NYU Langone researchers developed an artificial intelligence tool that identifies which patients will require additional support from nursing facilities at an 88% accuracy rate — with plans to bring it to a real-world clinical setting in the coming months.
The new model, tested as part of a January study, first prompts an AI model to scan a doctor’s note for seven predictive factors — including admission location, social support and cognitive status. It then produces a risk summary that an in-house large language model called NYUTron — which was trained on data from nearly 400,000 NYU Langone patients — uses to form recommendations for doctors based on variables such as how physically independent the patient is and whether they came from home or a nursing facility.
The technology can help medical professionals adjust their planning for complex care, so they avoid instances where patients are ready to be discharged from the hospital but need to find a nursing home first.
“There’s special forms, they have to send referrals out to a bunch of different facilities, get insurance acceptances — so it’s complex,” William Small, one of the study’s researchers, told WSN. “The earlier we can start it, the quicker we can get patients out of the hospital, and minimize the amount of time they’re just waiting around and at risk of infections.”
Researchers tested NYUTron, among other models, against nurse case managers who reviewed the same AI-generated summaries and independently predicted whether patients would require additional care. Patients that the model labeled high-risk were 13.5 times more likely to be flagged the same way by nurses, yielding a success rate that researchers said was promising.
NYU Langone Health has said that NYUTron can also help forecast in-hospital mortality, approximate duration of hospitalization and project denied insurance claims.
“Within the next two-to-three months, a bunch of providers will have access to it across the health care system,” Small said. “But that’s not to say we aren’t still improving it.”
The model currently only runs on one doctor’s note, but the team is working on improving the model to extract all relevant patient notes from the first 24 hours of intake. Researchers also hope to expand the project by training models on patients beyond just internal medicine.
Small ensured that all patient information processed by NYUTron and the other AI models used, including GPT 4.0, is protected by HIPAA regulations and will not be shared with OpenAI or any other database.
“We’re putting AI into the health system right now — not in an autonomous fashion, but in a partnership with clinicians,” Small said. “The idea is to give them the appropriate information at the appropriate times to facilitate things that they already do.”
Contact Jayshil Blomstedt at [email protected].