How Tomorrow’s Algorithms Will Help Treat Post-Traumatic Stress Disorder

By
A soldier makes radio contact during an operation in Fallujah, Iraq, August, 2005.
Photo: Scott Olson/2005 Getty Images

Let's say a 33-year-old woman is admitted to an emergency room after being in a car accident. The accident was violent and scary, but other than a mild concussion and some bruises and lacerations, she appears to be okay, physically. But once she's been examined and bandaged, once her vital signs have been confirmed to sit in the normal ranges, another, trickier question pops up: Is she likely to develop post-traumatic stress disorder?

Here's where current medical science falters a little bit. Sure, once PTSD symptoms manifest themselves, researchers have some ideas about how to treat them — assuming the patient has access to solid care, of course. But those same researchers are a lot less far along in their quest to predict who is most likely to develop the disorder in the first place — a task made much more difficult by the lack of understanding about why some people bounce back psychologically from trauma so much more easily than others. If researchers had clearer diagnostic procedures for PTSD risk, it would help alleviate a lot of suffering by quickly directing resources to those most at risk, particularly in big populations like returning soldiers.

So imagine that same woman a few years from now. After patching her up, the ER's attending physician punches some details about the patient into a program on a tablet: her age, gender, the nature of the injury, her responses to a few questions about her personality. The program quickly responds with an estimate of how likely she is to develop PTSD: In this case, an alarmingly high number. Immediately the doctor refers her to a specialist with PTSD expertise — a specialist who will hopefully nip in the bud what would have otherwise been years of disruptive, wrenching symptoms.  

At the moment, such a scenario is science fiction, but it might not be for long — at least if a new paper in BMC Psychiatry is any indication. For their research, a team led by Karen-Inge Karstoft of the Danish Veteran Centre used the results of the Jerusalem Trauma Outreach and Prevention Study, which gathered extensive information on Israeli trauma victims. It's a pretty useful set of data. For each trauma victim, up to 68 variables were recorded, ranging from their demographic details to the nature of their trauma to whether or not they lost consciousness during it — and, crucially, the severity of their PTSD symptoms, if they experienced any.

The details are a bit technical, but the researchers basically fed a computer algorithm data from 957 patients so it could determine which combinations of factors most accurately predicted whether a given patient developed PTSD symptoms, and then removed some of that data and repeated the process over and over. "By doing so," explained Arieh Shalev of the NYU School of Medicine, a co-author of the paper, in an email, "the team simulated the real-life condition in which one or few predictors are missing (e.g., the person did not have his or her heart rated measured soon after trauma) and described multiple and similarly efficient sets of predictors." The idea is to lend some flexibility to how trauma victims are evaluated: Maybe a given victim wasn't checked for a concussion at the time of his injury, to take a hypothetical example, but the nurse examining him does know his age, gender, and how much of a social support network he has, and taken together that data can be just as important for predicting how likely he is to develop PTSD.

This sort of research is still pretty early on, but Shalev explained that his team's goal is to develop easy, user-friendly ways to assess PTSD risk levels. "In the future, when the necessary knowledge base is accumulated, our algorithm will be translated into a web-based computational instrument that a practitioner — or a patient — will be able to use for evaluating his or her risk," he said. "The algorithm will be able to prompt the practitioner or the patient to look for additional information that is relevant to his or her case, and will deliver individual probability ... which is a leap forward." In other words, the program could say something like, "Given what we already know about this patient, if you ask them whether or not they've had nightmares since the trauma and how well they feel they can control their emotions, we can get a lot more accurate in our prediction of whether they'll develop PTSD."

So while we're not there yet, with the help of more finely tuned algorithms, the health-care system may soon be able to make treatment of PTSD much, much more effective.