YuLin Zhen, Photography Editor

A recent study by Yale researchers demonstrated the potential of a machine learning approach to predict symptoms of post-traumatic stress disorder, or PTSD, for recent trauma survivors.

Researchers have been studying the medical applications of machine learning for only around a decade, and the team focused their efforts on pushing the boundaries of this innovative tool with a unique experimental design. The research stands out as a crucial milestone, as their reported prediction strengths are relatively high for clinical measures. 

“While a lot of studies usually are using cross-sectional designs and comparing patients with PTSD compared to healthy controls or compared to trauma-exposed healthy controls, this study focused on recent trauma survivors during the first 14 months after trauma exposure,” said Dr. Ziv Ben-Zion, a Fulbright postdoctoral fellow at Yale and first author of the study.

According to Ben-Zion, the data used in the study was “quite unique” and collected as part of his doctoral research from 2015 to 2020, at Tel Aviv Sourasky Medical Center in Israel. 

Then, Ben-Zion recruited individuals who arrived at the emergency department after experiencing potentially traumatic events, the most common being car accidents. 

The patients who experienced high levels of PTSD one month after admission — who were most likely to develop chronic PTSD — were assessed one month, six months and 14 months after admission. To monitor each patient’s progress, clinical assessments and fMRI scans, recording brain structure and function, were performed. 

Ben-Zion shared good news: most of the patients recovered sometime during the 14 months of study. 

By the end of data collection, Ben-Zion had obtained a multi-domain data set detailing PTSD symptom severity — CAPS-5 total scores, on a scale of zero to 80 — as well as cognitive functioning and neural data for each of the 171 participants. 

This data set was used to develop the predictive machine learning model. The team used connectome-based predictive modeling, a machine learning technique originally developed in the Constable Lab at Yale that has gained popularity over the past decade.

The model works by applying 10-fold cross-validated regression models to whole-brain functional connectivity data derived from the fMRI BOLD signal to predict behavioral measures of interest, such as PTSD symptoms.

While the study showed no association between whole brain connectivity and symptoms at the six-month time point, there appeared to be high predictive ability at one month and 14 months. 

According to Ben-Zion, these findings align with current clinical knowledge about PTSD, which defines the six-month time point as a fragile and dynamic point in the recovery process.

After breaking down the PTSD symptoms into clusters based on the DSM-5, the team also noticed that different clusters were driving predictions at different time points, which suggests the connection of various regions in the brain to PTSD progression. 

For Dr. Scheinost, this finding will benefit the growing understanding of PTSD.

“I think it helps shift some of the neurobiological thinking about PTSD—moving away from characterizing a few key regions (like the amygdala) to more widespread, whole-brain alterations,” Scheinost wrote to the News. “That’s not to say that the amygdala or other single areas are not crucial to PTSD, just that we are likely only capturing a piece of the picture.”

A key aspect of the team’s work from the beginning was collaboration. Ben-Zion and his advisor, Dr. Ilan Harpaz-Rotem, first connected with AJ Simon, a doctoral student in Interdepartmental Neuroscience, and his rotation advisor at the time, Dr. Dustin Scheinost, around two years ago to expand their study on the initial data set. 

For Simon, who was a first-year rotating graduate student in Dr. Scheinost’s lab at the time, the project was an exciting new chance to explore machine learning models.

“I jumped on board because it was my opportunity to learn connectome-based predictive modeling and to apply it in a way where there was potential for translational impact,” said Simon.

The team worked on analyzing the data for six months and writing the paper for another six months before moving on to the review phase of their research, which took over a year before publication.

Looking to the future of the study, Dr. Ben-Zion hopes that other researchers will try to replicate the study with new data sets.

He noted that more researchers are currently publishing their own independent — and sometimes, inconsistent — findings, rather than focusing on replication to produce more robust results that build on prior studies.

While Harpaz-Rotem notes that it’s still a long way before MRI scans can be used as predictive clinical tools, the study shows promising results for the field’s future.

“I think the study demonstrates the capacity of [connectome-based predictive modeling] to be useful to identify the brain regions that are involved in the potential development of PTSD and think how we can intervene to prevent the development of PTSD based on this knowledge gained,” Harpaz-Rotem wrote.

To learn more about connectome-based predictive modeling, see here.