Artificial intelligence may eventually be asked to help with one of medicine’s most difficult questions: what would a patient want if they could not speak for themselves?
Researchers have considered that possibility for more than a decade. The idea is not to replace doctors or loved ones, but to create AI surrogates that could help estimate a patient’s values and goals when end-of-life decisions must be made under pressure.
The idea behind AI surrogates
An AI surrogate is a proposed system that would use available patient data to predict what an incapacitated person might choose in a critical medical situation. The concept sits at the intersection of health care AI, trauma care, family decision-making, and medical ethics.
Experts following the work told Ars that no hospital has yet deployed these systems. Still, AI researcher Muhammad Aurangzeb Ahmad is taking early steps toward a possible pilot at a US medical facility.
Ahmad is a resident fellow working with trauma department faculty at the University of Washington’s UW Medicine. His research is based at Harborview Medical Center in Seattle, a public hospital in the UW Medicine health system.
The work is still early. Ahmad said he has spent most of this year in the conceptual phase, focused on how to test machine learning models against Harborview patient data. No patient has interacted with the models, and UW Medicine spokesperson Susan Gregg said there is “considerable work to complete prior to launch” and that any system “would be approved only after a multiple-stage review process.”
“This is very brand new, so very few people are working on it,” Ahmad told Ars.
What the model would try to learn
Ahmad’s current models are focused on information Harborview already collects. That includes injury severity, medical history, prior medical choices, and demographic information.
The research question is whether those inputs can help predict what a patient would have wanted when a choice must be made about treatment. Ahmad described the process as feeding that information into a machine learning predictive model and then studying performance in retrospective data.
The first testing problem is basic but serious: the model’s accuracy can only be checked if patients survive and can later say whether the prediction matched what they would have wanted. Ahmad has said that this is only a first step. Testing could later expand to other facilities in the network, with the aim of developing AI surrogates that can predict patient preferences about “two-thirds” of the time.
In a future version, Ahmad imagines adding textual data. That might include patient-approved recorded conversations with doctors. Trusted human surrogates, such as family members, could also provide chats or texts with the patient. In the most “ideal” form, patients would interact with AI systems throughout their lives and give feedback as they age through the health system.
That vision depends on time and data. As Ahmad put it, “It takes time to get the relevant data.” Before any human subject testing, he said approval from an institutional review board (IRB) would be required.
Why end-of-life decisions are hard to model
The problem AI surrogates are meant to address is real. Doctors often must make urgent decisions when patients cannot communicate. A patient may have wanted to avoid a ventilator, dialysis, or cardiopulmonary resuscitation (CPR). Another patient may fear infections, discomfort, or relying on a machine for life support.
Some patients have never stated a preference. The source article gives the example of young people involved in accidents, who may not have previously discussed end-of-life care.
Emily Moin, a physician in an intensive care unit in Pennsylvania, told Ars that time matters in these cases, but human surrogates remain essential when they are available. In fast-moving emergencies, clinicians may provide CPR until they reach a clinical judgment that the effort is no longer indicated or until they can engage a surrogate decision maker.
“These decisions are dynamically constructed and context-dependent,” Moin said.
That point goes directly to the limits of AI. A model may be tested by asking recovered patients what they would have wanted before recovery. But Moin warned that this may not accurately represent the decision as it existed in the moment. Some patients have reported that their preferences changed after lifesaving treatments because they then knew what to expect.
Health systems have long encouraged patients to complete “advanced directives” to record preferences. But the source article notes that preferences can be unstable and sometimes change within days. That creates a hard target for any predictive model, even one built from medical history and prior choices.
The case for caution
Ahmad sees possible value in AI surrogates because these decisions can be “very emotionally taxing” for doctors and loved ones. Family members may second-guess whether they are honoring the patient’s wishes. Ahmad believes AI could help improve the odds of getting those wishes right.
But the medical setting makes caution unavoidable. Ahmad said practitioners in this space are more conservative, and he added that this is “rightfully so.” He does not describe AI surrogates as perfect copies of patients. Instead, he expects rigorously tested systems that doctors and loved ones could consult alongside all other known information.
Gregg told Ars that UW Medicine supports “thoughtful exploration of innovative ideas, such as the potential responsible and transparent use of AI surrogates in end-of-life care,” connecting the work to questions about honoring patient wishes when patients cannot communicate directly or have no next of kin.
The unresolved issue is not whether AI can produce a prediction. It is whether that prediction is meaningful enough, tested enough, and clinically appropriate enough to enter a life-or-death conversation.
What comes next
For now, AI surrogates remain research, not hospital practice. Ahmad’s work has not enrolled patients at Harborview, and the project is still defining scope and theoretical considerations.
The strongest version of the idea would require more than a model. It would require patient participation, review, testing, transparency, and a clear role beside doctors and human surrogates. It would also have to confront the central challenge raised by Moin: end-of-life preferences are not fixed data points waiting to be retrieved.
That is why the future of AI surrogates in end-of-life care remains uncertain. The technology may become a tool for decision support, but the source article makes clear that medicine has not yet crossed that line.