Early and well-coordinated discharge planning in a hospital can help ensure a smooth handoff by care providers, reduce the length of stay, and lower a patient’s risk of falling or acquiring an infection. However, determining which patients can be sent straight home and which need extra care at a skilled nursing facility can be difficult and time-consuming for clinical care teams.

“Sometimes it takes a few days for providers to decide what’s going to happen after the hospital stay because they’re so focused on what’s happening to the patient at that very moment,” says William R. Small, MD, associate medical director of clinical informatics and applied AI in the Medical Center Information Technology’s Department of Health Informatics at NYU Langone Health. “The problem is that a patient may be stuck in the hospital longer than they need to be, or unable to move on to the next phase of their care.”
 
A recent study published in the journal npj Health Systems describes a new AI tool that could help avoid this problem. Dr. Small and colleagues first prompted the chatbot ChatGPT to condense doctors’ clinical notes into concise risk snapshots—concise summaries of a patient’s health, living situation, and ability to perform daily tasks. The researchers then trained nine separate AI programs, known as large language models, or LLMs, to use the snapshots to predict whether patients were discharged to a nursing facility (as 15 percent of NYU Langone inpatients are) or sent home. The LLMs, they found, produced more accurate predictions from the ChatGPT-generated summaries—with the top model achieving an accuracy rate of 88 percent—than they did from the full-length notes.

“Our two-step approach acts like a fast but careful reader, turning a complex medical note into a simple summary of what matters most for discharge planning,” says senior study author Yindalon Aphinyanaphongs, MD, PhD, director of operational data science and machine learning.

Working with the electronic health records of hospitalized patients, the team first focused on the lengthy clinical summary called a history and physical, the first major doctor’s note after a patient is admitted. “That assessment discusses the patient’s relevant medical history and tells the story of how they got to the point of needing to be admitted,” says Dr. Small. While descriptions of a patient’s social support system, physical and cognitive status, and other medical details can help providers assess the need for skilled nursing care, the notes often include a large volume of extraneous information geared toward immediate care requirements.

“To filter out the noise, we developed a generative AI model to read the note, extract the information related to skilled nursing care risk factors, and generate the risk snapshot,” explains study coauthor Eric K. Oermann, MD, director of the Health AI Research Lab at NYU Langone.

For generating the risk snapshots, the researchers directed the AI model to focus on words and phrases that previous studies, discussions with doctors and nurse case managers, and their own experience with hospitalized internal medicine patients suggested were most relevant. The analysis found that the most predictive search terms were “nursing home” and “SAR,” the acronym for subacute rehabilitation; both refer to a skilled nursing facility. Other highly predictive words included “weakness,” “fall,” “dementia,” “rehab,” and “AMS,” the acronym for altered mental status.

The team trained the LLMs on both the doctors’ original notes and the AI-written risk snapshots for 3,000 patients. Then they tested how well the models used the longer or shorter notes to predict the posthospital destinations for 1,000 patients with known outcomes. To gauge whether each model could explain “why” it deemed a patient high or low risk, researchers asked nurse case managers to review the risk snapshots and determine—without seeing the model’s prediction—whether each patient went to a nursing facility. The case managers’ responses aligned well with the winning model’s predictions, suggesting that the risk snapshots contained good predictive information.

“An innovative aspect of our study is that we can provide a summary that gives the care team greater context about why a patient is at high risk for needing skilled nursing care,” says Dr. Small. “And that provides more nuance about whether to trust the information or not.”

Dr. Small hopes an accurate model might also enable productive conversations among providers who influence discharge decisions, including physical and occupational therapists and social workers. He notes that the model is most useful in the early stages of hospitalization. In fact, the prediction is designed to expire after five days to ensure that decision-making is not based on outdated information.

Since publishing the study, the research team has introduced the discharge prediction tool as an option in Epic, NYU Langone’s electronic health record system, to assess internal medicine patients. As providers become more comfortable using it, Dr. Small hopes that a randomized controlled trial that either hides or displays the prediction values will offer further proof of the model’s value.

The initial model is merely a starting point in using AI to refine postdischarge predictions. “Intuitively, we know that there’s information we’re not capturing,” says Dr. Small. “For the next phase, we’re looking at notes from the emergency department as well as physical therapists and wound care nurses to cast a much broader net.” With the added input, the model’s next iteration is making even more accurate predictions, he says.