AI more efficient than doctors? The question arises after a new tool demonstrated its ability to anticipate deaths, hospital readmissions and other complications. Created by a team from the Faculty of Medicine of Langone (New York), the Grossman School of Medicine, the software is now being tested in several partner hospitals.
A study on its potential interest in the medical community has been published in the scientific journal Nature. Its main author, Eric Oermann, neurosurgeon and computer engineer, explains that while non-AI-based predictive models have been around for quite some time, they are little used in practice because they require heavy input and formatting work. Datas.
“One thing that is common in medicine is that practitioners take notes on what they see, what they talk about with patientshe noted in an interview with AFP. Our basic idea was therefore to know if we could start from medical notes as data sources and build predictive models”he continues.
Data from 387,000 patients over nine years
The predictive model, dubbed NYUTron, was trained from millions of medical observations from the records of some 387,000 patients treated between January 2011 and May 2020. These observations included doctors’ written reports, notes on the changes in the condition of patients, X-rays and medical imaging, or even the recommendations given to patients on leaving the hospital, all forming a corpus of 4.1 billion words.
One of the main challenges for the software was to correctly interpret the language used by doctors, which varies greatly between professionals, especially in the abbreviations used. The tool was also tested in real conditions, in particular by training it to analyze reports from a hospital in Manhattan, then comparing the results to those of a hospital in Brooklyn, with different patients. By looking at what happened to the patients, the researchers were able to measure the number of times the software’s predictions turned out to be correct.
Not a substitute
The NYUTron software would thus have identified 95% of patients who died in partner hospitals before discharge authorization, and 80% of readmissions less than a month after the latter. Results that exceeded the predictions of most doctors, as well as those of non-AI-based computer models.
The software would also have predicted with a success rate of 79% the duration of hospitalization of patients, and at 87% the cases in which patients were refused reimbursement of care by their insurance. 89% of the cases in which the patient suffered from additional pathologies would have been identified. Artificial intelligence will never replace the patient-doctor relationship, says Dr Oermann, but it could “provide more information […] physicians to help them make informed decisions”.