Development of Deep Ensembles to Screen for Autism and Symptom Severity Using Retinal Photographs

Authors:

Abstract
Autism spectrum disorder (ASD) is characterized by 2 core areas of symptoms: social communication impairment and restricted and repetitive behaviors or interests.1 In 2020, the US Centers for Disease Control and Prevention estimated that the prevalence of ASD was 1 in 36; this number continues to grow, possibly due to increased awareness among the public, medical, and research communities.2,3 The 2019 Global Burden of Disease study reported an age-standardized global ASD prevalence of 369.39 per 100 000.4 Several ASD screening tools have demonstrated notable performance.5,6 For example, the Modified Checklist for Autism in Toddlers is primarily based on caregiver report, with sensitivity of 0.83 (95% CI, 0.77-0.88) and specificity of 0.94 (95% CI, 0.89-0.97)7; however, caregiver assessment is influenced by one’s understanding of the child’s developmental milestones.8 The Social Attention and Communication Surveillance is conducted by trained professionals, and it exhibits sensitivity of 0.96 (95% CI, 0.94-0.98) and specificity of 1.00 (95% CI, 0.99-1.00) for ages 12 to 42 months.9 The growing demands for ASD evaluation cannot be met with limited resources, including trained specialists.10 In addition, because of the substantial time required to evaluate individuals suspected of having ASD, inaccessibility to medical services has increased.11 Therefore, objective ASD screening methods are increasingly needed.

This publication has 45 references indexed in Scilit: