Monday, June 19, 2017


Sometimes, algorithms pick up on early signs of disease that humans wouldn't even know to look for. Last week, researchers at the University of North Carolina and Washington University reported an AI that can identify autistic infants long before they present behavioral symptoms. It's a thrilling opportunity: Early detection gives autism neuroscience a big leg up, as researchers try to understand what goes wrong during development. But now clinicians and researchers have to figure out what they’ll do with that information—is it just a research tool, or will they one day begin diagnosing and treating autism before symptoms start? Especially when it comes to infants, it won't be easy to entrust medical care to a computer-generated guess.
In this study, researchers scanned the brains of 59 6 month-olds whose older siblings were already diagnosed with autism. By age two, 11 of those infants had received a diagnosis of autism. By training a machine learning algorithm on their behavior and earlier MRI data, the scientists built a model that predicted 9 of those 11 autism cases, with no false positives. The AI predicted autism around a year before the earliest age—around 14 months—that clinicians diagnose it based on behavior.

So this new information is problematic to use: How can clinicians create an intervention for an infant who mightdevelop autism? All of the researchers interviewed for this story agree that early detection and intervention for autism is better. But current autism therapies for babies and toddlers focus on their specific behavioral deficits—teaching children to communicate needs, to play with toys, and to have positive interactions with caregivers. How do you design a treatment when you don't know what those specific deficits will be?