited from the community; 63 had ASD (32 with co-occurring ADHD) and 19 were NT. Exclusions included known genetic/neurological syndromes linked to autism, seizure disorder (except simple febrile or seizure-free past year), motor/sensory impairment interfering with measures, and neonatal brain damage. Additional exclusions: child did not understand the game (n=18; NT 13, ASD 5), caregiver popped bubbles (n=5), or insufficient engagement (<3 touches; NT 29, ASD 3). None in Study 2 failed to understand the game. Clinical assessments: Study 1 used the M-CHAT-R/F at the visit; children who failed and/or had developmental concerns were referred for diagnostic evaluation with ADOS-2 and the Mullen Scales of Early Learning (MSEL: ELC, fine motor, visual reception, receptive and expressive language). NT status: did not fail M-CHAT-R/F and no developmental concerns; NT children did not receive diagnostic/cognitive testing. Study 2 ASD diagnosis: ADOS-2 and ADI-R by a research-reliable psychologist. Cognitive ability: Differential Abilities Scales (DAS). ADHD diagnosis: MINI-Kid with ADHD supplements, brief child interview as appropriate, parent/teacher ADHD-RS, and clinical consensus. NT criteria: IQ > 70, Vineland scores in average range, and no clinical elevations on parent rating scales (CBCL, ADHD-RS, SRS). Task and apparatus: The bubble-popping game was administered on 7th/8th generation 10.2" iPads at 60 Hz. Children sat on caregivers’ laps; the iPad was mounted on a tripod ~50 cm away. After a brief demonstration, data collection started once the child independently popped two bubbles; data were recorded for 20 seconds. A bubble popped when the starting location of a touch was within 18.5 mm of its center; the same bubble (same lane/character) reappeared to assess repetitive vs exploratory behavior. Caregivers were instructed not to touch the screen or provide instructions. Touch events and on-device inertial/gyroscopic data were recorded; inertial data provided a proxy for applied pressure. Feature extraction: Nineteen features were computed, including: (1) number of touches; (2) number of pops; (3) bubble popping rate (pops/touches); (4) double touch rate; (5) screen exploratory percentage (touched area proportion); (6) number of targeted bubbles; (7) number of transitions (lane changes); (8) repeat percentage (consecutive same bubble); (9) touch duration (mean/median/std); (10) touch length (mean/median/std); (11) touch velocity (mean/median/std); (12) applied force (mean/median/std) approximated from accelerometer data; (13) distance to the center (mean/median/std); (14) popping accuracy (mean/median/std) relative to bubble area; (15–16) variability measures of popping accuracy (average variation; variability of average and maximum accuracy); (17) touches per target; (18) touch frequency while targeting (mean/median/std); (19) time spent on targeted bubble (mean/median/std). Statistical analysis: Differences in prior electronic game experience used a proportion Z-test. Group differences in age/IQ used two-sided Mann–Whitney U tests with rank-biserial r. Group comparisons of motor variables used one-way ANCOVA with diagnostic group as predictor; covariates were age (Study 1) and age plus IQ (Study 2). Effect sizes reported as eta-squared (η²). P-values were FDR-corrected via Benjamini–Hochberg at α=0.05. Correlations between motor features and clinical variables used Spearman’s rho with significance via Student’s t-distribution, controlling for age. Logistic regression with greedy feature selection and leave-one-out cross-validation evaluated discrimination (ROC AUC with 95% CIs; class imbalance addressed by up-sampling).

