AI Delirium Detection: How a Hospital Tool Helped Staff Find Delirium Twice as Often

Delirium is a sudden state of confusion that can strike older adults in the hospital, and it’s surprisingly easy to miss. A new peer-reviewed study in JAMA Network Open shows how AI delirium detection worked in real life at a New York hospital. After an AI risk tool went live at Mount Sinai Morningside, the share of delirium checks that found delirium roughly doubled, from about 9% to nearly 20% each month.

It’s an encouraging real-world result, with some important caveats. Let’s walk through both.

Why delirium is so often missed

Delirium can come on quickly and can look like tiredness, dementia or simple confusion. According to the study, between 55% and 70% of delirium cases are missed by the treating doctors and nurses.

That’s a real problem, because care teams can only treat what they recognize. Screening every older patient by hand takes trained staff and a lot of time, so hospitals need a smarter way to decide who to check first.

How the AI tool works

The Mount Sinai team built a machine learning model that pulls together several streams of information already in a patient’s electronic health record:

• structured data, such as lab results and vital signs
• semi-structured data, such as clinical assessments
• clinical notes, read with natural language processing (AI that understands written text)

From all of this, the model gives each patient aged 60 or older on non-intensive-care units a daily delirium risk score. High-risk patients show up color-coded in staff patient lists, and the hospital’s delirium team uses those flags to decide who gets a standard bedside delirium assessment first.

AI delirium detection in a real hospital

The model was first developed at The Mount Sinai Hospital. This study tested whether it would still work at a different hospital, Mount Sinai Morningside, without being retrained on that hospital’s data.

A tough test

That’s a bigger deal than it sounds. The study notes that AI delirium models often perform worse when moved to a new hospital. Morningside’s patients were noticeably different: older on average, more often medical rather than surgical patients, and more likely to have dementia.

The results

Across a year of live use, from December 2024 to December 2025, the model held up well:

• It scored 0.91 on a standard accuracy measure (AUROC) at Morningside, close to its 0.94 at the original hospital.
• It caught about 92% of patients who turned out to have delirium.
• The monthly rate at which delirium checks found delirium rose from 9.23% before the tool to 19.59% after it.

The headline: an AI tool moved to a new hospital, kept working without retraining, and was linked to staff finding delirium about twice as often.

Important limitations

This result is promising, but it needs careful reading.

Many false alarms

Only about 26% of patients the AI flagged as high risk actually had delirium. That means many false positives, and the authors warn this could lead to alert fatigue among staff.

A before-and-after comparison

The “before” numbers come from 2018 to 2019, years before the “after” period. Many things can change in that time, so the rise in detection is linked to the AI tool but doesn’t prove the tool caused all of it.

Built-in biases

Because the AI decided who got assessed first, the study can’t fully measure what it missed among patients who weren’t checked. The authors flag this verification bias themselves and say a prospective study with uniform assessments would be needed to measure it properly.

One health system

Both hospitals belong to the same health system and share records and workflows. The authors say testing in completely separate health systems is still needed.

People still matter

The tool’s success depended on trained staff who could reliably carry out delirium assessments. AI pointed the way, but people did the diagnosing and treating.

Why it matters

Older adults make up a large share of hospital patients, and catching delirium sooner gives care teams a chance to act sooner. A tool that can move to a new hospital without costly retraining could make that kind of help easier to roll out.

This study doesn’t change what patients or families should do on their own. If you’re worried about sudden confusion in a loved one in the hospital, tell their care team.

The bottom line on AI delirium detection

This peer-reviewed quality improvement study shows AI delirium detection can travel to a second hospital, keep its accuracy, and support staff in finding delirium in older patients far more often. With false alarms to reduce and wider testing still ahead, it’s a hopeful, practical step toward better hospital care for older adults.

Read the full study in JAMA Network Open: External Validation of a Machine Learning Model for Delirium Risk.