A Novel “AI Physician” Forecasts Hospital Readmission and Additional Health Results

A computer software with artificial intelligence (AI) can reliably assess a patient’s risk of mortality, duration of hospital stay, and other care-related characteristics by reading their physician’s notes. The program, which was created by a group of researchers from NYU Grossman School of Medicine, is presently being used at hospitals affiliated with NYU Langone Health to forecast the likelihood that a patient who is released may need to return within a month.

Researchers have long studied computer algorithms designed to enhance medical treatment; some of these algorithms have been demonstrated to produce useful clinical forecasts. Few, though, are in use because doctors usually write in imaginative, personalized language that mirrors human thought processes, but computers are better at processing information presented in tidy tables.

Researchers claim that laborious data restructuring has been a barrier, but large language models (LLMs), a novel kind of artificial intelligence, are able to “learn” from text without the requirement for specifically prepared data.

The research team created an LLM known as NYUTron that can be taught using unmodified text from electronic health data to generate helpful judgments about patient health condition. The work was published online in the journal Nature on June 7. The program’s ability to forecast 80% of readmissions was demonstrated by the findings, which represents an improvement of around 5% over a normal, non-LLM computer model that necessitated reformatting medical data.

“Our results demonstrate the possibility of utilizing extensive language models to advise doctors on patient care,” stated Lavender Jiang, BSc, the study’s primary author and a doctorate candidate at the NYU Center for Data Science. “In order to address or even prevent readmissions, programs like NYUTron can notify healthcare providers in real time about factors that may lead to them and other concerns.”

Jiang continues, “The technology may expedite workflow and free up more time for doctors to interact with patients by automating routine tasks.”

Using sophisticated computer algorithms, LLMs determine which word would be most appropriate to complete a phrase by considering the likelihood that actual people would use the term in question. According to Jiang, the computer’s estimations are increasingly accurate over time the more data it is fed to “train” it to detect certain word patterns.

The researchers used millions of clinical notes from the electronic health records of 336,000 men and women who had gotten care at NYU Langone hospital system between January 2011 and May 2020 to train NYUTron for their study. Any medical document authored by a physician, including radiological reports, patient progress notes, and discharge instructions, was included in the resultant 4.1 billion word language “cloud.” Interestingly, medical professionals did not use standardized language, and the algorithm was able to decipher writer-specific abbreviations.

The results showed that NYUTron accurately assessed 79 percent of patients’ actual duration of stay (a 12 percent improvement over the usual model) and correctly identified 85 percent of those who died in the hospital (a 7 percent improvement over normal techniques). The comorbidity index, which measures the probability of coexisting diseases with a main ailment, and the possibility of an insurance refusal were also successfully evaluated by the instrument.

Eric K. Oermann, MD, a neurosurgeon and research senior author, said, “These results demonstrate that large language models make the development of’smart hospitals’ not only a possibility, but a reality.” “NYUTron’s predictive models can be easily built and quickly implemented through the healthcare system because it reads data directly from the electronic health record.”

Future research, according to Dr. Oermann, an assistant professor in NYU Langone’s departments of neurosurgery and radiology, may examine the model’s potential uses in extracting billing codes, predicting infection risk, and determining which prescription to purchase.

He issues a warning, saying that NYUTron should not be used in place of provider judgment that is specific to each patient and is only meant to be a support tool for healthcare professionals.

The National Institutes of Health grants P30CA016087 and R01CA226527 funded this research. An additional grant of W.M. Keck Foundation Medical Research was given.

Dr. Oermann works as a consultant for Sofinnova Partners, a venture financing company that focuses on the advancement of medicines, biotechnologies, and sustainability. In addition, he has stock in Artisight Inc., a provider of software for healthcare organizations, and his spouse works at Mirati Therapeutics, a firm that creates cancer treatments. These agreements’ terms and conditions are being handled in compliance with NYU Langone Health’s policies and procedures.