Underdogs Take on DARPA Triage Challenge, Secure 3rd Place

Researchers from the University of Virginia, including two graduates of the School of Data Science’s M.S. in Data Science (MSDS) program, made an impressive showing in the DARPA Triage Challenge: Secondary Triage Data Competition. Despite competing as a self-funded team against well-resourced rivals, they secured third place in the opening event, demonstrating the power of applied data science in high-stakes environments.

The competition, designed to improve trauma care decision-making, challenged participants to develop algorithms that detect physiological signatures of injury using anonymized, multi-modal patient data.

During secondary triage, medics must quickly assess and prioritize casualties for medical evacuation and care. The challenge, organized by the Defense Advanced Research Projects Agency (DARPA), aimed to advance this process by leveraging data-driven solutions to predict which patients require life-saving interventions (LSIs). Competitors worked with real-world physiological data from trauma patients, provided by the DARPA Research Infrastructure for Trauma with Medical Observations (RITMO) program, to refine predictive models that could support medics in pre-hospital environments.

UVA Team’s Resourceful Approach

Unlike other teams with dedicated DARPA funding, the UVA group was self-backed and therefore had to make it through a qualification round to even get to the main event. Representing the UVA School of Medicine, and holding eight UVA degrees among them, members included Dr. Andrew Barros (CLAS ’05, MD ’14, MS ’16, MSDS ‘24), assistant professor of medicine; Dr. Thomas Hartka (MD ’10, MSDS ’21), associate professor of emergency medicine; and Will Ashe (CLAS ’18, PhD ’22), data scientist and data manager in the Center for Advanced Medical Analytics. Their combined expertise in clinical medicine and data science provided a unique advantage, enabling them to interpret complex physiological data and develop predictive models tailored to trauma care challenges.

They relied on Python, a core programming language in the data science community, to build and refine their models. By leveraging data analysis tools and libraries such as scikit-learn, pandas, and XGBoost, they efficiently processed massive datasets and developed algorithms that could provide real-time, actionable insights.

The three-person team was also one of the smallest teams participating in the competition, relying on their skilled backgrounds in lieu of additional resources. 

“Most of the teams had clinicians as consultants, but we had clinician who were also coders and developers,” Hartka said. “I think that gave us an advantage, having both been trained in medicine and also having the opportunity to have completed the MSDS program at UVA.”

Innovative Model Development

The team’s approach involved extracting time-series features from heart rate variability and other vital sign data, using decision tree models and algorithms to forecast eight different LSIs. Their predictions covered a range of critical needs, from IV fluids and blood transfusions to chest tube placement and resuscitation efforts. Working within a strict framework, they had to generate real-time predictions based on incoming 15-minute data increments while ensuring their model was both accurate and computationally efficient.

Ultimately, the hope is that the data and insights gleaned from this competition could be used to help predict what LSIs that a patient that has experienced a particular trauma would need, and even what facility would best be able to support their emergency care.

“I think that’s why this competition was so natural for us was that a lot of the same things that they're worried about in battlefield evacuations, I worry about in my ICU or Tom worries about in the emergency department,” Barros said.

Hartka and Barros agreed that their MSDS training and knowledge of statistical modeling played a big role in developing a streamlined, scalable approach to the challenge. “I do think part of the reason why we were so successful is that sort of combination of domain knowledge and then data science background,” Barros said.

With the second phase of the three-part competition on the horizon, the team plans to expand on their work, potentially incorporating deep-learning techniques to further refine their models. They also hope to secure additional funding and possibly expand the team to continue their research and enhance their capabilities in the next round.

But the experience validated the team’s approach and demonstrated that well-directed effort and expertise can rival the advantages of heavy funding. And as this group looks ahead to the next phase, they were excited for the chance to push their models even further and showcase the potential of data science to transform medical decision-making in high-stakes environments.