WASHINGTON – Artificial intelligence has been proven to help read medical images and has been proven to pass medical licensing exams.
Now, a new AI tool has demonstrated the ability to read physician records and accurately predict patient mortality risk, readmissions, and other outcomes important to care.
Designed by a team at New York University’s Grossman School of Medicine, the software is now being used at the university’s affiliated hospitals across New York in hopes of becoming a standard part of healthcare.
The predictive value study was published Wednesday in the journal Nature.
Lead author Eric Orman, a neurosurgeon and computer scientist at New York University, told AFP that while non-AI predictive models have long existed in medicine, the data they require involves cumbersome reorganization and formatting. He said it was rarely used in practice because it was needed.
But “one thing that is common in all medical practice is that doctors write notes about what they see in the office and what they discuss with their patients,” he said.
“So our fundamental insight was, can we start with medical records as a data source and build a predictive model on top of that?”
A large-scale language model, called NYUTron, was trained on millions of clinical notes from the health records of 387,000 people treated within New York University Langone Hospital from January 2011 to May 2020. it was done.
These include patient follow-up notes, radiology reports, discharge orders, and anything else a doctor creates, resulting in a corpus of 4.1 billion words.
One of the main challenges for this software was interpreting the natural language written by doctors. Natural language, including the abbreviations doctors choose, varies greatly from person to person.
By looking back at what happened, the researchers were able to calculate how often the software’s predictions turned out to be accurate.
They also tested the tool in a real-world setting, training the tool on records from a Manhattan hospital, for example, and then seeing how it performed at a Brooklyn hospital targeting different patient demographics. .
-not a substitute for humans-
Overall, NYUTron identified 95% of people who died in hospital before discharge and 80% of patients readmitted within 30 days to anxiety.
This prediction outperformed most physicians as well as non-AI computer models currently in use.
But to the team’s surprise, “the superlative doctor, who was actually a very famous doctor, had a superhuman performance better than the model,” Oermann said.
“Technology and medicine’s sweet spot doesn’t always deliver superhuman results, but it really raises that baseline.”
NYUTron also accurately estimated 79 percent of patients’ actual hospital stays, 87 percent of cases where patients were denied coverage, and 89 percent of cases where patients had additional symptoms to their primary disease.
AI will never replace the doctor-patient relationship, says Oermann. Rather, it will help “seamlessly provide more information to the point of care so that doctors can make more informed decisions.”
