A Stanford AI watches ICU patients around the clock and spotted a quiet warning sign of delirium
Patients who developed delirium moved and communicated with purpose far less, the system found across 280 Stanford Hospital patients. It is a prototype, not a diagnosis tool.
Stanford researchers have built an AI system that watches intensive care patients 24 hours a day and lets doctors ask how a patient’s behavior has changed, a step toward catching delirium earlier, according to Stanford’s Institute for Human-Centered Artificial Intelligence. Across 280 patients at Stanford Hospital, the system found that those who developed delirium did far fewer purposeful things, such as using objects, gesturing or handing things to staff.
Delirium is a sudden state of confusion. It is among the most common complications in intensive care units, especially for patients on breathing machines, and it is linked to lasting thinking problems and a higher risk of death, Stanford HAI said. “Most patients on a ventilator develop delirium during their ICU stay, and for many, cognitive impairment may last for years,” said Dr. Paul Tang, a co-first author and adjunct professor at Stanford’s Clinical Excellence Research Center.
What this means
Doctors usually check a patient’s thinking only a few times a day, with quick questions such as “What day is it?” The most severe form of delirium shows up as a lack of activity, which is easy to miss in a short visit, Stanford HAI said. A camera system that notices a patient has gone quiet could flag the problem sooner, giving care teams a chance to step in.
For now, it is a research prototype. The finding is a pattern across groups of patients, not proof that the system can tell whether a single patient has delirium, and doctors, not the software, make the call. The study has been posted on arXiv, a site for research papers that have not yet been peer reviewed.
How it works
The system, called TRACER, uses two cameras aimed at each ICU bed. Once a minute, it notes which of 15 basic actions a patient is doing, such as talking, moving a hand to the face or lying still. It rolls those notes into hourly reports and combines them with the patient’s electronic health record, including the medicines given.
At the bedside, a doctor can ask questions in plain language. In one example from the paper, a doctor asked how the patient’s behavior changed over three days, and TRACER answered that activity peaked overnight from 8 p.m. to 4 a.m. Doctors can then jump to the matching video clips to check for themselves.
TRACER was not trained on patient data. It is built on existing general-purpose AI models, which the team says should make it easier to share with other hospitals without privacy concerns.
What the study found
The researchers used video recorded in two Stanford Hospital ICUs between August 2021 and June 2025, about 60,000 hours of camera footage. Of the 280 patients, 53 had at least one positive delirium screening. In those patients, purposeful actions showed up in a median of about 5% of observed minutes, compared with about 15% for the others, the paper reports.
TRACER identified actions correctly about 85% of the time in a test set. It answered doctors’ questions correctly 61% of the time, better than the 49% scored by the strongest comparison system, but those questions covered only two patients. The team, led by Stanford computer scientist Fei-Fei Li, is refining the system.
Source: Stanford HAI, Nov. 20, 2026; Durante, Chandrasegaran, Tang et al., arXiv preprint.