Early detection of autism using digital behavioral phenotyping

Area of Science:

• Neurodevelopmental Disorders
• Pediatric Health Technology
• Machine Learning in Medicine

Background:

• Early autism detection is crucial for intervention, but current screening tools show lower accuracy in real-world settings, particularly for girls and children of color.
• Traditional screening methods often face challenges in diverse populations, leading to potential disparities in diagnosis and access to services.

Purpose of the Study:

• To assess the accuracy of a novel autism screening digital application (app) in a diverse pediatric population during routine well-child visits.
• To evaluate the app's performance across different demographic subgroups, including sex, race, and ethnicity.
• To explore the potential of digital phenotyping for objective and scalable autism screening.

Main Methods:

• A prospective, multiclinic study involving 475 children aged 17-36 months.
• Administration of a digital screening app during pediatric well-child visits, displaying stimuli to elicit behavioral signs of autism.
• Quantification of behavioral signs using computer vision and machine learning, with an algorithm combining multiple digital phenotypes.

Main Results:

• The autism screening algorithm demonstrated high diagnostic accuracy (AUC=0.90, sensitivity=87.8%, specificity=80.8%).
• The app exhibited similar sensitivity across subgroups defined by sex, race, and ethnicity, suggesting reduced bias.
• High negative predictive value (97.8%) indicates strong ability to rule out autism.

Conclusions:

• Digital phenotyping offers a promising objective and scalable method for autism screening in primary care settings.
• Combining digital phenotyping with caregiver questionnaires may further enhance screening accuracy and equity.
• This technology has the potential to improve early identification and reduce disparities in autism diagnosis and intervention access.