Machine Learning Tool Flags PTSD Risk in Traumatized Kids

Megan Brooks

July 13, 2017

A machine learning (ML) algorithm accurately predicts which children are at risk for posttraumatic stress disorder (PTSD) after trauma, according to a first proof-of-concept application of ML to predict childhood PTSD, researchers say.

The algorithm also provides insights into causal factors for PTSD after trauma (including the child's pain and the parent's level of stress). This suggests that possible approaches to PTSD prevention may include better pain treatment for the child and more support for parents in the days and weeks after the trauma, the researchers say.

The study was published online July 10 in BMC Psychiatry.

"Considerable" Predictive Signal

Research has shown that more than 20% of children in the United States will suffer a traumatic event before age 16 years and that 10% to 40% will develop PTSD as a result of that trauma.

"The goal of our research is to be able to predict – as early as possible after trauma, with information available from the child and about the child – which child would develop PTSD and which wouldn't, and then intervene early," Glenn Saxe, MD, professor of child and adolescent psychiatry at NYU Langone's Child Study Center in New York City, told Medscape Medical News.

Data for the study came from 163 patients aged 7 to 18 years. The data were collected as part of a National Institute of Mental Health–funded study on risk factors for PTSD in children hospitalized with injuries. Injured children were assessed within hours or days after their hospitalization and were reassessed 3 months following discharge.

For their predictive model, the researchers considered 105 variables measured during the hospitalization. These variables belong to such domains as childhood development, demographics, parent symptoms, stress, magnitude of injury, candidate genes, neuroendocrine and psychophysiologic response, and child symptoms and functioning.

Eleven of the 163 (7%) injured children were classified as having PTSD after scoring 38 or higher on the UCLA PTSD Reaction Index. The patients were assessed 3 months after discharge. This cut-off score is based on a high level of symptoms of PTSD and is strongly related to a DSM-IV diagnosis of PTSD, the researchers explain.

Using five ML classification methods, a predictive classification model was obtained with "considerable predictive signal." The best model had an average area under the receiver operating characteristic curve of 0.79. "These are much stronger than the performance yielded by the conventional classification methods, where predictive signal was at the chance level," the researchers note.

In addition, using ML "causal discovery" feature selection methods, they identified 10 variables that "may be on a causal pathway to PTSD," said Dr Saxe. The variables include having a prior history of PTSD; having externalizing symptoms; loss and help seeking; acute pain in the child; acute stress symptoms in the parent; mutations in several candidate genes (ie, COMT, CRHR1, and FKBP5); prior ketamine exposure; and two protective factors (child history of breast feeding and attending religious services).

"The specific variables that were found to be most stable in the predictive classification model are interesting and may shed light on the process by which PTSD emerges – and the possibility of prevention – in acutely injured children," the authors write. "The early identification of a child's level of risk – and specific vulnerabilities – opens the possibility of preventative intervention tailored to the child's specific needs," they add.

"There obviously needs to be more research, and this needs to be replicated in larger samples," said Dr Saxe. "Ultimately, machine learning is ideal for being able to know the risk of an individual person, and this really has translational potential to clinical care."

The study was supported by the National Institute of Mental Health. The authors have disclosed no relevant financial relationships.

BMC Psychiatry. Published online July 10, 2017. Full text


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.