Box 1.

Considerations for future studies evaluating digital tools for ASD risk assessments.

1. Specificity to developmental delay in general, not to autism in particular: Although the addition of other NDD comparison groups can demonstrate digital assessment tools’ specificity to ASD symptoms, at the stage of community screening and referral, it may be impractical to focus specifically on autism risk alone: a child who is at risk of a developmental delay but not autistic still needs a referral. 2. Participant characteristics: Future studies should report participants’ demographic characteristics in detail. Studies would also benefit from using standardised measures to rule out ASD symptoms in the TD group, especially as we move towards more dimensional and nuanced measures proposed to characterise heterogeneity within groups. Mental age-matched TD comparison groups should be included in studies involving ASD participants with comorbid ID. 3. Choice of device: While Android devices could be prioritised since they are cheaper and more widely available in LMICs, care must be taken to ensure that the sensitivity and accuracy of the device-derived metric is appropriate for the task delivered on the device, and that the delivery of stimuli and collection of responses will be robust to the fast pace of hardware development and marketing. 4. Individual risk measures: Finally, while group differences are adequate to evaluate the potential of these novel technologies, the analytical methods should be refined to allow quantifying individual risk. Bayesian classification and ML methods have been employed by a few studies reviewed. Discriminative features from multiple developmental domains could serve as individual features in an ML algorithm designed to predict autism risk (Dawson & Sapiro, 2019; Jin et al., 2015; Kang et al., 2020; Liu et al., 2016). This multidimensional strategy would be akin to the standard practice of observing multiple behaviours for ASD diagnosis. Therefore, while any one of the features may not be enough to capture the full heterogeneity of the spectrum, the combination as determined through an ML approach may achieve higher degrees of sensitivity or specificity. 5. Ecological validity: Taking advantage of rapid advancements in computer vision, future assessments should focus on computing social and motor metrics relevant to the autism phenotype from brief, automated tasks portable into homes, schools or other ecologically valid settings. Some examples include reciprocal social interactions, repetitive behaviours and sensory sensitivity during regular interactions of the child with their peers, teachers and parents in home or school settings or during solitary play, captured using cameras on tablet computers or smartphones. 6. Patient and public involvement: Stakeholders and community members (e.g. community health workers) must be involved in planning and execution of the research, from the beginning.

ASD: autism spectrum disorder; TD: typically developing; NDD: neurodevelopmental disorders; ID: Intellectual Disability.