I woke up one Saturday morning with double vision. Of course I wanted to know why it was happening. But three other questions were more urgent: What kind of doctor did I need to see? How quickly did I need to see that doctor? And, of course, how serious was the situation – was I at risk of dying?
The why – that is, the diagnosis – gets a lot of attention in medicine and in medical AI, and for good reasons. We want our doctors to nail common diagnoses 100% of the time, and we want them to be the “Sherlock Holmes of diagnostics” for medical mysteries when necessary.
But in AI specifically, what the excitement about diagnostics obscures is a critical layer of infrastructure that urgently needs to be built: triage. Triage is the process of safely getting the right patient to the right care provider at the right time. I learned this first-hand with my double vision. I was appropriately sent to urgent care by a triage nurse when it initially presented. I was otherwise feeling well, and the main concern was to make sure that I wasn’t having a stroke. When it was determined that I wasn’t, the plan was to see a neuro-ophthalmologist during the upcoming work week.
But my condition deteriorated over the next 24 hours. I was sent to the ED for a full emergency work up, including MRI head imaging. In the end, the best diagnostic guess was cranial nerve palsy related to Covid. Later, in the setting of another viral illness, I had another, distinct cranial nerve palsy that led to an even more extensive work up over many months that has, to date, not yielded satisfactory answers, only a seemingly unlucky run of viral-induced cranial nerve palsies: one fully resolved, one still healing.
When I reflect back on the whole endeavor over 18 months, though, I think about the fact I always felt safe. I am fortunate enough to have access to good doctors and hospitals. But access isn’t enough for good care. But what makes them good is not only their diagnostic acumen – it’s also their ability to understand risk. To be focused on my safety and what is, in the face of scary and perhaps unresolvable diagnostic uncertainty, the next best step in my care. In short, they are very good at triage.
I work at a healthcare AI company, where I’m a medical AI informaticist. On the clinical team we say, “Diagnosis is besides the point.” It’s an unexpected statement coming from doctors and nurses, but given my lived experience, it makes sense. Though it’s an exaggeration, it underscores something counterintuitive – and vital.
Yes, AI is very good at diagnosing, and it isn’t surprising that a lot of attention in AI goes to diagnostics. Once the deep neural network era took hold, it became clear that methods such as convolutional neural networks could achieve “superhuman” performance in tasks such as identifying pneumonia in chest x-rays. But some of the thorniest problems in medicine that AI can improve have nothing to do with diagnosis. They have to do with triage.
What’s changing is where triage happens. Though we think of it as practiced in clinical settings, triage happens anywhere people are presenting with symptoms and concerns, and conversational AI agents powered by LLMs meet patients where they are. The agents are always available and always updated. They even can have the context awareness of the data in a patient’s electronic health record when fielding questions from a nervous patient late at night.
But as AI agents enter healthcare settings, it is important to realize that any conversation can become a clinical conversation. When that happens, the agent needs to be able to quickly make a decision about what to do, just like the doctors did with me during my urgent care visit. A scheduling agent for a primary care office needs to route people to 9-1-1 if a patient wants to be seen for their crushing chest pain. An informational agent for a specific drug needs to hand off calls to the right person if the caller says they took the wrong dose and now don’t feel quite right.
“Urgent” and “emergency” are the most time-sensitive triage settings, and the split-second decisions they necessitate will be hindered by doubt if AI’s systematic hunches are proven to be unworthy of enhancing the clinical team’s judgment. Additionally, aside from those fixed-moment scenarios, triage is a longer, fluid process that responds to evolving symptoms and information. I go back to how my care actually happened. Clinically, it’s rare that everything becomes clear at a point-in-time. Mostly, care evolves over time as conditions change, as mine surely would (and did). That common scenario, which happens millions of times daily in the U.S. makes it essential that the initial triage direction be as true as the available information can point to. Of course diagnosis matters, and AI’s role in it will only grow. But, first, we need it safe for triage.
Photo: Issarawat Tattong, Getty Images
Dr. Michael Choma is a physician-engineer and Medical Informaticist at Infinitus Systems, where he helps ensure that large-language-model (LLM)-powered agents are clinically safe, effective, and aligned with real-world healthcare needs. Drawing on over two decades of experience at the intersection of medicine, engineering, and artificial intelligence, he bridges clinical insight with technical rigor to advance responsible AI in healthcare.
This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.
