‘AI Doctor’ Predicts Patient Outcomes and Hospital Readmission

Some of these algorithms have demonstrated their potential in providing crucial clinical predictions. However, their adoption has been limited as they are best suited to analyze neatly structured data, while healthcare providers often express their observations in individualized and creative language, mirroring human thought processes. Now, a breakthrough AI program can interpret physicians' notes to reliably calculate patients' mortality risk, hospital stay duration, and other significant care factors. This technology, which automates basic tasks, could streamline workflows and free up more time for doctors to communicate with their patients.

This innovative tool has been developed by a research team from NYU Grossman School of Medicine (New York, NY, USA) and is currently in use at NYU Langone Health hospitals to estimate the likelihood of a patient getting readmitted within 30 days post-discharge. Amidst the challenges posed by laborious data restructuring, a novel AI type, known as large language models (LLMs), has emerged, which can "understand" text without requiring specially formatted data. LLMs utilize specific computer algorithms to predict the most suitable word to complete a sentence based on the probable use of that term by humans in the given context. As the computer receives more data to "learn" these word patterns, its accuracy in making predictions improves over time. The NYU Langone researchers created an LLM, named NYUTron, which can be trained using unmodified text from electronic health records to evaluate patient health conditions effectively.

In their study, the researchers trained NYUTron using millions of clinical notes gathered from the electronic health records of 336,000 individuals who received care within the NYU Langone hospital system from January 2011 to May 2020. The resulting language "cloud" of 4.1 billion words included all types of doctor-written records, such as radiology reports, patient progress notes, and discharge guidelines. Interestingly, the program was able to handle the non-standardized language used by different doctors, even deciphering unique abbreviations used by individual writers. The study revealed that NYUTron was able to accurately pinpoint 85% of in-hospital deaths (an improvement of 7% compared to traditional methods). The tool estimated the actual length of stay of 79% of the patients (an improvement of 12% compared to the standard model) and was also successful in assessing the possibility of other conditions accompanying a primary disease (comorbidity index) as well as the likelihood of denial of insurance.

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

Related Links:
NYU Grossman School of Medicine