Early intervention is key in the healthy development of children with autism spectrum disorder (ASD), and the lack of effective screening is a major cause of delayed diagnoses and misdiagnoses. A recent study proposed a means of detection via machine learning.

Studies have indicated that there are significant associations between certain retinal features and ASD, such as retinal nerve fiber layer thinning and significantly larger optic disc and cup diameters. With this in mind, the researchers incorporated retinal image analysis into their machine learning approach. “Retinal images can be obtained from very young children instead of relying solely on lengthy clinical and behavioral assessment,” they wrote in their paper. “This technique provides an objective screening method that can be implemented in a community setting.”

The investigators recruited school-aged participants and assembled an age-matched cohort of controls. They captured retinal images with a nonmydriatic fundus camera and implemented machine learning technology to optimize retinal information and develop a classification model for ASD.

The sensitivity and specificity of the cloud-based algorithm were 95.7% and 91.3%, respectively. The area under the receiver operating characteristic curve was 0.974. Importantly, the researchers noted that specificity tended to be lower in females.

The study authors concluded that their fully automated system can be an effective screening tool for ASD. They suggested further confirmation by professionals.