Social work and machine learning study aims to identify individualised PTSD interventions for US veterans
The concept of using AI for social good is a relatively new one. In fact, in a Google search for “artificial intelligence and social work” among the top related questions is “Will social workers be replaced with robots?” The short answer is no, but the mission of the USC Center for AI in Society (CAIS) in the United States, is how to use AI to impact social work, perhaps the most human-centered of disciplines.
Jordan Davis, assistant professor of social work and associate director of USC Center for AI in Society (CAIS), is leading a transdisciplinary team of researchers, including researchers from the UK, to use machine learning to address an overlooked population: recently discharged veterans with a history of Post-Traumatic Stress Disorder (PTSD) and cannabis use who are not under treatment for either condition.
The research team is a collaboration between three USC schools and an international university. Davis is joined by Bistra Dilkina, associate professor at USC Viterbi and co-director of CAIS, Eric Pedersen, associate professor of psychiatry and the behavioral sciences at Keck School of Medicine of USC, and Daniel Leightley, a Research Fellow at the King’s Centre for Military Health Research at King’s College London, who was introduced to the team through a cultural exchange relationship between USC Social Work’s Military and Veterans Programs (MVP) and the King’s Centre for Military Health. The study applies machine learning to identify significant behavioral triggers for escalating PTSD symptoms to develop preventive, just-in-time interventions.
A population that is not seeking care, but needs it
According to US Veterans Affairs (VA), about 11-20% of those who served in Iraq or Afghanistan have PTSD in any given year, and of those approximately 13% of those meet the diagnostic criteria for cannabis use disorder. Fewer than half of eligible veterans make use of VA health care. This leaves a clear unmet need among veterans who are often self-medicating, a compelling opportunity to better understand and help this hidden population.
“We’re focused on the most vulnerable veterans, those that are returning and have diagnostic criteria, but who are not willing or able to seek treatment,” Davis said. “If we can predict escalations in symptomology among those that aren’t seeking care, then we can certainly do it for those who have intermittently sought care.”
This is the first study of its kind to combine user input, passive data and active data to predict clinically meaningful outcomes for PTSD and problematic cannabis use in US veterans. Eligible participants will be non-treatment-seeking veterans with a history of trauma exposure and past-month cannabis use who are within their first six months of civilian reintegration.
Participants will complete baseline and monthly follow-up surveys, be provided with a wearable activity tracker, have a passive and active data collection app installed on their smartphone, and log brief daily surveys of PTSD symptoms and cannabis use. Machine learning algorithms will then analyze all of these data points to predict symptom escalation and understand daily/weekly symptom interplay.
“A lot of times veterans have PTSD symptoms, but they don’t recognize them until they get out of the military and they’re faced with everyday life,” Pedersen said. “They start to realize that they don’t need to be constantly hypervigilant to look for threats in their surroundings as the pace of life slows down. Then the symptoms that may have been manageable, or even helpful, when they were in the military start to negatively impact their relationships and other areas of their life.
They may learn that one way to manage their symptoms is to get high, which can then come with a host of other problems in the long term. With this study, we’re going to be able to look at how PTSD symptoms and use of cannabis change during those initial months after separation from the military, so we can design better interventions to help veterans during this vulnerable time.”
A new approach to research
Machine learning is not just a more efficient way to perform analysis. Davis points out that it provides an entirely new perspective on the data that cannot be inferred with standard social science methods. Rather than being a dichotomy, machine learning is adding to the conversation.
“If you rely on linear modeling, you’re going to get linear associations and if you rely on machine learning you’re going to get something different,” Davis said. “They’re two different things. Machine learning models can potentially discover more complex relationships that can result in more accurate predictions.”
Over the three-month period of the study, pulling from multiple data sources, the team will collect millions of data points, a volume that would be intimidating for even the most sophisticated traditional analysis methods. According to Leightley, who leads the machine learning and software development components of the study, the volume of data science collected on an individual is incredibly massive and relying on traditional methods for expert review of that data becomes increasingly difficult “Machine learning enables us to try and find those needles in a haystack,” Leightley said. “We can find signals and patterns that can give us indications of worsening PTSD symptoms and intervene.”
Machine learning also allows interventions to be individualized so that veterans are receiving genuine, personalized help, rather than using one-size-fits-all treatment packages. Probability models created by algorithms provide faster predictions that can produce real-time interventions.
So, if a particular veteran has a 90% probability of meeting relapse criteria over the next few days based on their current data, that could trigger a proactive intervention, via either technology in the form of an app or human outreach. Prioritizing care using predictive models can help cash-strapped social service agencies to direct their limited resources in the most targeted and effective ways possible.
“We’re testing the boundaries of what we can do with this technology and using it to make actual impact on people’s lives,” Leightley said.
Davis notes that this research is happening at a critical time, with the enormous influx of veterans returning from Iraq and Afghanistan. He hopes that this novel approach may help to lift the stigma around treatment and engage with veterans before symptoms escalate.
“I believe data-driven models that leverage both rich information as well as advanced machine learning methods can help make progress on some really challenging societal problems, such as PTSD among veterans,” Dilkina added. “However, it is only possible through the close interdisciplinary collaboration between social work experts with machine learning experts. And this is exactly what is very exciting and promising about this project – we have an amazing team with the right ingredients of expertise.”
This article was written by Michele Carroll, Research Writer at USC Suzanne Dworak-Peck School of Social Work, originally published here.