Age effects (N=233): Strong correlations between age and performance: number of touches (rho=0.62, p<1e-25), bubble popping rate (rho=0.50, p<1e-17), median distance to center (rho=-0.48, p<1e-16), average touch duration (rho=-0.70, p<1e-36), and average touch length (rho=-0.63, p<1e-28). Study 1 (18–36 months; ASD vs NT): Groups had similar prior tablet experience and high completion (>95%); age distributions matched. No difference in number of touches. ASD toddlers showed: lower bubble popping rate (F(1,148)=15.14, p=7.7e-4, η²=0.09); larger median distance to center (F(1,148)=20.14, p=1.7e-4, η²=0.12); longer average touch length (F(1,148)=23.56, p=5.5e-5, η²=0.14) and higher variability in touch length (F(1,148)=32.70, p=2e-6, η²=0.18); longer average time spent to pop a bubble (F(1,148)=18.56, p=4.6e-4, η²=0.11). Average touch duration also differed (p=1.8e-3, η²=0.07). Study 2 (3–10 years; ASD±ADHD vs NT): With age and IQ covaried, engagement did not differ (number of touches F(1,78)=0.428, p=0.77). ASD showed lower average touch frequency (F(1,57)=14.77, p=1.1e-2, η²=0.21) and lower median time spent targeting a bubble (F(1,57)=10.79, p=2.0e-2, η²=0.16) versus NT. Study 2 (ASD+ADHD vs ASD without ADHD): No group differences in age, prior gaming, or IQ. Engagement similar (number of touches F(1,60)=0.02, p=0.90). ASD+ADHD group showed: less accuracy (higher average distance to center; F(1,60)=12.76, p=1.2e-2, η²=0.12); lower bubble popping rate (F(1,60)=8.98, p=1.7e-2, η²=0.13); more touches per target (F(1,60)=10.0, p=1.4e-2, η²=0.14); greater variability in touches per target (F(1,60)=13.10, p=2.1e-2, η²=0.18), distance to center (F(1,60)=11.26, p=9.9e-3, η²=0.16), and average popping accuracy (F(1,60)=12.71, p=8.6e-3, η²=0.18). Double touch rate also differed (p=1.2e-2, η²=0.12). Feature combination (classification): Study 1 (ASD vs NT): AUCs improved from 0.67 (95% CI 0.56–0.78; average touch length) to 0.71 (0.61–0.81; +average touch duration) to 0.73 (0.63–0.83; +average time spent). Study 2 (ASD+ADHD vs ASD): AUCs improved from 0.68 (0.55–0.81; average distance to center) to 0.74 (0.58–0.84; +number of targets) and remained 0.74 (0.62–0.86; +screen exploratory percentage). Correlations with clinical measures: Study 1 (ASD toddlers, age-controlled): MSEL fine motor T-score positively correlated with pop rate (rho=0.59, p=3.2e-3), double touch rate (rho=0.43, p=4.8e-2), and average popping accuracy (rho=0.62, p=2.0e-3); negative associations were noted with touch velocity and touch duration metrics. MSEL ELC correlated with number of pops (rho=0.51, p=1.5e-2) and average popping accuracy (rho=0.49, p=1.9e-2). Expressive and receptive language T-scores correlated positively with screen exploratory percentage (rho≈0.47–0.48), and expressive language with number of targets; visual reception T-score correlated with repeat percentage. No significant correlations with ADOS calibrated severity. Study 2 (ASD 3–10 years, age-controlled): IQ positively correlated with number of pops (rho=0.35, p=4.9e-3) and negatively with screen exploratory percentage (rho=-0.34, p=6e-3) and variability of touch frequency (rho=-0.32, p≈3e-2). DAS verbal score correlated with number of touches (rho=0.31, p=1.4e-2). DAS spatial score correlated positively with pops (rho=0.39, p=2.1e-3) and negatively with screen exploratory percentage and with average touch duration/velocity and force variability, as well as time spent targeting a bubble. DAS non-verbal composite correlated positively with pops and negatively with double touch rate, screen exploratory percentage, and time spent targeting a bubble. No significant correlations with ADOS severity.

The findings support that an engaging, brief tablet-based bubble-popping game can yield objective, quantitative measures of visual-motor skills across early childhood. Despite similar engagement (number of touches), autistic toddlers showed slower, less accurate, and more variable touch behavior than neurotypical peers, consistent with visual-motor integration difficulties reported in autism. In older children, co-occurring ADHD contributed to further reductions in accuracy and increases in motor variability, highlighting comorbidity as an important determinant of motor performance heterogeneity. The extracted touch-based features tracked developmental changes with age and correlated meaningfully with independent clinical assessments of fine motor and cognitive abilities, particularly spatial and nonverbal skills, underscoring construct validity. Aggregating features in simple machine-learning models improved discrimination between diagnostic groups, suggesting potential utility for scalable screening or assessment tools that complement caregiver-report measures.

This study demonstrates that a simple, scalable tablet game can quantify visual-motor performance differences in autistic children as young as 18 months and in school-age children, including those with co-occurring ADHD. Several features distinguished autistic from neurotypical toddlers and differentiated autistic children with vs without ADHD, and game-derived measures correlated with standardized fine motor and cognitive assessments. These results support touch-based, game-derived digital phenotypes as components of broader, objective screening and assessment approaches for early autism-related differences. Future work should expand sample sizes, examine longitudinal trajectories, integrate multimodal digital measures (e.g., gaze, facial dynamics, posture), and evaluate longer gameplay to capture learning, focus, and anticipatory behaviors, ultimately advancing toward comprehensive, multi-feature digital phenotyping for developmental assessment.

Limitations include limited sample sizes for subgroup analyses by sex and demographics and relatively small autistic samples, constraining generalizability and the evaluation of machine learning models. Studies 1 and 2 used different clinical batteries, limiting cross-study comparisons. The 20-second game duration may miss information about learning, focus, and anticipation. In Study 1, NT children did not receive full diagnostic/cognitive evaluations, so undetected developmental concerns cannot be completely ruled out.