Artificial intelligence identification of autism using a smart tablet serious game for preschool children : results from a phase 3 diagnostic trial of 779 children in Sweden and the United Kingdom

Delafield-Butt, Jonathan and Lu, Szu-Ching and Sobota, Krzystof and Young, Robin and Thompson, Lucy and Thorsson, Max and Tachtatzis, Christos and Andonovic, Ivan and Rowe, Philip and Wilson, Phil and Minnis, Helen and McConnachie, Alex and Gillberg, Christopher (2023) Artificial intelligence identification of autism using a smart tablet serious game for preschool children : results from a phase 3 diagnostic trial of 779 children in Sweden and the United Kingdom. In: International Society for Autism Research (INSAR) 2023 Annual Meeting, 2023-05-03 - 2023-05-06, Stockholmsmässan.

[thumbnail of Delafield-Butt-etal-INSAR-2023-identification-of-autism-using-a-smart-tablet-serious-game-for-preschool-children]
Text. Filename: Delafield_Butt_etal_INSAR_2023_identification_of_autism_using_a_smart_tablet_serious_game_for_preschool_children.pdf
Final Published Version
License: Strathprints license 1.0

Download (2MB)| Preview


Background -- Early detection of autism spectrum disorder (ASD) allows early intervention and potentially the best lifelong health outcomes. However, ASD’s complex symptomatology makes diagnosis complex, time-consuming and requires specialist clinical input. Waiting times for diagnosis can be many months or years. Recent evidence suggests the motor system is linked with autism aetiology, providing an accessible modality for computational assessment. Further, most children with or without autism are attracted to smart tablet gameplay. Its touch screen and inertial sensors enable collection of reliable motor kinematic and behavioural data, suggesting a promising new route for accessible, scalable early assessment. This study set out to test promising pilot results of an iPad serious game assessment paradigm (Anzulewicz, Sobota, & Delafield-Butt, 2016) with a gold-standard blinded, multi-site phase 3 diagnostic trial (Millar, McConnachie, Minnis, et al., 2019). Objectives -- To determine the predictive accuracy of a serious smart tablet game for the early identification of autism using pre-trained algorithms naïve to trial data collected in two sites using blinded comparison against clinical diagnosis. Methods -- A registered phase 3 prospective, diagnostic classification study (NCT03438994) tested the predictive accuracy of a smart tablet serious game with artificial intelligence data analytics to identify autism. Three cohorts aged 3 – 6 years participated: children typically developing (TD); children with a clinical diagnosis of autism (ASD); and children with diagnoses of other non-autism neurodevelopmental disorders (OND). 779 children were recruited from Scotland (Glasgow) and Sweden (Gothenburg). Children played two 5-minute games on an iPad. One commercial algorithm and four research algorithms were trained on a previous cohort of children collected prior to this trial (n=767). Algorithms were tested naïvely on these new, blinded trial participant data to classify gameplay patterns as positively or negatively associated with an ASD diagnosis. Classification results then compared against medical diagnosis by a clinical trial unit. A Socio-Emotional Questionnaire (SEQ), ESSENCE-Q, and adaptive function scores were collected for a subset of participants. Sensitivity and specificity of the algorithms to differentiate ASD children from TD children are reported. Results -- 694 participants (331 TD, 185 ASD, and 178 OND) were included in the final analysis. Research algorithms produced up to 0.80 sensitivity and 0.84 specificity. The commercial algorithm underperformed with sensitivity of 0.29 and specificity of 0.47. Peak performance was obtained by the combination of research algorithms and SEQ, yielding up to 0.95 sensitivity and 0.95 specificity. Differences in accuracy between genders, ages, sites, and severity levels will be discussed. Conclusions -- This phase 3 diagnostic accuracy study of a digital health smart tablet assessment of autism demonstrates sensitivities and specificities useful in clinical diagnostic pathways. Furthermore, the assessment is based on computational analysis of motor patterns, indicating strong predictive value of the motor system in early identification of autism. Performance can be enhanced by the inclusion of a brief questionnaire additionally taking into account social and emotional factors. Future work is required to further improve algorithms and develop this system into a clinic- or school-ready tool amenable to integration with screening, assessment or diagnostic pathways.