Skip to content

AI-based autism detection shows promise across cultures

New Stanford research suggests a method of analyzing cell-phone videos of children could alleviate the bottleneck in autism diagnosis around the world.

California families who suspect their child has autism can face months-long waits for a diagnosis. In developing countries, where specialist pediatricians are rare, the problem is much worse.

New Stanford research suggests technology could alleviate the bottleneck in autism diagnosis around the world. Biomedical data scientist Dennis Wall, PhD, and his team are developing a diagnostic method that relies on analyzing short home videos of children playing and interacting with others.

Their newest paper, published recently in the Journal of Medical Internet Research, suggests the technique, which uses artificial intelligence, can be transferred across cultures to help kids worldwide, and may also work to identify other delays of speech or development.

Wall's team collaborated on the latest study with researchers in Dhaka, Bangladesh. They collected 159 short videos of Bangladeshi children who were filmed by their parents using mobile phones. Even in many low-resource settings, mobile phones are ubiquitous; in Bangladesh, 95 percent of the population have phone subscriptions.

The videos had to meet the same criteria Wall's team used for their earlier studies in American kids: The videos showed the child's face and hands, showed social interaction or attempts at interaction, and showed interaction between the child and a toy or other object. About a third of the children in the videos had been previously diagnosed with autism, one third had other forms of speech or developmental delay, and the remaining third were typically developing.

Trained raters watched the videos and scored several aspects of the children's behavior. The team fed these raw scores into machine-learning algorithms they developed. The goal was to produce algorithms that weight the raw scores to predict which children have developmental delays, and distinguish those children from kids with autism.

The team's algorithm identified children with some form of developmental delay with 76 percent accuracy and 76 percent sensitivity, and distinguished autism from other developmental delays with 85 percent accuracy and 76 percent sensitivity. (Accuracy is a composite measure of the test's ability to correctly identify those with or without a diagnosis; sensitivity is the ability to correctly identify those with a diagnosis.)

One important caveat was that the raters -- the people watching the videos -- did not speak the same languages as the Bangladeshi children in the videos. Not surprisingly, compared with the team's prior studies in American children, the algorithms that predicted which Bangladeshi children had developmental delays were less skewed toward verbal interaction and more toward other behaviors, such as whether the kids made eye contact.

The team has several ideas for how to refine the approach, including training video raters who share the cultural and linguistic backgrounds of the children in Bangladesh. The training required for the raters is brief, and the potential impact is large. The authors write:

"This strategy could provide important resources for developmental health in developing countries with few clinical resources for diagnosis, helping children get access to care at an early age."

Image courtesy of Dennis Wall

Popular posts