Machine Learning Successfully Predicts Childhood PTSD

In recent years, machine learning has been a fun diversion for most of us, bringing joy and horror equally with things like Cooking With Cthulhu and Google Deep Dream’s obsession with dogs. But while we were busy training computers to recognize the difference between kittens and ice cream cones, researchers have been using machine learning in vastly more complicated ways — including mental health diagnoses.

In a proof-of-concept study published this week in BMC Psychiatry, researchers were able to use machine learning to predict PTSD with significant accuracy. This could be a huge change for post-traumatic treatment for kids. We know that between 10 and 40 percent of children who experience traumatic events will develop PTSD, but narrowing down the risk factors to a reliable predictive model has so far been less than entirely successful.

Even the largest meta-analysis of childhood risk factors for traumatic stress was only able to look at 25 risk factors, the only ones that had been identified in more than one of the 64 studies examined. Among those risk factors, only six were assessed in more than 10 studies. The data for a reliable model just isn’t there.

But with machine learning, researchers were able to sidestep that problem. They took 105 potential biopsychosocial risk factor variables — data that could be collected during a child’s hospital visit, like demographics, child symptoms, parent symptoms, stress, magnitude of injury, and several genetic and neurological variables. With that information, their machine learning method was able to rank each child’s risk factors using analysis much more complex than any researcher could calculate alone.

The machine learning protocol able to accurately predict which children were most at risk for developing PTSD after their injuries, which has huge implications for the possibility of early treatment and even prevention. And the study went further, identifying potential causal factors. This points to areas where further research might lead to a change in treatment protocols, and better outcomes.

For example, children anesthetized with Ketamine (which has been shown to have a positive impact on treatment-resistant depression) were more likely to develop PTSD. That doesn’t mean anesthetists should immediately stop using Ketamine with children, but it does mean that more research is required. It’s just as possible that Ketamine itself isn’t the issue — but that it’s replacing another pharmacological factor that helps prevent the development of PTSD, or that it’s interacting with some other factor that wasn’t studied. Machine learning can probably be applied to that, as well.

Mental health issues are complex, and diagnosing them can require taking a staggering number of symptoms and factors into account. With this study, machine learning shows its value as a possible diagnostic tool capable of finding links even the most robust meta-analysis can’t identify.

[Machine learning methods to predict child posttraumatic stress: a proof of concept study]