DARPA Pushes AI Into the Hardest Medical Decisions
DARPA is advancing AI that can navigate disaster zones, interpret noisy real-world vital signs, and support medics with faster, more consistent triage decisions. By training algorithms on complex casualty data and rigorously keeping humans in the loop, the DARPA Triage Challenge aims to transform how first responders and trauma teams prioritize care—on the battlefield, in disasters, and beyond.
Transcript
00:00:01 What if artificial intelligence could assist first responders and making split-second medical triage calls?
>> Mass casualty events can overwhelm human decision-m DARPA is pushing artificial intelligence to bring speed, consistency, and life-saving precision to emergency medical triage.
>> Teams are using their expertise in machine learning to train their models
00:00:23 on real world data. These models are designed to detect patterns in the data and ultimately these patterns will be used to inform their triage decisions. A lot of the models that are available right now off the shelf that have been trained for heart rate and respiration rate and things like that are all models that have been done in a clean environment. It's a lot to teach a a
00:00:49 robot to reason.
>> It's very very difficult. So for example some fields like respiratory rate, heart rate, they're very easy to calculate when we are in a controlled environment. On the other hand, when we're in the wild, when we cannot touch the people, it's very very difficult to get the vitals. In the systems challenge, sophisticated algorithms guide
00:01:07 semiautonomous systems to navigate chaotic disaster environments, identify casualties, pinpoint heart rates, respiration rates, and vital signs, and prioritize victims for rescue. In the data challenge, powerful algorithms sift through massive realworld, often noisy data sets. Identifying features and patterns in the vitals data that could indicate the need
00:01:31 for an intervention.
>> The goal is to prepare the AI for anything. So, having as much casualty data as possible is important in training and tuning the AI. So one of the things that we're doing is working with the trauma surgeons to see if these algorithms could be installed potentially in trauma bays to help not just in field environments but also in
00:01:50 hospital settings. Making decisions in these situations is incredibly difficult. Information is limited, time pressure is high, data is often times confusing and uncertainty is everpresent. These nuanced decisions are often highly variable between any two individuals. That's why we are rigorously testing and keeping humans in the loop.
00:02:11
>> A lot of times a real medic will take a person, roll them over and do a blood sweep on their back. Um, and our robots don't do that. So, I would say that is probably one of the most difficult things to tell and one of the most important
>> to validate uh vital signs like heart rate and respiratory rate. We have our human actors wear uh standard
00:02:32 off-the-shelf uh equipment that more commonly you would use when you're running or cycling. And that calculates your heart rate and your respiratory rate and those values are used to compare to what the robots assess.
>> Most exciting about this battle of algorithms is that the lessons learned from the DARPA Triage Challenge could change the way we save lives in mass
00:02:51 casualty incidents.