Diagnosing autism in children is an important task. Various studies show that early intervention can significantly improve the outcome for children in later life.

The idea is to teach autistic children important social skills and behavioural patterns before other patterns become ingrained. And to be most effective, that needs to be done as early as possible, at say 2 or 3 years old.

And yet the average age for diagnosis in the US is 5. That’s largely because the process of diagnosis in young children is particularly difficult and taxing. It requires a psychologist with specialist expertise in autism who analyses a child’s behaviour in a one-on-one setting over a significant period of time. 

Video plays an increasingly important role in this process but again it requires detailed frame-by-frame analysis by an expert. That’s why this kind of diagnosis is time consuming and expensive. 

That looks set to change. Today, Jordan Hashemi at the University of Minnesota and a few pals say they’ve developed a computer vision technique to automatically identify behaviour that shows an increased risk of autism.

The system relies on video footage of the child in an ordinary setting and automatically works out the position of the child’s head and how it changes during activities, as well as the position of the arms, torso and legs for analysis of body position and gait.

Autism is generally associated with impaired social interaction and impaired communication. This is most easily diagnosed in children beyond the age of 5 when the differences with their peers become most apparent.

But in recent years, child psychologists have found various other indicators they can use to spot toddlers and infants with a higher risk of autism.

For example, one test is to shake a rattle on one side of an infant’s field of vision, to wait for the child to look at it and then move it to the other side of the field of vision. The question is how well the child tracks the object. Infants with autism tend to show delayed or discontinuous tracking.

Another test is shake a noisy toy on one side of the infant and wait till it engages his or her attention and then to shake another noisy toy on the other side of the child. The question here is how quickly the child switches his or her attention to the second toy. A delayed response shows an increased risk of autism.

Yet another test that is less well established, is to look for unusual body movement such as walking on tip toes, flapping the arms and hands and various kinds of asymmetric body posture such as walking with one arm pointing straight ahead and the other pointing down by the side.

Hashemi and co have designed their computer vision system to look for exactly these behaviours. 

These guys measure attention by monitoring head movement, using the position of the left ear, left eye and nose. They also generate a 2D stick skeleton of the child to monitor body movement.

One of the big challenges in this kind of research is validating these kinds of models and building up a statistically significant corpus of data. These guys tested their approach by videoing the behaviour of 15 infants and toddlers of both sexes between the ages of 5 and 18 months. All the children were either siblings of children already diagnosed with autism or showing developmental delay themselves.

Each child was assessed in four different ways: by a specialist psychologist in child autism,  a child psychiatrist; two psychology students; and the new software  system.

The results are impressive. Hashemi and co take the expert’s assessment as a kind of gold standard that the others must match. In these tests, the computer vision system outperformed the child psychiatrist and psychology students by more often agreeing with the expert’s assessment.

“This work is the first achieved milestone in a long-term project for non-invasive early observation of children in order to aid in diagnosis of neurodevelopmental disorders,” say Hashemi and co.

 That’s certainly a good start but these guys will want to gain a significantly greater body of evidence to back up this work. 

That may not be too difficult to gather. Hashemi and co point out that their computer vision system can be tested on historical videos taken both at home and in a clinical setting, of which there are significant volumes.

Once the behaviour of the children in these videos is quantified, it will become much eaiser to crunch the data looking for patterns that may otherwise be hidden from human observers. That could reveal other indicators for autism. 

A human expert will have to be involved in the final diagnosis to help avoid false positives and false negatives. And of course, this will never be a perfect tool, given the dazzling variety of behaviour that children demonstrate.

Nevertheless, computer vision techniques could help to significantly reduce the time it takes to diagnose infants and toddlers with autism. And the consequent ability to intervene earlier could significantly improve the outcome for these children.

Important work.

Ref: arxiv.org/abs/1210.7014: Computer Vision Tools For The Non-Invasive Assessment Of Autism-Related Behavioral Markers